doi
stringlengths 28
28
| title
stringlengths 19
311
| abstract
stringlengths 217
5.08k
| plain language summary
stringlengths 115
4.83k
| article
stringlengths 3.87k
161k
|
|---|---|---|---|---|
10.1371/journal.pgen.1002760 | Protective Coupling of Mitochondrial Function and Protein Synthesis via the eIF2α Kinase GCN-2 | Cells respond to defects in mitochondrial function by activating signaling pathways that restore homeostasis. The mitochondrial peptide exporter HAF-1 and the bZip transcription factor ATFS-1 represent one stress response pathway that regulates the transcription of mitochondrial chaperone genes during mitochondrial dysfunction. Here, we report that GCN-2, an eIF2α kinase that modulates cytosolic protein synthesis, functions in a complementary pathway to that of HAF-1 and ATFS-1. During mitochondrial dysfunction, GCN-2–dependent eIF2α phosphorylation is required for development as well as the lifespan extension observed in Caenorhabditis elegans. Reactive oxygen species (ROS) generated from dysfunctional mitochondria are required for GCN-2–dependent eIF2α phosphorylation but not ATFS-1 activation. Simultaneous deletion of ATFS-1 and GCN-2 compounds the developmental defects associated with mitochondrial stress, while stressed animals lacking GCN-2 display a greater dependence on ATFS-1 and stronger induction of mitochondrial chaperone genes. These findings are consistent with translational control and stress-dependent chaperone induction acting in complementary arms of the UPRmt.
| Defects in mitochondrial function are associated with numerous age-related diseases including cancer and Parkinson's. Mitochondrial function relies upon maintenance of the mitochondrial proteome, which is comprised of nuclear and mitochondrial-encoded proteins. Nuclear-encoded polypeptides are translated in the cytosol and must be transported into the mitochondrial matrix, where resident chaperones facilitate folding into their functional conformation. In order to protect against dysfunction arising from an accumulation of misfolded or unfolded mitochondrial proteins, cells employ mechanisms to maintain the folding environment. One such signaling pathway is mediated by the bZip transcription factor ATFS-1, which upregulates mitochondrial chaperones to accommodate an overwhelming misfolded protein load. Here, we describe a complementary pathway that couples the mitochondrial functional status with the rate of cytosolic protein synthesis to protect the organelle from incoming unfolded protein substrates during mitochondrial stress. This pathway is regulated by the cytosolic kinase GCN-2, which phosphorylates the translation initiation factor 2α (eIF2α) subunit to slow general translation. GCN-2 responds to ROS emitted from dysfunctional mitochondria to promote growth and extend lifespan during mitochondrial stress.
| Mitochondrial dysfunction and altered protein homeostasis are associated with numerous developmental and age-related diseases as well as the general process of aging [1]. The mitochondrial protein-folding environment is maintained by nuclear-encoded mitochondrial chaperones, which promote efficient protein folding, and proteases that degrade those proteins that fail to fold or oligomerize correctly [1], [2], [3]. Protein folding is compartmentalized in eukaryotic cells and facilitated by compartment-specific folding machinery in the cytosol, endoplasmic reticulum (ER) and mitochondria. As threats to protein homeostasis affect the folding compartments differently, each compartment has dedicated stress responses or unfolded protein response (UPR) signaling pathways to transcriptionally regulate organelle-specific molecular chaperones and reduce the protein-folding load on the resident protein folding machinery.
Dysfunction and accumulation of misfolded proteins in the ER triggers a multi-pronged unfolded protein response (UPRER) that combines the upregulation of molecular chaperones to accommodate the folding requirements in the organelle with a reduction of cytosolic translation and ER protein import [4]. Activation of the transmembrane kinase PERK phosphorylates the cytosolic translation initiation factor eIF2α, thus attenuating general mRNA translation and reducing the load of incoming unfolded polypeptides [5]. In a complementary branch of the UPRER, the transcription factor XBP-1 is activated and mediates the induction of ER-resident chaperones [6]. Thus, by coordinating signaling through parallel pathways, stress is relieved and organelle function restored. In contrast to these ER-protective mechanisms, signaling pathways that protect the mitochondrial protein-folding environment are only beginning to emerge.
Maintenance of mitochondrial metabolic function depends on the efficient assembly of the mitochondrial proteome, which is comprised of nuclear-encoded as well as mitochondrial-encoded polypeptides [7]. Those proteins encoded by the nucleus are translated in the cytosol and post-translationally imported into mitochondria in an unfolded or unstructured state where they interact with the network of mitochondria-resident molecular chaperones. Failure of mitochondrial proteins to properly fold or oligomerize can result in electron transport chain (ETC) defects and accumulation of ROS, which further impacts additional mitochondrial activities including metabolic function. In order to respond to mitochondrial-specific stresses caused by the accumulation of unfolded proteins, depletion of mtDNA, defects in respiration or altered ROS metabolism, mitochondria have evolved stress response pathways that upregulate mitochondrial molecular chaperones to restore organelle homeostasis [8], [9], [10]. One of these pathways, termed the mitochondrial unfolded protein response (UPRmt), couples the status of the mitochondrial protein-folding environment to the transcription of mitochondrial chaperone genes [8], [9]. The complement of nuclear-encoded mitochondrial chaperones, such as mtHsp70 and HSP-60, assist in import, folding, and assembly of multi-protein complexes in the matrix and on the matrix side of the inner mitochondrial membrane [2]. Increased levels of mitochondrial dysfunction perturb the balance between chaperones and their client proteins, leading to activation of the UPRmt and upregulation of mitochondrial chaperone genes to re-establish homeostasis [8], [9], [10].
Our previous genetic studies in C. elegans have identified several proteins required for signaling the response including the mitochondrial inner membrane-localized peptide transporter HAF-1 and the bZip transcription factor ZC376.7 [11], which was recently renamed ATFS-1 (Activating Transcription Factor associated with Stress-1). Mitochondrial dysfunction triggers the HAF-1-dependent nuclear accumulation of ATFS-1, resulting in the upregulation of mitochondrial chaperone genes including HSP-60 and mtHsp70. Activation of this pathway occurs in response to elevated levels of mitochondrial stress, which can be the result of accumulation of unfolded proteins beyond the capacity of mitochondrial molecular chaperones [8] as well as increased levels of oxidative stress [9], respiratory chain dysfunction and by mtDNA depletion [10]. Thus, this mitochondrial stress response pathway, although termed a UPR because of conceptual similarities with the XBP-1 branch of the UPRER, responds to diverse insults to mitochondrial function.
In addition to chaperone induction, the UPRER also mediates the attenuation of cytosolic translation to protect the ER during stress. Similarly, inhibition of cytosolic translation has been suggested to promote mitochondrial function in yeast and Drosophila models of mitochondrial stress, although a potential regulatory mechanism(s) remained to be elucidated [12], [13]. Cytosolic translation attenuation via PERK-1-mediated eIF2α phosphorylation promotes ER function during stress by reducing the client load on ER-resident chaperones [5], [14]. Additionally, in C. elegans genetic manipulations that reduce cytosolic translation rates provide resistance to numerous stresses including heat shock and also extend lifespan [15], [16], [17]. Several signaling pathways are known to regulate translation rates in eukaryotic cells including TOR-regulated phosphorylation of S6 kinase and 4E-BP [16], [17], [18], however a mechanism to couple cytosolic translation rates to mitochondrial function has not been demonstrated.
Phosphorylation of eIF2α by four dedicated kinases (GCN2, PERK, HRI and PKR) serves to attenuate cytosolic translation in response to a variety of cellular stresses including starvation, oxidative stress, viral infection and unfolded protein stress in the ER [19], [20], . In yeast and mammals, GCN-2 phosphorylates eIF2α in response to conditions of low free amino acid levels and oxidative stress [22], [23]. Here we describe experiments demonstrating that in C. elegans, translation attenuation via GCN-2-dependent eIF2α phosphorylation acts in a responsive and adaptive protective pathway during mitochondrial stress to promote mitochondrial function. Phosphorylation levels of eIF2α are increased during mitochondrial stress, which requires ROS generated from dysfunctional mitochondria. Our data demonstrate that GCN-2-dependent translational control acts in a mitochondrial protective signaling pathway complementary to the regulation of mitochondrial chaperone gene expression mediated by HAF-1 and ATFS-1.
We have previously described a mitochondrial stress response pathway that upregulates mitochondrial chaperone genes in response to multiple perturbations in mitochondrial function [8], [11], [24]. RNAi experiments indicated a requirement for the bZip transcription factor ATFS-1 in mitochondrial chaperone induction as demonstrated by quantitative PCR experiments as well as using reporter strains where the hsp-60 promoter regulates expression of GFP (hsp-60pr::gfp) [11]. In order to corroborate the requirement for ATFS-1, we obtained the atfs-1(tm4525) deletion strain which lacks 432 base pairs and most of exons 2–4, and crossed it into the reporter strain. Unlike wild-type worms, atfs-1(tm4525) animals were unable to induce hsp-60pr::gfp when raised on spg-7(RNAi), a mitochondrial protease required for ETC quality control and mitochondrial ribosome biogenesis [25]. These results confirm the requirement for ATFS-1 in stress-induced mitochondrial chaperone gene induction (Figure 1A).
We next investigated ATFS-1-dependent hsp-60pr::gfp activation in strains harboring the well-characterized clk-1(qm30) or isp-1(qm150) mutations [26], [27]. clk-1 encodes a mitochondrial protein required for ubiquinone synthesis [28], which acts as a lipid antioxidant throughout the cell and an electron transporter within the electron transport chain. isp-1 encodes an iron-sulfur component of complex III in the ETC. As both mutations affect respiration and display impaired development [22], [23], we hypothesized that they would cause activation of the UPRmt. Indeed, hsp-60pr::gfp expression was consistently elevated in both strains consistent with the presence of mitochondrial stress. The isp-1(qm150) mutation caused considerably stronger hsp-60pr::gfp induction suggestive of a larger impact on mitochondrial function [29] (Figure 1B). Chaperone induction in both mutants required ATFS-1 as animals raised on atfs-1(RNAi) were unable to induce expression of hsp-60pr::gfp (data not shown).
To determine if the ATFS-1-dependent regulation of mitochondrial chaperone genes has a protective role during mitochondrial stress we examined the effect of atfs-1(RNAi) on the development of wild-type and mitochondrial stressed worms. As previously demonstrated, both clk-1(qm30) and isp-1(qm150) worms developed considerably slower than wild-type animals [22], [28]. Consistent with ATFS-1 being a stress responsive transcription factor, wild-type worms fed atfs-1(RNAi) developed at similar rates to wild-type animals (data not shown). However, feeding clk-1(qm30) and isp-1(qm150) worms atfs-1(RNAi) dramatically impaired their developmental rates (Figure 1C and Figure S1), indicating a requirement for ATFS-1 during development in the presence of mitochondrial stress.
In addition to ATFS-1-regulated mitochondrial chaperone expression, we sought to identify additional components that promote mitochondrial protein homeostasis by acting in complementary pathways. To identify signaling pathways that act in parallel to ATFS-1 we generated and screened an RNAi sub-library consisting of all C. elegans kinases and phosphatases [30]. We took advantage of hsp-60pr::gfp activation as a sensitive readout for the status of mitochondrial function to identify signaling components that promoted or impaired mitochondrial protein homeostasis. The clk-1(qm30) strain was chosen for the RNAi screen as it displayed mild hsp-60pr::gfp induction, potentially allowing for the identification of candidates whose knockdown by RNAi either decreased or further increased hsp-60pr::gfp expression (Figure 1B). We hypothesized that RNAi knockdown of candidates that act in a complementary protective signaling pathway would show enhanced hsp-60pr::gfp activation in the presence of stress because of an increased substrate load on the mitochondrial protein folding machinery. Alternatively, if components exist whose knockdown somehow enhances the protein-folding capacity, then those RNAi may suppress hsp-60pr::gfp activation in the clk-1(qm30) background.
Interestingly, RNAi of several kinases required for protein synthesis reduced hsp-60pr::gfp expression in the clk-1(qm30) background (Figure 2A), which was also confirmed using the isp-1(qm150) strain (Figure S2A), suggesting that knockdown of these components protected the mitochondrial folding environment. The reduced hsp-60pr::gfp expression was not simply due to a reduction in translation as separate GFP reporters under the myo-3, ges-1 or ER stress-inducible hsp-4 promoters were unaffected by the RNAi candidates (data not shown). These findings are consistent with previous experiments demonstrating that translation attenuation is protective against mitochondrial stress in yeast and Drosophila [12], [13]. Similarly, reduced translation has been associated with longevity and stress resistance in C. elegans. For example, knockdown of the C. elegans target of rapamycin ortholog (TOR), CeTor, which regulates mRNA translation in response to nutrient cues [31], or knockdown of rsks-1, the ribosomal S6 kinase, slows development and extends lifespan in C. elegans [16]. Because the long-lived clk-1(qm30) mitochondrial mutants have increased levels of mitochondrial stress and the stress responsive hsp-60pr::gfp reporter was specifically reduced by CeTor and rsks-1(RNAi), these findings support the hypothesis that reduced translation is beneficial to mitochondrial protein homeostasis. However, because the TOR-signaling pathway impacts many biological processes in addition to translation, other possibilities exist. Because all components identified in our RNAi screen affect protein synthesis, we sought to further characterize the role of translation attenuation in maintaining the mitochondrial protein-folding environment.
In addition to CeTor, rsks-1 and cel-1, we identified components that are known to regulate translation initiation by modulating the phosphorylation status of the translation initiation factor eIF2α. RNAi-knockdown of the eIF2α kinase General Control Non-derepressible-2 (GCN-2) further increased hsp-60pr::gfp expression in clk-1(qm30) animals, suggesting a role for GCN-2 in promoting mitochondrial protein homeostasis or function (Figure 2A and 2B). The effect of gcn-2(RNAi) on hsp-60pr::gfp expression was not due to direct effects on GFP translation as gcn-2(RNAi) did not cause induction of the ER stress reporter hsp-4pr::gfp (Figure 3A) suggesting a specific role for GCN-2 in promoting mitochondrial protein homeostasis. In unstressed animals, gcn-2(RNAi) did not effect hsp-60pr::gfp expression, suggesting its primary role is during stress (Figure 2B). Contrary to gcn-2(RNAi), our RNAi screen identified gsp-1(RNAi), which reduced hsp-60pr::gfp expression in both the clk-1(qm30) and isp-1(qm150) strains (Figure 2A, 2B and Figure S2B). GSP-1 encodes a protein phosphatase (PP1) required for numerous cellular dephosphorylation events [32], [33] and is homologous to the yeast phosphatase required for eIF2α dephosphorylation [34].
To determine if GCN-2 and GSP-1 regulate eIF2α phosphorylation in C. elegans, we examined the phosphorylation status of eIF2α in whole worm lysates. We utilized an antibody that specifically recognizes the highly conserved serine that is phosphorylated by the repertoire of eIF2α kinases (S51 in mammals and S49 in C. elegans (Figure 3B)). Consistent with previous reports, we detected phosphorylated eIF2α in otherwise unstressed worms (Figure 3C and 3D) [35], [36], [37], which was reduced when the lysate was incubated with calf-intestine phosphatase (CIP), confirming the specificity of the antibody for the phosphorylated form of eIF2α (Figure 3C). Furthermore, in a deletion mutant lacking 1482 bases of gcn-2 (gcn-2(ok871)), the level of steady-state phospho-eIF2α was reduced relative to wild-type worms (Figure 3D). In C. elegans, the only other known eIF2α kinase is PEK-1 (homologous to mammalian PERK [38]). Indeed, phospho-eIF2α was further reduced relative to levels of total eIF2α protein and mRNA in a strain lacking both kinases (Figure 3D and Figure S3B), further supporting the specificity of the phospho-eIF2α antibody and demonstrating the contribution of both kinases to steady state levels of eIF2α phosphorylation.
In contrast to inhibition of GCN-2 and PEK-1, GSP-1 knockdown resulted in increased levels of phospho-eIF2α consistent with it acting as a constitutive eIF2α phosphatase (Figure 3E). In either the gcn-2(ok871) or pek-1(zcdf2) deletion strains fed gsp-1(RNAi) there was still an increase in steady state levels of eIF2α phosphorylation likely reflecting the ability of both kinases to constitutively phosphorylate eIF2α in the absence of exogenous stress (Figure 3E). As increased phospho-eIF2α results in reduced cytosolic translation [5], [39], these data suggest that gsp-1(RNAi) reduces hsp-60pr::gfp induction through attenuation of cytosolic translation, thus reducing the load on the mitochondrial protein folding machinery similar to eIF2α phosphorylation and translation attenuation in the UPRER [5], [6].
The data presented above suggest that GCN-2 activity promotes mitochondrial protein folding during mitochondrial stress. Therefore, we hypothesized that eIF2α phosphorylation would increase in a GCN-2-dependent manner in response to mitochondrial dysfunction. Indeed, phospho-eIF2α levels were increased relative to total eIF2α protein levels in the clk-1(qm30)mutant, which was absent in the gcn-2(ok871) mutant strain (Figure 4A). In contrast, gsp-1(RNAi) caused a further increase in phospho-eIF2α levels (Figure 4A). A similar result was observed in the isp-1(qm150) mutant, supporting the role of GCN-2 in eIF2α phosphorylation in response to stress (Figure 4B). As gcn-2(RNAi) perturbs the mitochondrial protein folding environment and GSP-1 knockdown promotes mitochondrial protein homeostasis as indicated by reduced hsp-60pr::gfp expression (Figure 2B and Figure S2B), these data suggest a correlation between an increase in phospho-eIF2α and a more favorable mitochondrial protein-folding environment.
It should be noted that deletion or knockdown of the other C. elegans eIF2α kinase PEK-1 had no obvious effect on hsp-60pr::gfp induction during mitochondrial stress (data not shown). Furthermore, the increase in eIF2α phosphorylation observed in the clk-1(qm30) animals was not dependent on pek-1 indicating GCN-2 is the primary eIF2α kinase involved in maintaining mitochondrial protein homeostasis (Figure 4C). In pek-1 deletion worms, steady state levels of phospho-eIF2α were reduced (Figure 3E), however these animals still induce eIF2α phosphorylation in response to mitochondrial dysfunction supporting the specific role for GCN-2 during mitochondrial stress (Figure 4C). A similar relationship has been described with PEK-1 and the induction of ER chaperones during ER stress. PEK-1 is specifically activated during ER stress and animals lacking PEK-1 display stronger induction of ER chaperone genes including hsp-4 during ER stress [38]. The UPRER reporter hsp-4pr::gfp is induced during mild heat stress, a condition known to activate the UPRER but not the UPRmt [9], [11]. Incubation of hsp-4pr::gfp animals at 30°C for 3 hours mildly induced GFP expression (Figure 3A and Figure S3A). However, worms raised on pek-1(RNAi) displayed a much stronger induction of the UPRER reporter upon heat exposure consistent with PEK-1 activity protecting ER protein homeostasis [19], [38]. Unlike pek-1(RNAi), gcn-2(RNAi) had no impact on hsp-4pr::gfp during heat stress. These results indicate that the effect of gcn-2(RNAi) on hsp-60pr::gfp induction was not due to dysregulation of global translation further supporting a mitochondrial stress-specific role for GCN-2.
As our data indicated that GCN-2 phosphorylates eIF2α in response to mitochondrial stress, we sought to determine the role of GCN-2 in development and mitochondrial maintenance during mitochondrial stress. gcn-2 deletion or RNAi had no observable effect on worm development in the absence of stress (Figure S4). However, in the presence of mitochondrial stress caused by either the isp-1(qm150) or clk-1(qm30) mutations, gcn-2 deletion significantly slowed development (Figure 5A and 5B). Furthermore, exposure to the NADH ubiquinone oxidoreductase (complex I) inhibitor rotenone or spg-7(RNAi) also significantly delayed development of gcn-2(ok871) worms relative to wild-type worms (Figure 5C and data not shown) indicating a protective role for GCN-2 during mitochondrial stress.
To further assess the role of GCN-2 in maintaining mitochondrial function during mitochondrial stress we examined the effect of gcn-2 deletion on oxygen consumption in wild-type as well as mitochondrial stressed worms. We observed no difference in the rates of oxygen consumption between wild-type worms and those lacking gcn-2 (Figure 5D), consistent with the gcn-2(ok871) deletion having no effect on worm development (Figure S4). clk-1(qm30) worms displayed a slight reduction in oxygen consumption when compared to wild-type worms consistent with mild mitochondrial dysfunction (Figure 5E) [40]. Impressively, clk-1(qm30) worms lacking gcn-2 had a much lower rate of oxygen consumption than worms harboring either the gcn-2-deletion or clk-1(qm30) alone (Figure 5E), supporting a role for GCN-2 in promoting mitochondrial function during mitochondrial stress.
Elevated ROS produced by dysfunctional mitochondria can damage proteins through the formation of irreversible carobonyl modifications on lysine, cysteine, proline and threonine residues [41], [42], [43]. In order to examine levels of oxidative damage in mitochondrial stressed worms, we visualized the accumulation of carbonylated proteins using the Oxyblot system [44]. Consistent with the clk-1(qm30) mutation causing mitochondrial dysfunction, significantly more carbonylated material was detected in lysates from clk-1(qm30) worms than lysates from wild-type worms (Figure 5F). clk-1(qm30);gcn-2(ok871) worms displayed even more oxidative damage than worms harboring clk-1(qm30) alone. Because oxidative damage can perturb protein folding, these data support a role for GCN-2 in protecting the folding environment as well as mitochondrial function.
To further assess the contribution of GCN-2 in maintaining mitochondrial protein homeostasis we targeted GFP to the mitochondrial matrix via the strong muscle-specific myosin promoter (myo-3). High-level expression of mitochondria-targeted GFP challenges the organelle's protein folding environment by increasing the load of unfolded proteins [24], [45]. While wild-type worms were able to accommodate the increased folding load and maintain mitochondrial morphology, myo-3pr::gfpmt worms raised on gcn-2(RNAi) displayed severely perturbed mitochondrial morphology consistent with a loss of protein homeostasis and mitochondrial function [24], [46] (Figure 5G). Furthermore, in the absence of GCN-2, developmental rates (data not shown) and muscle cell function were severely reduced as determined by a motility or thrashing assay (Figure 5H). Together these data indicate that GCN-2 protects mitochondrial function during increased load of mitochondrial unfolded proteins.
clk-1(qm30) and isp-1(qm150) animals, which activate ATFS-1-dependent hsp-60pr::gfp expression and GCN-2-dependent eIF2α phosphorylation, are among the numerous C. elegans mitochondrial mutants that exhibit lifespan extension [47], [48], [49]. It was recently reported that ubl-5, a small ubiquitin-like protein required for UPRmt signaling [45], was required for lifespan extension in several mitochondrial mutants highlighting the importance of maintaining mitochondrial protein homeostasis [50]. Consistent with these studies, knockdown of ubl-5 prevented hsp-60pr::gfp induction in the long-lived clk-1(qm30) worms (Figure S5) similar to atfs-1(RNAi).
Additionally, cytosolic translation attenuation also contributes to longevity in several animal models [15], [16], [17]. As GCN-2 slows cytosolic translation [51] in response to mitochondrial dysfunction, we examined the role of GCN-2 in lifespan extension associated with mitochondrial dysfunction. Interestingly, GCN-2 knockdown in clk-1(qm30) animals reduced their lifespan to that of wild-type worms (Figure 6A) consistent with a role for GCN-2 in lifespan extension associated with mitochondrial dysfunction. gcn-2(RNAi) was not generally toxic, as it did not affect lifespan or development in the absence of stress (Figure 6B and Figure S4). gcn-2(RNAi) also shortened the lifespan of isp-1(qm150) animals, but because the animals were very sick with considerable developmental defects, we were unable to determine a role for GCN-2 in longevity of these animals (data not shown). These data are consistent with GCN-2 and increased eIF2α phosphorylation contributing to the lifespan extension observed in mitochondrial mutants and further emphasizes the importance of protein homeostasis in aging.
Because gsp-1(RNAi) caused an increase in eIF2α phosphorylation in the absence or presence of mitochondrial stress (Figure 3E, Figure 4A and 4B), we hypothesized gsp-1(RNAi) would promote lifespan extension. However, the lifespan of wild-type or clk-1(qm30) worms on gsp-1(RNAi) were severely shortened (Figure S6). Interpretation of this result is complicated by the pleitropic, non-specific effects of GSP-1 knockdown. gsp-1(RNAi) also prevents C. elegans germline formation (data not shown) and is required for a variety of cellular processes including mitosis [33]. Therefore, we were unable to determine if increased eIF2α phosphorylation was sufficient to extend lifespan.
We next sought to determine how phospho-eIF2α status is linked to mitochondrial dysfunction. While the most well-studied mechanism of GCN-2 activation is through starvation or amino acid depletion, hydrogen peroxide exposure also stimulates GCN-2 activity through a mechanism that requires the tRNA synthetase domain [52], [53]. Because clk-1(qm30) and isp-1(qm150) worms produce increased levels of ROS (Figure 5F) that are also required for their extended longevity [22], [23], we hypothesized that ROS generated from dysfunctional mitochondria act as an upstream signaling molecule coupling mitochondrial dysfunction to GCN-2 activation. If ROS are required for the observed increase in eIF2α phosphorylation during mitochondrial stress, then treatment with ROS scavengers would phenocopy GCN-2 inhibition with respect to hsp-60pr::gfp activation and the reduced accumulation of phospho-eIF2α in the presence of mitochondrial stress. Impressively, incubation of clk-1(qm30) animals with the ROS scavenger ascorbate resulted in increased hsp-60pr::gfp activation, similar to gcn-2(RNAi) (Figure 7A and Figure 2B). Ascorbate had no effect on the induction of hsp-60pr::gfp in unstressed animals (data not shown) as observed with gcn-2(RNAi) (Figure 2B). We next examined the impact of ascorbate on eIF2α phosphorylation in clk-1(qm30) and isp-1(qm150) animals. Ascorbate treatment, like GCN-2 inhibition, caused a reduction of eIF2α phosphorylation in both mutants supporting a role for ROS in GCN-2 signaling during mitochondrial stress (Figure 7B).
Our data indicate a requirement for ROS in GCN-2-dependent eIF2α phosphorylation observed in response to mitochondrial dysfunction in clk-1(qm30) and isp-1(qm150) mutants. In addition to these two mutants which generate ROS, the herbicide paraquat is known to generate excessive ROS and extend C. elegans lifespan [22], [23]. Interestingly, similar to the clk-1(qm30) and isp-1(qm150) mutants, exposure of wild-type worms to paraquat increased eIF2α phosphorylation in a GCN-2-dependent manner (Figure 7C). In sum, these data support a protective upstream signaling role for mitochondria-generated ROS in GCN-2 activation during mitochondrial stress.
The above data are consistent with GCN-2-dependent eIF2α phosphorylation and translation attenuation playing a protective role in maintaining mitochondrial function similar to the protection provided by the induced mitochondrial chaperone expression regulated by HAF-1 and ATFS-1 [11]. Therefore, we sought to determine the potential interaction or relationship between GCN-2 and ATFS-1/HAF-1. If they act in complementary pathways, we hypothesized that loss-of-function of both should be more detrimental than loss of either individual pathway. Inhibition of one pathway would cause more stress placing additional burden on the other pathway to maintain the mitochondrial protein-folding environment resulting in further activation of the complementary pathway. As indicated in Figure 2B, in the presence of stress, GCN-2 inhibition results in further activation of hsp-60pr::gfp expression. Similarly, reducing eIF2α phosphorylation by inhibiting ROS accumulation resulted in increased activation of hsp-60pr::gfp activation by ATFS-1 (Figure 7A). To determine if inhibition of mitochondrial chaperone induction during stress caused a further upregulation of the GCN-2 pathway and an increase in eIF2α phosphorylation, we examined phospo-eIF2α levels in clk-1(qm30) animals lacking HAF-1 or ATFS-1. clk-1(qm30) animals displayed an increase in eIF2α phosphorylation which was further increased in combination with the haf-1(ok705) deletion or when fed atfs-1(RNAi) (Figure 8A), consistent with GCN-2 acting in a separate and complementary pathway to that of ATFS-1 and HAF-1.
We next investigated the potential synthetic interaction between GCN-2 and ATFS-1 during worm development. Either individual mutation had no obvious growth or developmental defect in the absence of stress (Figure 8B and Figure S4). However, gcn-2(ok871);atfs-1(tm4525) animals or the gcn-2(ok871) strain fed atfs-1(RNAi) developed somewhat slower in the absence of exogenous stress (Figure 8B and 8C). These data suggest the presence of low levels of mitochondrial stress during development that required the activity of either GCN-2 or ATFS-1.
In the presence of stress caused by spg-7(RNAi) or the clk-1(qm30) mutation, development of worms lacking gcn-2 and atfs-1 was severely compromised. When raised on spg-7(RNAi), most worms arrested at the L1 or L2 larval stage and no animals were able to reach adulthood (Figure 8C and Figure S7A). Furthermore, clk-1(qm30) (Figure 8D and Figure S7B) or isp-1(qm150) (data not shown) animals lacking both GCN-2 and ATFS-1 developed more slowly than worms lacking either individual gene. No synthetic interactions were observed between ATFS-1 and PEK-1 as atfs-1(tm4525) animals raised on pek-1(RNAi) developed at similar rates to atfs-1(tm4525) animals raised on a control RNAi in the absence or presence of stress (data not shown). Despite the developmental defect observed in animals lacking both ATFS-1 and GCN-2 (Figure 8B and 8C), animals lacking both genes had similar lifespans to those of wild-type worms (Figure 8E) suggesting the primary role for each pathway in the absence of exogenous stress is during development, when the majority of mitochondrial biogenesis occurs [54]. Together, these results support a model in which ATFS-1 and GCN-2 act in different yet complementary mitochondrial stress response pathways to regulate mitochondrial chaperone expression and cytosolic translation to protect mitochondrial function (Figure 8F).
The experiments described here implicate the eIF2α kinase GCN-2 in the maintenance of mitochondrial function and protein homeostasis. Development of worms lacking GCN-2 was impaired in the presence of mitochondrial stress which caused further induction of ATFS-1-dependent mitochondrial chaperone genes consistent with perturbed mitochondrial protein homeostasis. Furthermore, simultaneous deletion or knockdown of GCN-2 and ATFS-1 has a negative synergistic effect on animal development suggesting that GCN-2-dependent translation attenuation and ATFS-1-dependent mitochondrial chaperone gene induction act in parallel pathways to maintain mitochondrial protein homeostasis. Additionally, GCN-2 was required for development and lifespan extension in the presence of mitochondrial stress suggesting it is responsive to and protective against mitochondrial dysfunction. gcn-2(RNAi) or deletion inhibited eIF2α phosphorylation during mitochondrial stress. These results, along with recent experiments in yeast and flies [12], [13], support our conclusion that attenuation of cytosolic translation is protective during mitochondrial dysfunction.
Our results demonstrate that ROS generated from stressed or dysfunctional mitochondria [22], [23] are required for GCN-2-dependent eIF2α phosphorylation (Figure 7B and 7C). Furthermore, treatment with ROS inhibitors phenocopied gcn-2(RNAi) further exacerbating mitochondrial chaperone induction in the presence of stress (Figure 7A) suggesting ROS are required for GCN-2-dependent eIF2α phosphorylation but not ATFS-1-mediated induction of mitochondrial chaperone gene transcription. While these data support a model in which ROS act as an upstream signaling molecule, the mechanism of GCN-2 activation remains unclear. GCN-2 activation through amino acid depletion is thoroughly characterized, and requires interaction between uncharged tRNA and the tRNA synthetase domain of GCN-2 [55]. GCN-2 activation by peroxide exposure is less well understood, however it also requires the tRNA binding domain [53], [56]. Interestingly, both increased ROS and alterations in amino acid levels are known to occur in clk-1(qm30) and isp-1(1qm150) mutant worms, suggesting that ROS could participate in GCN-2 activation either through a direct interaction with the GCN-2 tRNA synthetase domain or through effects on amino acid metabolism [22], [23], [57], [58], [59]. Regardless, our results suggest a protective role for mitochondrial-generated ROS by influencing eIF2α phosphorylation, consistent with recent data indicating that low levels of ROS participate in beneficial cyto-protective stress-signaling pathways [22], [23], [60].
Attenuation of cytosolic translation slows mitochondrial import, thus reducing the folding load on mitochondrial chaperones. However, continued translation of proteins encoded by the mitochondrial genome could become detrimental when the expression of cytosolic components required for ETC complex formation is reduced. Interestingly, mitochondrial translation is tightly linked to the accumulation of imported ETC subunits and complex assembly. In their absence, mitochondrial translation is also attenuated [61]. We hypothesize that translation attenuation in the cytosol slows mitochondrial protein import leading to translational repression within mitochondria, thus reducing the overall burden on the mitochondrial protein folding and complex assembly machinery.
Protection of mitochondrial protein homeostasis and function appears to be a novel role for GCN-2 in addition to its established role during starvation [62]. The GCN-2 signaling pathway is complementary to the signaling pathway that transcriptionally upregulates mitochondrial chaperone genes during stress, which requires the mitochondrial peptide transporter HAF-1 and transcription factor ATFS-1 (Figure 8F) [11]. This parallel relationship between a reduction in organelle protein folding load and the regulation of organelle-specific protein folding machinery is similar to mechanisms that regulate ER protein homeostasis, in which another eIF2α kinase, PEK-1 (PERK in mammals), responds directly to unfolded protein stress within the ER. PEK-1-mediated translation attenuation complements the IRE-1/XBP-1 branch of the UPRER, which regulates expression of ER chaperones and additional protein handling machinery [38].
In addition to a protective role during development, GCN-2 also contributes to the lifespan extension of clk-1(qm30) animals (Figure 6A). These mutants have disrupted mitochondrial function and elevated levels of mitochondrial chaperones (Figure 1B), consistent with a recent report that indicated a requirement for mitochondrial chaperone induction in the lifespan extension of several mitochondrial mutants [50]. Additionally, clk-1(qm30) animals display elevated levels of ROS that have also been shown to contribute to longevity [22], [23]. Our data support a model in which ROS and GCN-2 activate a pathway that contributes to lifespan extension, in parallel to the requirement for transcriptional induction of mitochondrial chaperone genes. The contribution of GCN-2 is most likely through cytosolic translation attenuation, which promotes stress resistance and extends lifespan in C. elegans [15], [16], [17].
An additional eIF2α-dependent protective activity not addressed here involves the preferential translation of mRNAs with small upstream open reading frames (uORFs). A number of uORF containing transcripts have been identified in S. cerevisiae [63] including the well-characterized transcription factor Gcn4 [64]. Homology searches did not reveal an obvious Gcn4 orthologue in C. elegans, and this avenue was not further pursued. Our RNAi screen identified components that when knocked down slow cytosolic translation as suppressors of hsp-60pr::gfp activation in stressed animals supporting a role for translation attenuation in promoting mitochondrial protein homeostasis.
Our finding that GCN-2-dependent eIF2α phosphorylation protects mitochondrial protein homeostasis raises the possibility that manipulation of phospho-eIF2α status may be a therapeutic entry point for the diverse number of degenerative diseases associated with mitochondrial dysfunction [65]. At least two strategies to accomplish this seem plausible: (1) caloric restriction to reduce cytosolic amino acid levels and activate GCN-2 to increase eIF2α phosphorylation independent of mitochondrial stress or (2) small molecule inhibition of stress-dependent eIF2α dephosphorylation to increase phospho-eIF2α levels through phosphatase inhibition [37]. It will be interesting to determine the viability of these possibilities in future studies.
Reporter strains hsp-60pr::gfp(zcIs9)V, myo-3pr::gfpmt(zcIs14) and hsp-4pr::gfp(zcIs4)V have been described previously [9], [24], [45]. Where indicated, the hsp-60pr::gfp(zcIs9)V transgene was crossed into individual mutant strains of interest, with the exception of atfs-1(tm4525)V, which was backcrossed with N2 three times prior to crossing into the hsp-60pr::gfp background. The clk-1(qm30), isp-1(qm150) and gcn-2(ok871) strains were obtained from the Caenorhabditis Genetics Center (Minneapolis, MN) and the atfs-1(tm4525) strain was obtained from the National BioResource Project (Tokyo, Japan). RNAi feeding experiments were performed as described [9] with constructs from the Ahringer and Vidal libraries [66], [67].
Worms were synchronized via bleaching and allowed to develop on the described RNAi plate or condition. For development in the presence of oxidative stress, rotenone was applied to vector(RNAi) plates and allowed to soak in prior to seeding eggs. At the time points indicated, the numbers of L1, L2, L3, L4, young adult (non-gravid) or gravid adult worms were counted on each plate and quantified as a percent of the total number of animals. For each plate, the worms in 6–8 individual fields of view were counted, and the total number combined.
For lifespan analysis, worms were synchronized as eggs and allowed to develop under the described condition for two days. At that point, 100 L4 animals were transferred to new RNAi plates and subsequently transferred to fresh plates every day for the next 5–6 days and every two days thereafter. The numbers of dead and censored worms were counted every second day for the duration of the assay [68]. Survival curves and statistical analysis were generated using Prism 5.0b software (Graphpad). Each experiment was repeated 3 times.
Worms were grown under the described conditions and collected at the L4 stage for analysis as previously described [9], [24]. Phospho-eIF2α antibody (#3597S) was obtained from Cell Signaling Technology (Danvers, MA) and observed using SuperSignal West Femto Maximum Sensitivity Substrate (Thermo Scientific, Rockville, IL). GFP and HDEL immunoblots were visualized using Odyssey Infrared Imager (Li-Cor Biosciences, Lincoln, NE). Total eIF2α was assessed as described [36]. Due to limited amounts of pan-eIF2αantibody, western analysis was only performed during select experiments to confirm specificity of the phospho-eIF2α antibody in wild-type lysates and during analysis phospho-eIF2α levels in clk-1(qm30) animals.
For the eIF2α dephosphorylation assay, 150 µg of worm lysate was treated with calf intestinal phosphatase (CIP) for 30 minutes at 30°C prior to SDS-PAGE analysis. For ascorbate treatment, synchronized worms were grown to adulthood in liquid medium, when 25 mM ascorbate was added for 16 hours prior to western analysis.
Total RNA was isolated using RNA STAT (Tel-Test Inc, Friendswood, TX). RNA samples were prepared from the described worms at the L4 stage. cDNA was then synthesized from total RNA using a iScript cDNA Synthesis Kit (Bio-Rad Laboratories, Hercules, CA). Following mRNA isolation and cDNA synthesis, qPCR was used to determine the expression level of eif2α using iQ sybr green supermix and MyiQ2 Two-Color Real-Time PCR Detection System (Bio-Rad). Actin was used as a control. Fold changes in gene expression were calculated using the comparative CtΔΔCt method.
Fluorescent photomicrographs were obtained using a Zeiss AxioCam MRm mounted on a Zeiss Imager.Z2 microscope or Zeiss M2 Bio stereo microscope (Carl Zeiss Imaging, Thornwood, NY).
Oxygen consumption assays were performed as described [11] using a Clark type electrode [40]. To determine the accumulation of oxidative protein modifications, synchronized wild-type, clk-1(qm30) and clk-1(qm30);gcn-2(ok871) worms were harvested once they reached the L4 stage. Worm lysates were separated by SDS-Page and treated according to the Oxyblot manufacturer (Millipore, Billerica, MA).
|
10.1371/journal.pntd.0002598 | Rapid, Serial, Non-invasive Assessment of Drug Efficacy in Mice with Autoluminescent Mycobacterium ulcerans Infection | Buruli ulcer (BU) caused by Mycobacterium ulcerans is the world's third most common mycobacterial infection. There is no vaccine against BU and surgery is needed for patients with large ulcers. Although recent experience indicates combination chemotherapy with streptomycin and rifampin improves cure rates, the utility of this regimen is limited by the 2-month duration of therapy, potential toxicity and required parenteral administration of streptomycin, and drug-drug interactions caused by rifampin. Discovery and development of drugs for BU is greatly hampered by the slow growth rate of M. ulcerans, requiring up to 3 months of incubation on solid media to produce colonies. Surrogate markers for evaluating antimicrobial activity in real-time which can be measured serially and non-invasively in infected footpads of live mice would accelerate pre-clinical evaluation of new drugs to treat BU. Previously, we developed bioluminescent M. ulcerans strains, demonstrating proof of concept for measuring luminescence as a surrogate marker for viable M. ulcerans in vitro and in vivo. However, the requirement of exogenous substrate limited the utility of such strains, especially for in vivo experiments.
For this study, we engineered M. ulcerans strains that express the entire luxCDABE operon and therefore are autoluminescent due to endogenous substrate production. The selected reporter strain displayed a growth rate and virulence similar to the wild-type parent strain and enabled rapid, real-time monitoring of in vitro and in vivo drug activity, including serial, non-invasive assessments in live mice, producing results which correlated closely with colony-forming unit (CFU) counts for a panel of drugs with various mechanisms of action.
Our results indicate that autoluminescent reporter strains of M. ulcerans are exceptional tools for pre-clinical evaluation of new drugs to treat BU due to their potential to drastically reduce the time, effort, animals, compound, and costs required to evaluate drug activity.
| The discovery and development of new drugs to improve the treatment of BU is greatly hampered by the slow growth rate of the organism, which requires up to 3 months to form countable colonies on solid media. Here, we engineered a virulent reporter strain of M. ulcerans with intrinsic bioluminescence to enable serial, real-time, non-invasive in vivo monitoring of viable bacterial counts and demonstrate its utility for high-throughput in vitro and in vivo screening of antibiotic efficacy using a panel of anti-mycobacterial drugs with various mechanisms of action. We show: 1) that a drug's in vitro activity against M. ulcerans is detectable in real time after as little as 2 days of exposure, and 2) that a drug's in vivo activity against M. ulcerans is detectable in as little as 1 week through rapid, serial, non-invasive assessment in live mice performed with a benchtop luminometer. This method promises dramatic reductions in time and effort as well as requirements for animals, compounds and other supplies. We believe such a strain is capable of transforming drug discovery and development efforts for BU.
| Buruli ulcer (BU) caused by Mycobacterium ulcerans is the world's third most common mycobacterial disease with cases occurring on every continent, especially in certain humid tropical regions of the world [1], [2]. M. ulcerans secretes an immunosuppressive macrolide toxin, termed mycolactone [3] whose biosynthetic enzymes are encoded on a giant plasmid [4]. Mycolactone is responsible for deep and necrotizing skin ulcers and occasionally bone lesions. No vaccine is available for BU [5]. Based on experiments in the mouse footpad model [6], [7] and subsequent clinical studies [8]–[12], a regimen of streptomycin (STR) plus rifampin (RIF) for 2 months is recommended for treatment of BU [13], although additional surgery including skin grafting may be necessary to repair large ulcers, contractures and deformities [10]. Though this 2-month drug regimen does reduce numbers of colony-forming units (CFU), lesion size, and mycolactone levels [14], [15], it has significant disadvantages, including STR's requirement for parenteral administration and potential for oto- and vestibulotoxicity, while RIF causes challenging drug-drug interactions with many drugs, including anti-mycobacterial and anti-retroviral agents. Therefore, entirely oral regimens and/or regimens capable of treating BU in 1 month or less are sought [16].
Efforts to evaluate the therapeutic potential of drugs for BU in the pre-clinical setting are hampered by the very slow growth of M. ulcerans which necessitates up to 3 months of incubation at 32°C for colonies to form on solid media. Light production by various luciferase enzymes has been used as a real-time biomarker of bacterial viability for high-throughput screening of antibiotics and drug susceptibility testing against mycobacteria [17]–[22]. The bacterial luciferases encoded by luxAB catalyze the oxidation of reduced flavin mononucleotide using a long-chain fatty aldehyde substrate, producing H2O and light (∼490 nm wavelength) in the process. Other genes of the lux operon (luxCDE) encode enzymes for the synthesis of the aldehyde substrate [23]. We have recently demonstrated that recombinant bioluminescent reporter strains of M. ulcerans expressing luxAB genes from Vibrio harveyi are useful for real-time evaluation of drug activity in vitro [21] and in vivo [22]. The endpoints used to measure efficacy in this mouse footpad model were relative light units (RLU) detected ex vivo in the footpad tissue or in vivo in live mice. However, these recombinant strains require the exogenous addition of substrate to produce light, making serial monitoring of the infection in live mice challenging. More recently, we created a virulent and stable autoluminescent Mycobacterium tuberculosis strain and demonstrated its utility for high-throughput in vitro and in vivo screening of antibiotic efficacy [24]. Using this reporter strain for serial, non-invasive monitoring of live mice in an acute infection model, drug activity against M. tuberculosis was evident within 3 days of treatment. In the present study, we created a virulent and stable autoluminescent M. ulcerans strain and evaluated its utility for real-time evaluation of antimicrobial effects in vitro and rapid, serial, non-invasive assessment of efficacy in two mouse footpad infection models.
All animal procedures were conducted according to relevant national and international guidelines. The study was conducted adhering to the Johns Hopkins University guidelines for animal husbandry and was approved by the Johns Hopkins Animal Care and Use Committee, protocol MO08M240. The Johns Hopkins program is in compliance with the Animal Welfare Act regulations and Public Health Service (PHS) Policy and also maintains accreditation of its program by the private Association for the Assessment and Accreditation of Laboratory Animal Care (AAALAC) International.
Rifampin (RIF), streptomycin (STR) and isoniazid (INH) were purchased from Sigma (St. Louis, MO). Kanamycin (KAN) and hygromycin (HYG) were purchased from Invitrogen (Carlsbad, CA) and Roche Diagnostics (Indianapolis, IN), respectively. Moxifloxacin (MXF), linezolid (LZD), clarithromycin (CLR), and bedaquiline (BDQ, formerly known as TMC207), were kindly provided by Bayer (Leverkusen, Germany), Pfizer (New York, NY), Abbott (Abbott Park, IL), and Tibotec (Beerse, Belgium), respectively.
STR, RIF and MXF were dissolved in distilled water, and CLR and LZD were dissolved in distilled water with 0.05% agarose for administration to mice. BDQ was formulated in an acidified cyclodextrin suspension as previously described [25]. All drugs were administered 5 days per week in 0.2 ml. RIF, MXF, CLR and LZD were administered by esophageal gavage. STR was administered by subcutaneous injection. The daily dosages are indicated for each experiment.
We previously constructed plasmids for expressing the luxCDABE operon from Photorhabdus luminescens [26] in mycobacteria, including the episomal (pTYOEH) and integrative (pTYOK, pTYZOK1/pTYZOK2 and pOAIK1/pOAIK2) constructs [27]. pTYOK contains only one copy of luxCDABE from P. luminescens under control of one hsp60 promoter. pTYZOK contains one copy of luxAB from V. harveyi and one copy of luxCDABE from P. luminescens, each under one hsp60 promoter. pOAIK1 and pOAIK2 contain an additional copy of luxB from V. harveyi downstream of luxCDABE [27], with the only difference between plasmids being the orientation of the fragment inserted. We transformed these plasmids into colony suspensions of M. ulcerans Mu1059 (WtMu, a clinical isolate from Ghana [28]) by electroporation, as described previously [21]. Two months later, colonies isolated on KAN- or HYG- containing selective 7H11 plates were individually tested for luminescence. Positive transformants for each plasmid were picked, homogenized in 2 ml Dubos broth with 0.07% Tween 80, and incubated at 32°C. When the OD600 nm reached more than 0.3, luminescence was detected using a TD-20/20 luminometer (Turner BioSystems), measuring light production over 3 sec. Strains were compared on the basis of relative light units (RLU) per ml of culture and the ratio of RLU to colony-forming units (CFU).
Serial dilutions of drug-containing solutions and M. ulcerans broth culture (OD600 of 0.3 to 0.7) were prepared as previously described [21]. RLU counts from the same batch of triplicate samples were measured daily over the first 7 days of exposure. The TD20/20 luminometer provides RLU values 1,000 times lower than those provided by the 20/20n luminometer. For consistency with prior studies, we defined 1 RLU as 1 unit in the 20/20n luminometer, equivalent to 0.001 units in the TD20/20 luminometer [21].
Colony suspensions of AlMu and WtMu were made by vortex-mixing 30 mg of colony material in 15 ml PBS. After allowing the clumps to settle, the resulting suspension was used to inject the right hind footpads of six-week old, female BALB/c mice. The inoculum volume was 0.03 ml, containing approximately 4 log10 CFU. The left hind footpads served as negative controls for observation of swelling. On the day after infection and at each time point thereafter, 5 mice per strain were sacrificed after non-invasive in vivo RLU measurement to determine RLU and CFU counts in the right hind footpad suspension. The mice were first anesthetized by isoflurane inhalation and the in vivo RLU count was determined non-invasively by placing the foot into the detection hole of the luminometer and measuring light production for 4 sec. To assure reproducibility, three separate measures were obtained for each mouse. The mice were then euthanized. The footpads were carefully cleaned with antimicrobial soap and water followed by an alcohol swab. After cleaning, the footpads were harvested using scalpel and forceps, minced in one drop of PBS using scissors and suspended in 1 ml PBS. After being shaken several times, the suspension was allowed to stand for 10 minutes to allow larger tissue debris to settle down. The resulting supernatant was used for ex vivo RLU detection, which was performed twice. Series of 100-fold dilutions of the suspension (undiluted, 10−2, 10−4) were plated in duplicate directly onto selective 7H11 plates (0.5 ml sample/plate).
Mice infected as described above were examined on a weekly basis to determine the footpad lesion index, which is defined as follows: 0 = normal footpad; 1 = non-inflammatory footpad swelling; 2 = inflammatory footpad swelling; 3 = inflammatory hindfoot swelling; 4 = inflammatory leg swelling; and 5 = death of the mouse [6]. For the purpose of assessing time-to-swelling, swelling was defined as a lesion index of grade 2 or higher. Treatment began 31 days after infection, when all footpads reached a swelling index ≥2 and continued for 3 or 4 weeks. At each time point, 5 mice from each group were sacrificed for RLU (in vivo and ex vivo) and CFU counts. Colonies of AlMu isolated from untreated control mice 8 weeks after infection were assessed for autoluminescence to confirm stability of the construct in vivo.
Mice were infected as described above. RLU counts determined non-invasively on the day after infection as described above were used to allocate mice to treatment groups (4 mice per group) with comparable distributions of RLU counts. Treatment began either 1 day or 11 days after infection and was administered for 2 weeks. The following 3 dose levels of each drug were used: 40, 10 and 2.5 mg/kg of body weight for RIF; 150, 75 and 19 mg/kg for STR; 100, 25 and 6.3 mg/kg for CLR; 25, 6.3 and 1.6 mg/kg for BDQ; and 200, 100 and 25 mg/kg for MXF. RLU counts from live mice were measured non-invasively in whole footpads of live mice twice weekly during treatment and on the footpad homogenate at the time of sacrifice, when CFU counts were also determined.
RLU and CFU counts were log10 transformed before analysis. Group means were compared by unpaired t test or by one-way analysis of variance (ANOVA) with Dunnett's posttest when multiple comparisons were made. An alpha value of 0.05 was used to determine statistical significance. Time-to-swelling curves were compared using the log rank test. As 5 pairwise comparisons were made in the time-to-swelling analysis, an alpha value of 0.01 was used to determine statistical significance. All statistical tests were performed with Prism 4 software (GraphPad Software, Inc., San Diego, CA).
The natural luxCDABE operon from P. luminescens [26] under control of the constitutive hsp60 promoter [24] was successfully cloned into four different vectors. However, only colonies transformed with the integrative plasmids pTYOK and pTYZOK2 exhibited strong luminescence that was visible to the naked eye and captured by a digital camera (Figure 1). Similar results were observed when engineering an autoluminescent M. tuberculosis strain [27]. In addition, the colonies grew as fast as their parent strain on agar plates. In vivo, the doubling time for WtMu was calculated at 4.93 days (95% confidence interval, 3.5–8.3) and for AlMu it was 3.63 days (95% confidence interval, 2.4–7.6). Colonies transformed with the other vectors produced weaker or no light, implying that the orientation of the luxCDABE operon is important and that increasing expression of luxAB with additional copies does not increase light production. The strain containing pTYOK was selected for further study because it did not contain additional copies of luxAB in addition to the luxCDABE operon.
Serial RLU counts were obtained daily, without adding exogenous substrate, from tubes of 7H9 broth containing AlMu in the presence of each of 6 anti-mycobacterial drugs with different mechanisms of action. The activity of each drug increased with concentration. BDQ had the greatest effect on RLU counts over the first 1–2 days, but the maximal effect observed over 7 days was similar for each drug. We defined the MIClux as the lowest concentration that prevented an increase in RLU after 7 days of incubation, i.e., RLUday7≤RLUday0. MIClux values against AlMu were similar to the MIC against the WtMu parent strain (Table 1) [21]. Even closer correspondence may have been achieved if doubling dilutions of drug had been used. Time-kill curves are shown in Figure 2. The MIClux could be discriminated from the drug-free control by the 2nd day of incubation. In each case, RLU remained detectable above background after 7 days of drug exposure, demonstrating a suitable dynamic range to the assay.
The mean RLU count of the AlMu suspension used to infect mice was 6.52 (SD 0.12) log10 RLU/ml. The light produced from footpads of live mice infected with AlMu increased continuously before reaching a plateau around 38 days post-infection, largely parallel to the change in RLU and CFU counts measured from the corresponding footpad suspension obtained at sacrifice (Figure 3). The RLU counts obtained non-invasively from live mice and those from footpad suspensions collected at necropsy in the AlMu groups were strongly correlated with the CFU counts, with correlation coefficients (r2) of 0.91 and 0.95, respectively. No significant light was detected from live mice or footpad suspensions from the WtMu infected group. The number of bacilli injected into the footpad was 3.76 (SD 0.11) for WtMu and 4.44 (SD 0.24) for AlMu. Swelling occurred 4–5 weeks post-infection for most mice (Figure S1). These results were comparable to a past experiment with similar infectious doses [22]. Thus, the AlMu strain appears to grow just as well and be just as virulent as the wild-type parent strain.
By day 31 after infection, the infected right hind footpads of all mice showed swelling. After treatment initiation with either STR+RIF or CLR alone, the impact of treatment on RLU counts could be observed within 2–4 days. Within one week of treatment, mice treated with STR+RIF experienced an approximately 1 log10 greater reduction in RLU counts compared to those receiving CLR alone (Figure 4). For AlMu-infected mice treated with STR+RIF or CLR alone, the effect on RLU counts measured non-invasively correlated well with the effects of treatment on RLU and CFU results determined ex vivo from footpad suspensions from the same animals, even though STR+RIF was highly bactericidal whereas CLR alone reduced the CFU counts by roughly 1 log10 CFU from the peak CFU count (Figure 4). For example, the correlation coefficients relating the non-invasive RLU counts to the ex vivo RLU and CFU counts were 0.99 and 0.98, respectively. Moreover, the responses of AlMu and WtMu to the regimens in this established disease model were very similar (Figure 4). We did not find any colony that lost the ability to produce light, suggesting that the AlMu construct is stable in animals even in the face of treatment with a highly bactericidal regimen.
The RLU counts from live mice treated with STR+RIF for only 3 weeks continued to decrease at week 4 and week 5 despite no treatment (Figure S2), which indicated a strong post antibiotic effect. In the case of CLR, the RLU counts remained largely stable for 2 weeks after treatment, suggesting the post-antibiotic effect was not as great. The mean log10 CFU count per footpad from AlMu-infected mice treated for 4 weeks with RIF+STR or CLR alone were 2.59 (SD 0.82) and 3.72 (SD 1.09), respectively. In support of the observed post-antibiotic effects, the mean CFU counts from AlMu-infected mice treated for 3 weeks with RIF+STR or CLR alone and sacrificed 2 weeks later were 2.11 (SD 0.53) and 5.25 (SD 0.32) (Figure S2).
Using a preventive model of testing drug activity by initiating treatment 11 days after infection, before the onset of footpad swelling, we observed dose-dependent activity of drugs with differing mechanisms of action, as measured by RLU obtained non-invasively (Figure 5A). The AlMu strain has a KAN resistance marker; KAN was used as a negative control and showed no activity at 150 mg/kg. RIF and STR showed bactericidal activity at 40 and 150 mg/kg, respectively. However, RIF at 10 mg/kg and STR at 75 mg/kg displayed only bacteriostatic activity and no activity at lower doses. Similarly, MXF at 200 mg/kg and BDQ at 25 mg/kg only had bacteriostatic activity and weaker activity at lower doses. CLR at 100 mg/kg had bacteriostatic activity in this experiment but none at lower doses. Drug activities can be observed within one week of treatment initiation in live mice (Figure 5A). Similar results were obtained using RLU and CFU counts from footpad suspensions (Figure 5B).
We also attempted to examine drug activity more rapidly by initiating treatment the day after infection. Differentiation of bactericidal and bacteriostatic activity in this system could be assessed after 10–14 days, but not after 7 days, for STR and CLR, whereas high-dose RIF activity could be discerned in the first week whether by RLU from the footpads of live mice or by RLU and CFU from footpad suspensions (Figure 6).
M. ulcerans requires up to 3 months of incubation at 32°C to form colonies on solid media, hampering the pre-clinical evaluation of new drugs and drug regimens to improve BU treatment. We have engineered an autoluminescent M. ulcerans strain and shown the advantages of using it as a real-time surrogate marker for viable bacilli to overcome this impediment. We have improved upon our previous studies [21], [22] in this regard by developing a truly non-invasive method for more rapid (e.g., approximately 60 mice can be assessed per hour by one person) and simple detection of light emitted from viable M. ulcerans infecting mouse footpads without the requirement for exogenous administration of substrate and using serial, real-time in vivo monitoring of M. ulcerans infection to evaluate the response to treatment.
As reported previously with M. tuberculosis [27], creation of an autoluminescent M. ulcerans strain could not be achieved using an extra-chromosomal vector but was successful using an integrative plasmid, indicating that M. tuberculosis and M. ulcerans have similar regulatory mechanisms for producing light with luxCDABE and maintaining extra-chromosomal plasmids. Growth, virulence, and drug susceptibility (with the necessary exception of KAN due to the inserted resistance marker) of AlMu are equivalent to its parental strain.
Our in vitro test system, using only 200 µl of the AlMu strain in broth, can discern the bacteriostatic and bactericidal activity of a drug within 2–3 days of exposure, indicating its potential for high-throughput screening. It is possible to obtain the MIClux within 1 week and, through the generation of dynamic time-kill curves, it may be possible to obtain useful information on drug pharmacodynamics, such as concentration- or time-dependence [29]. None of the tested drugs was able to reduce RLU below the limit of detection (2.7 log10/ml) in 7 days, indicating a suitable dynamic range for the assay. In addition, as also shown with autoluminescent M. tuberculosis as well as recombinant M. ulcerans expressing luxAB alone, the RNA- and protein-dependent generation of light production is not affected by antibiotics inhibiting transcription or translation out of proportion to the effects on CFU counts. Interestingly, BDQ, which inhibits ATP synthesis, did reduce RLU counts more sharply than other drugs over the first 2 days of exposure. AlMu displayed the same drug susceptibility as its parent strain. Taken together, these results indicate that AlMu is a suitable reporter strain for testing drug activity in vitro. Although the construction and evaluation of only a single autoluminescent M. ulcerans strain is a clear limitation of the present study, the limited published data regarding in vitro drug susceptibility among M. ulcerans strains do not indicate marked strain-to-strain differences [30]–[32].
The AlMu strain showed the same growth rate, virulence, and drug susceptibility in mice as its parent strain, indicating its utility as a reporter strain for in vivo drug efficacy studies. In these studies, we determined that a drug could be shown to be inactive, bacteriostatic, or bactericidal as early as two weeks after infection and treatment initiation in a preventive model. Treatment time can be further reduced to one week if treatment initiation is delayed to two weeks after infection, allowing for the use of a smaller amount of compound. If there is activity in the preventive model, testing in the established disease model may more closely mimic clinical presentations with swollen tissue and inflammation. In this model, in which the bacterial burden is more-or-less stable in the absence of treatment, RLU obtained non-invasively from live mice indicated that the STR+RIF regimen is bactericidal but the impact of the bacteriostatic CLR monotherapy regimen was more difficult to discern. With the established disease model, one can also keep mice for an additional 2–3 weeks to see if treatment has succeeded in killing bacteria and preventing relapse. Finally, it is likely that this autoluminescent strain could also be helpful in evaluating vaccine efficacy by monitoring the same mice over time after challenge.
|
10.1371/journal.pntd.0001338 | Gene Regulation in Giardia lambia Involves a Putative MicroRNA Derived from a Small Nucleolar RNA | Two core microRNA (miRNA) pathway proteins, Dicer and Argonaute, are found in Giardia lamblia, a deeply branching parasitic protozoan. There are, however, no apparent homologues of Drosha or Exportin5 in the genome. Here, we report a 26 nucleotide (nt) RNA derived from a 106 nt Box C/D snoRNA, GlsR2. This small RNA, designated miR5, localizes to the 3′ end of GlsR2 and has a 75 nt hairpin precursor. GlsR2 is processed by the Dicer from Giardia (GlDcr) and generated miR5. Immunoprecipitation of the Argonaute from Giardia (GlAgo) brought down miR5. When a Renilla Luciferase transcript with a 26 nt miR5 antisense sequence at the 3′-untranslated region (3′ UTR) was introduced into Giardia trophozoites, Luciferase expression was reduced ∼25% when synthetic miR5 was also introduced. The Luciferase mRNA level remained, however, unchanged, suggesting translation repression by miR5. This inhibition was fully reversed by introducing also a 2′-O-methylated antisense inhibitor of miR5, suggesting that miR5 acts by interacting specifically with the antisense sequence in the mRNA. A partial antisense knock down of GlDcr or GlAgo in Giardia indicated that the former is needed for miR5 biogenesis whereas the latter is required for miR5-mediated translational repression. Potential targets for miR5 with canonical seed sequences were predicted bioinformatically near the stop codon of Giardia mRNAs. Four out of the 21 most likely targets were tested in the Luciferase reporter assay. miR5 was found to inhibit Luciferase expression (∼20%) of transcripts carrying these potential target sites, indicating that snoRNA-derived miRNA can regulate the expression of multiple genes in Giardia.
| Giardia lambia is a deeply branched parasitic protozoan and the pathogen causing the diarrhetic disorder, giardiasis. The mechanism of gene regulation in this organism is largely unknown. Here, we identified a 26 nucleotide (nt) small RNA from the 3′-end of a 106 nt small nucleolar RNA (GlsR2) in Giardia. GlsR2 is processed through the action of a Dicer protein in Giardia to generate the 26 nt RNA. The latter becomes associated with the Argonaute protein. The protein-RNA complex can repress the translation of messenger RNAs carrying the antisense sequence of the 26 nt RNA at the 3′-untranslated region. This small RNA, designated microRNA5 (miR5), has several potential targets identified in Giardia, among which four were further tested in Giardia and found their translation repressed by miR5. This is the second functioning microRNA we have indentified in Giardia. The microRNAs could be thus important regulators of gene expression in this ancient single cellular organism.
| The intestinal parasitic protozoan Giardia lamblia, one of the earliest branching eukaryotes, is the causative pathogen of a diarrheal disease giardiasis throughout the world. It has been included as part of the WHO Neglected Disease Initiative since 2004 [1]. Though the transcripts in Giardia are produced in the nucleus and transported to the cytoplasm for translation like the other eukaryotes, few consensus promoters have been identified. Most of the transcription factors identified in other eukaryotes are missing in Giardia. There is but one highly divergent TATA-binding protein [2]. An AT-rich region commonly located near the beginning of open reading frames is believed to function as the promoter [3]. The transcripts in Giardia have exceedingly short 5′-untranslated regions (UTRs) ranging mostly from 0 to 14 nucleotides and similarly short 3′-UTRs of 10 to 30 nucleotides [3]. This has ruled out some mechanisms of translational regulation, which are essential in higher eukaryotes, such as ribosome scanning in translation initiation [4]. Moreover, the machinery of RNA interference (RNAi) is absent from Giardia. A specific double-stranded (ds)RNA virus (Giardiavirus) with a 6,277 basepairs (bp) has been found actively multiplying in the cell [5], and long dsRNAs were not degraded in Giardia [6]. Apparently, foreign dsRNA cannot be degraded in Giardia.
microRNAs (miRNAs) are an ancient class of small RNAs that mediate post-transcriptional regulation of mRNAs in animals and plants, which is critical for many biological processes [7]–[9]. Functional 18∼24-nucleotide (nt) mature miRNAs are derived from primary transcripts (pri-miRNAs) that are usually several kilobases long non-coding RNAs [9]–[11]. Cleavage at the stem of a hairpin structure in the pri-miRNA by the nuclear RNase III Drosha releases a smaller hairpin structure, the precursor miRNA (pre-miRNA) [9], [11], which is exported to the cytoplasm by Exportin5, a member of the nuclear transport receptor family [12]. In the cytoplasm, pre-miRNAs are cleaved near the terminal loop by Dicer, releasing a ∼22 nt double-stranded (ds) miRNA intermediate bearing a 2 nt overhang at the 3′-end [13], [14]. This duplex, often imperfectly paired, is then associated with an Argonaute protein plus other proteins to release the antisense strand and assemble into a protein-RNA complex, RISC (RNA-induced silencing complex) [15], [16]. The complex binds to a partially complementary sequence in the 3′- UTR of the target mRNA and represses its translation [7], [8].
Homologues of Drosha or Exportin5 have not been found in Giardia, but a single functional Dicer (GlDcr) and a single Argonaute (GlAgo) homologue have been identified in the Giardia genome database [17], [18]. The crystal structure of GlDcr was recently resolved, representing the first Dicer structure available [17]. It consists of two coupled RNase III domains associated with a canonical PAZ domain. But it lacks the N-terminal DExD/H helicase, C-terminal double-stranded RNA binding domain and a DUF283 domain identified in the Dicers of higher eukaryotes [17]. The distance between the PAZ and the processing center predicts a small RNA product of ∼25 nt. GlDcr was shown to cleave short dsRNA (155 bp) in vitro to generate 25∼27 nt RNA and support RNAi in a Schizosaccharomyces pombe Dicer deletion mutant [17]. In view of the inconclusive indications on whether a small RNA-mediated posttranscriptional regulation could be operational in Giardia, we isolated, cloned and sequenced some of the small RNAs in the size range of 20∼30 nts from Giardia and identified one of the 26 nt small RNAs, miR2, as a Dicer-digested product from the 3′-end of GlsR17, an orphan snoRNA containing a C/D box without an antisense sequence to ribosomal RNAs (rRNAs) [18], [19]. A conserved putative target site for miR2 was identified at the 3′-UTRs of 22 variant surface protein (VSP) mRNAs. Expression of a reporter mRNA in Giardia carrying this putative target site at the 3′-UTR was repressed by miR2 without altering the mRNA level. This repression, which was dependent on the presence of GlAgo, indicated the ability of this snoRNA-derived small RNA to function as a miRNA in RISC-mediated translational repression in Giardia [18].
In order to expand this observation to see if more functional miRNAs are derived from the snoRNAs in this organism [19], we pursued the identification and characterization of a new Giardia miRNA, miR5, from the 3′ end of another Giardia box C/D snoRNA, GlsR2. The result provided extensive evidence for the presence of another functional snoRNA-derived miRNA in Giardia.
Giardia lamblia (WB clone C6, ATCC50803) trophozoites were cultured as described previously [20]. Cells were grown anaerobically in plastic culture tubes at 37°C in the modified TYI-S-33 medium supplemented with antibiotics.
Total RNA was isolated from Giardia trophozoites using Trizol (Invitrogen), while size-fractioned small RNAs (<200 nts) were isolated using the mirVana kit (Ambion).
MAXIscript Kit (Ambion) was used to incorporate α-32P-UTP (Perkin Elmer) into the RNA probes. DNA probes for RL, RL-TS (see below), and 16S rRNA were generated from PCR using γ-32P-ATP (Perkin Elmer) end-labeled gene specific primers. Fifteen µg of size-fractionated small RNAs (<200 nts) were separated in a 12% denaturing polyacrylamide gels (8 M Urea, 1× TBE (Tris-Borate-EDTA) buffer) and capillary blotted onto a Hybond-N membrane (Amersham) followed by UV light irradiation. Blots were hybridized with the radiolabeled probes overnight at 42°C in a solution containing 50% formamide, 0.5% SDS, 5xSSC (150 mM NaCl, 15 mM sodium citrate), 5xDenhardt's solution and 100 µg/ml denatured salmon sperm DNA. The blots were washed twice with 2xSSC and 0.1% SDS for 15 min at room temperature followed by two washes with 0.1xSSC and 0.1% SDS for 15 min at 42°C. The hybridization signal was monitored with a PhosphorImager screen and scanned with a GE Storm 860 (Amersham).
Primer Extension System–AMV Reverse Transcriptase (Promega) was used for primer extension.
A cDNA sequencing ladder was obtained using fmol DNA Cycle Sequencing System (Promega). Primer (5′-GGC TCG GAC ATC CAA GG-3′) (IDT) used for cDNA sequencing and primer extension was PAGE-purified and γ-32P-ATP (Perkin Elmer) end-labeled using T4 polynucleotide kinase (New England Biolab). The cDNA thus synthesized was analyzed by electrophoresis in 8% polyacrymide/8 M urea gel along with the GlsR2 cDNA sequencing ladder. The gel was exposed to a PhosphorImager and scanned with GE Storm 860 (Amersham).
Giardia small RNAs (20∼30 nts) were cloned as described previously [18]. The RT-PCR products thus derived were used as the template for PCR aimed at amplifying the 3′ end of miR5. The products were then cloned into pGEM-T Easy vector using the pGEM-T Easy kit (Promega). Escherichia coli colonies containing the inserts were collected and the plasmid DNA was isolated and sequenced.
The N-terminal tagged (3×c-myc tags) GlDcr was integrated into the plasmid pNlop4 and expressed in Giardia cells as described [21]. For immunoprecipitation of GlDcr, 10 µl of anti-c-myc beads (Pierce) were incubated with 1% BSA for 12 hrs at 4°C, and then incubated with Giardia cell lysate of 5×107 cells in 0.5% BSA for 12 hrs at 4°C. The beads were extensively washed with Tris buffered saline (TBS), suspended in 50 µl Dicer storage buffer (50 mM Tris-HCl, pH 7.6, 50 mM NaCl, 5 mM MgCl2, 20% glycerol) and stored at −20°C. For in vitro dicing assay, 2 µl of the beads was incubated at 37°C with 500 ng of a substrate in the presence of 3 mM MgCl2, 30 mM NaCl and 100 mM Hepes, pH 7.5. In vitro transcribed GlsR2 (106 nts) and pre-miR5 (75 nts) were each trace-labeled using the MAXIscript Kit (Ambion). The final volume of each reaction was 10 µl. Reactions were stopped by adding 10 µl of formamide gel loading buffer. RNA fragments were resolved by denaturing polyacrylamide (12%) gel electrophoresis and visualized by phosphor imaging. The same reactions and electrophoresis were carried out using unlabeled in vitro transcribed GlsR2 and pre-miR5. The digested products were transferred onto a Hybond-N membrane (Amersham). An anti-miR5 sequense (5′ AAG GCT CGG ACA TCC AAG GAA GCA TC 3′) was γ-32P-ATP (Perkin Elmer) end-labeled and used as the probe in a Northern.
The N-terminal 3-HA tagged GlAgo was immunoprecipitated as described [21]. For RT-qPCR of miR5, the extracted ∼26 nt RNA band co-immunoprecipitated with GlAgo was extracted from the gel and reverse transcribed using the SuperScript III RT (Invitrogen) with the RT primer 5′ GTC GTA TCC AGT GCA GGG TCC GAG GTA TTC GCA CTG GAT ACG ACA AGG CTC G 3′. miR5 cDNA was then amplified using iQ Supermix (Bio-Rad), with a forward primer 5′ ACG ATG CTT CCT TGG ATG TC 3′, a reverse primer 5′ TAT CCA GTG CAG GGT CCG A 3′ and a TaqMan probe 5′ CTG GAT ACG ACA AGG CTC GGA CA 3′. The PCR products were analyzed by 2% agarose gel electrophoresis.
Giardia WB strain wild-type cells, GlAgo-knockdown cells or GlDcr-knockdown cells [18] were grown in modified TYI-S-33 media to a density of 107 per ml, washed twice in phosphate buffered saline (PBS), once in electroporation buffer (10 mM K2HPO4–KH2PO4 (pH 7.6), 25 mM HEPES (free acid), 120 mM KCl, 0.15 mM CaCl2, 2 mM EGTA, 5 mM MgCl2, 2 mM ATP, 4 mM glutathione), and then suspended in the electroporation buffer. Capped mRNA (4 µg), yeast tRNA (125 µg), synthetic 5′-phosphate-miR5 RNA (miR5, 1 µg, IDT) or synthetic 2′-O-methylated antisense of miR5 (ASO-miR5, 1 µg, IDT) were added in various combinations to the cell suspension, incubated on ice for 10 min and subjected to electroporation using a Bio-Rad Gene Pulser Xcell (Voltage: 450 V, Capacitance: 500 mF, Resistance: ∞). Cells were then incubated on ice for 10 min, added to pre-warmed culture medium, incubated at 37°C for 4 hrs, pelleted, washed once in PBS, and lysed using the Renilla Luciferase assay kit (Promega). The lysate was centrifuged at 12,000×g for 2 min to remove cellular debris. The cleared lysate was used to assay for Renilla Luciferase activity. The protein concentration of the cleared lysate was measured by the Bradford method (Bio-Rad) and used to normalize the Luciferase activity.
The first and only functional miRNA found in Giardia thus far, miR2, was a 26 nt small RNA derived from an orphan box C/D snoRNA GlsR17 [18]. GlsR2 is another box C/D snoRNA identified among the cDNA clones of a Giardia small RNA library [19]. The cloned GlsR2 cDNA has an 11 nt antisense element in the 5′-domain and was postulated to guide the 2′-O-methylation of Cm334 in the 16S rRNA of Giardia. Primer extension of the 16S rRNA, however, failed to show methylation at position C334 and the real function of GlsR2 in Giardia remains unclear [19].
To see if GlsR2 could be processed to a smaller RNA species in Giardia, Northern blot assays were performed using 15 µg of size-fractionated small RNAs (<200 nts) probed with the full-length GlsR2 antisense RNA (Figure 1). The results showed that a band with an estimated size of about 75 nt and a very faint band of about 26 nt could be detected in addition to the full length GlsR2 band. To determine which part of GlsR2 may have generated the two smaller RNA species, the sequence of GlsR2 cDNA was divided into three overlapping portions; the 5′-portion (1∼60 nt), the mid-portion (31∼80 nt) and the 3′-portion (61∼104 nt) (Figure 1), and their antisense RNAs were used as probes in Northern blots. All three probes were capable of hybridizing to the full-length GlsR2 band (Figure 1). The 75 nt band was intensely stained by the mid-portion probe and the 3′-portion probe but only weakly hybidized with the 5′-portion probe. Only the 3′-portion probe was able to detect the 26 nt RNA band. Thus, the 26 nt RNA is probably derived from the 3′ portion of GlsR2.
To determine the precise 5′ ends of both the 75 nt RNA and the 26 nt RNA located at the 3′ portion of GlsR2, primer extension assays were performed (Figure 2). A 17 nt end-labeled primer complementary to the 3′ end of GlsR2 [19] was used. Four µg of size-fractionated small RNAs (<200 nts) was used for the primer extension. An in vitro transcribed GlsR2 (0.75 ng) was used as a control template in order to distinguish nonspecific stops caused by the secondary structures of GlsR2 from the real stop. GlsR2 sequencing reactions were run in parallel to identify the precise stopping sites of the primer extension products. Figure 2 shows the products from two separate primer extension reactions run in the same gel along with the GlsR2 sequencing ladder. Two bands of 24 nt and 73 nt were identified in the two primer extension reactions but were absent from the control lanes. The 24 nt band could correspond to the ∼26 nt RNA band whereas the 73 nt band could appear as the ∼75 nt band in the Northern (see Figure 1 and below). From the flanking sequence ladders, the 5′-end of the 24 nt RNA was positively identified to be G81, whereas that of the 73 nt RNA was A32.
The detection of a 73 nt and a 24 nt RNA fragment by primer extension instead of the 75 nt and 26 nt pieces seen in the Northern (Figure 1) was puzzling. The 3′-ends of the primer extension products were defined by the primer, which was synthesized according to the published sequence of GlsR2, with a defined full-length of 104 nt [19]. We decided to re-examine the 3′-ends of GlsR2 and the two smaller RNA species derived from it. A 20 nt primer complementary to the 3′-linker and a 20 nt primer starting from the 5′ end (G81) of the 24 nt small RNA defined by the previous primer extension were used to amplify a Giardia small RNA cDNA library. The PCR products were cloned and sequenced. All the sequencing results from the 12 isolated clones showed an additional pair of UU added to the reported GlsR2 3′-terminus. The UU comes from the genomic sequence. Apparently, a slight degradation happened at the 3′ end of GlsR2 when it was originally cloned and sequenced by Yang et al. [19], causing the loss of the UU pair. The full length of GlsR2 is thus most likely 106 nts.
Thus, the sequence of GlsR2-derived small RNA is 26 nts and defined as: 5′- GAU GCU UCC UUG GAU GUC CGA GCC UU -3′. It was subsequently designated as miR5 due to its biological activity (see below). Its precursor (pre-miR5) is most likely the 75 nt RNA which starts with A32 and ends with the UU pair. In our recent deep sequencing data for GlAgo-associated small RNAs, there are 1121 hits for the 26 nt miR5 sequence and 107 hits for the 25 nt sequence from the 5′ end of pre-miR5 (5′-AGG CGA UGG AGA CAA AAG CAG UUA C-3′), which forms double strand RNA with miR5 in the hairpin structure (Figure 3). The latter could be thus the miRNA*, which is probably also incorporated into Ago and regulate the expression of target genes [9], [11]. This miR5* has a 2 nt overhang at the 3′-end, suggesting the specific feature from GlDcr processing. No apparent iso-miR5 sequence was detected in the deep sequencing data.
The secondary structures of GlsR2 were analyzed by MFOLD (Figure 3). The putative box C (nts 38–43) and box D (nts 98–101) form a stem suitable for binding to the snoRNPs. The 75 nt putative precursor of miR5 (pre-miR5) is folded into a hairpin structure, which could serve as a substrate of GlDcr (Figure 3, black boxed). The enzyme for converting GlsR2 to the 75 nt pre-miR5 remains to be further verified (see below).
N terminal tagged GlDcr (3×c-myc GlDcr) was over-expressed in Giardia, pulled down by anti-c-myc beads, washed thoroughly (Figure 4A), and used in the in vitro dicing assay. GlDcr was shown to cleave short dsRNA (155 bp) to generate 25∼27 nt RNA in vitro [17]. A 32P-labeled double stranded (ds) RNA (106 bp) was used as a substrate to test the tagged-GlDcr beads and was found to be processed to smaller RNA fragments with a 25∼27 nt RNA band as the primary product (Figure 4B). A time course of the processing was followed at 8 and 16 hrs of incubation, and the results showed that 16 hrs are required for a thorough digestion of the substrate (Figure 4B). We then tested the GlDcr beads on 32P-labeled full-length GlsR2 or pre-miR5 and both were processed to a 26 nt RNA band, suggesting that both RNA species could be processed by GlDcr to produce miR5 (Figure 4C, left panel). To confirm that the 26 nt RNA is the miR5 itself, we repeated the same experiments with unlabeled GlsR2 and pre-miR5 and monitored the products with an anti-miR5 sequence in a Northern analysis. The result, presented in the right panel of Figure 4C, indicates that miR5 is a product from GlDcr digestion of either GlsR2 or pre-miR5. There is, however, no apparent conversion from GlsR2 to pre-miR5, suggesting that this particular reaction could be catalyzed by an enzyme other than GlDcr. Other apparent RNA bands were also found among the products from GlDcr-beads digestion as well as the control (Figure 4C). They could be derived from some potential minor contaminating RNases in Giardia not fully washed off the beads.
Additionally, we analyzed the biogenesis of GlsR2 in a GlDcr knockdown strain of Giardia trophozoites. GlDcr-antisense-hammerhead ribozyme RNA was used to partially knock down the expression of GlDcr [18]. Western blot showed that GlDcr protein was decreased by 29% in GlDcr knockdown cells comparing with the control (Figure 4D, upper panel). Size-fractionated small RNAs (<200 nts) were extracted from both the GlDcr knockdown cells and the control cells and examined for miR5 in a Northern blot. The results indicate that miR5 is decreased by 39% in the GlDcr knockdown cells (Figure 4D, lower panel). The study was repeated three times with similar outcomes, which underwent statistical analysis and resulted in a p value<0.0006 for the reduction in GlDcr and <0.0005 for the corresponding drop in miR5. It is thus likely that GlDcr is required for in vivo biogenesis of miR5.
To further ascertain that miR5 is a functional miRNA that binds to GlAgo to form a potential RISC in Giardia, we immunoprecipitated GlAgo and examined if miR5 was associated with it. We have recently established a method of immunoprecipitating N-terminal 3×HA-tagged GlAgo from Giardia lysate and identified a single ∼26–30 nt RNA band associated with the protein that contained all the miRNAs we have identified thus far [21]. The miR5 level in this ∼26–30 nt RNA band was analyzed by RT-qPCR. The Ct value of the GlAgo pulled-down sample was about 9 cycles earlier than the Ct of the control sample (WB), suggesting that miR5 was significantly enriched in the GlAgo immunoprecipitate (Figure 5). miR5 is thus associated with GlAgo in Giardia, most likely in forming a RISC.
Most of the functional miRNAs reported to-date have been shown to bind to the 3′ UTR of their target mRNAs through base pairings and to exert an inhibition of translation of the mRNA [22]. Since miRNAs have usually multiple targets in a cell, the individual sites have typically reduced protein output by less than a half and often by less than a third from the action of a single miRNA [8]. To test the potential function of miR5 in vivo, the miR5 antisense sequence was incorporated into the 3′-UTR of a Renilla Luciferase (Rluc) cDNA, transcribed with T7 RNA Polymerase and capped in vitro (Figure 6A). The transcript RL-TS was co-transfected with synthetic miR5 into Giardia WB trophozoites and expressed at 37°C for 4 hrs. The cell lysate was then assayed for Rluc activity.
With the Luciferase expression from control mRNA with a miR5 sense sequence at the 3′UTR (RL) set at 100%, a co-transfection of synthetic miR5 with RL mRNA showed little effect on Luciferase expression (Figure 6B). A transfection with RL-TS mRNA alone, however, reached only 82% of Rluc expression, suggesting presence of endogenous miR5. A co-transfection of RL-TS with exogenous miR5 further reduced the Rluc expression to 61%. Apparently, the presence of endogenous miR5 in Giardia reduces the RL-TS expression by 18%, whereas the presence of additional synthetic miR5 further decreased the expression by a total of 39%. The inhibitory effect of the exogenously introduced miR5 is thus ∼21%.
To verify if the repressed Rluc expression requires hybridization between miR5 and the potential target site, a 2′-O-methylated antisense RNA sequence of miR5 (ASO-miR5) was synthesized [23]. The 2′-O-methylated antisense sequences are known to bind to the corresponding miRNAs in RISCs with extremely high affinity, effectively out-competing the target mRNA [23]. They represent a reliable tool in validating miRNA targets and studying the cellular functions of miRNAs. [23] When 1 µg of miR5 and 1 µg of ASO-miR5 were simultaneously introduced into the cells expressing RL-TS, the inhibition caused by exogenous miR5 was abolished and the Rluc activity increased back to the 80% level, equivalent to that induced by the endogenous miR5 alone (Figure 6B). Thus, the inhibitory effect of exogenous miR5 is most likely attributed to its binding to the potential target site (TS) and can be abolished by an equivalent amount of ASO-miR5. When the expression of an Rluc mRNA with 2 TSs separated by a 10 nt spacer in the 3′-UTR (RL-2TS) was assayed (Figure 6C), it was repressed by miR5 (1 µg) to 71%, a little more efficient than that observed with the single target site mRNA RL-TS (75%). When 1 µg of ASO-miR5 was introduced into the RL-2TS cells without introducing exogenous miR5, the Luciferase activity increased to >130% of the no ASO-miR5 control, suggesting that ASO-miR5 is also capable of inhibiting the function of endogenous miR5 (Figure 6C).
Levels of RL and RL-TS transcripts in the transfected Giardia were monitored by Northern blots (Figure S1). The results show little change in the levels of the mRNAs when the exogenous miR5 was introduced into the cells, indicating that the inhibited expression of Rluc by miR5 was not attributed to mRNA degradation, but more likely due to translation repression of the mRNA.
The Argonaute protein is an integral component of RISC and plays an essential role in miRNA-mediated repression of mRNA translation [24]. To knock down Giardia Argonaute mRNA, a specific GlAgo-antisense-hammerhead ribozyme RNA was introduced into Giardiavirus-infected Giardia trophozoites as previously described [18]. The result indicated that the level of GlAgo mRNA was significantly reduced (Figure 7, right panel). The GlAgo knockdown cells were transfected with the RL-TS mRNA and the synthetic miR5, and the expression of Rluc was monitored. The results show that miR5 can no longer inhibit the expression of RL-TS in the GlAgo knock down cells and the Luciferase activity went back to the RL control level (Figure 7, left panel). Thus, GlAgo is required for the miR5-mediated repression of translation in Giardia.
Current methods for predicting a target site of a miRNA have been diverse and require more room for improvement [25], [26]. However, three agreements appear to have emerged among them: (1) the need of conserved Watson-Crick pairing to the 5′ region of the miRNA centered on nucleotides 2–7 designated the seed-matched site; (2) Conserved pairing to the seed region can be sufficient on its own for predicting the targets; (3) an 8 nt match or multiple matches to the same miRNA increase the authenticity of the target. John et al. developed the miRanda algorithm to predict miRNA targets in different species. Three validation approaches (retrospective, statistical, and indirect experimental) showed that when the predicted targets are conserved and the gene contains more than one miRNA target sites or a single high-scoring site (the alignment score >110), genuine miRNA targets can be predicted at a reasonable accuracy [27]. We thus tried to predict the putative miR5 target sites in the Giardia genome using the miRanda program. Since Giardia mRNAs have relatively short 3′-UTRs and a growing number of mRNAs in other organisms have been found to be targeted by miRNAs within the open reading frames (ORFs) rather than the 3′-UTRs [28]–[30], segments of 100 nts with 50 nts upstream and 50 nts downstream from the stop codon of each ORF were extracted from the 4,889 ORFs in GiardiaDB [31], [32]. Twenty-one potential target sites for miR5 that bear a perfect complementation with nucleotides 2–7 in miR5 with a score threshold >120, and an energy threshold <−20 kcal/mol were selected (Table S1). Among them, 12 are located in the ORF region close to the stop codon, 9 are partially or totally in the 3′-UTR of the target mRNAs, whereas all 21 are upstream from the polyadenylation site (data not shown). For the 21 ORFs, 14 are hypothetical proteins and 7 are annotated proteins consisting of four kinases and three variant surface proteins (VSPs) (Table S1).
GlsR2 is a homologue of snoRNA U14 in yeast and vertebrates. The 3′ end, where miR5 is derived from, is the most conserved region. We went to deepBase [33], a platform for annotating and discovering small and long ncRNAs from deep-sequencing data, and found that there is a large number of small RNAs coming from the 3′ ends of human, mouse, chicken, Ciona and Drosophila U14 homologues, suggesting that miR5 may be conserved. We could not, however, find any homologue of the 21 potential Giardia mRNA targets of miR5 in other eukaryotes. Recent release of genome sequences and annotations of Giardia intestinalis GS [34] and Giardia E isolate P15 showed the presence of GlsR2 homologue in all three Giardia isolates. Comparative analysis of the 21 predicted mRNA targets among three isolates showed that 15 target sites were well conserved in at least one of the two other isolates, while the remaining 6 were found only in WB. Thus, a similar miR5-mediated regulation of gene expression could be functional among the three Giardia isolates.
Four of the 21 potential target sites were selected for further study. The first site (TS1) is from a hypothetical protein gene, which has the highest score. The second site (TS2) is from one of the NEK kinase genes and has the lowest free energy. The third site (TS3) is from a VSP gene and has the lowest score, while the fourth site is from a hypothetical protein gene with the highest free energy (Figure 8A and Table S1). Duplicate copies of the four target sites were cloned into the 3′-UTR of Rluc mRNA (designated RL-2TS1, 2, 3 and 4) and the capped and in vitro transcribed RNA was transfected into Giardia. Luciferase assay showed that exogenous miR5 reduced the expression of the four mRNAs to 83.7%, 83.5%, 77.8% and 84.3% respectively, compared to the no miR5 control (Figure 8B). Another Rluc mRNA with two copies of the target site of a newly identified 26 nt miR4 at the 3′-UTR [21] was included as another control in the study. Like miR5, miR4 is playing a role in regulating VSP gene expression in Giardia [21]. This miR4 target site is from another VSP (GL50803_36493), which carries no target site for miR5. The result shows that miR5 has no detectable effect on the expression of this transcript carrying double target sites for miR4 (Figure 8B), suggesting that the miR5 repression of Rluc carrying TS1, 2, 3 or 4 is not a nonspecific phenomenon. The similar levels of repression from 15.7 to 22.2% among the four target sites suggest that the scores and free energies predicted by miRanda do not play a pivotal role in identifying miRNA target sites in Giardia. This could be attributed to the stringent criteria we used in predicting the target sites. All the conserved canonical 7–8 nt seed-matched sites enlisted in Table S1 could belong to the same family of authentic binding sites of miR5.
The presence of a miRNA pathway in Giardia has now been demonstrated by two examples of functional miRNAs, miR2 [18] and miR5, both of which are derived from snoRNAs. There is thus the probability that the snoRNAs may constitute a significant source of miRNAs in Giarida. Among the 16 putative Giardia box C/D snoRNAs identified thus far [19], 11 carry a single antisense element of rRNAs, whereas the other 5 have none, and could be classified as orphan snoRNAs. The corresponding rRNA targets postulated for the 11 box C/D snoRNAs were examined by primer extension [19]. Only GlsR1 showed an rRNA modification at the snoRNA-targeted site [19]. This suggests that most, if not all, of the 16 box C/D snoRNAs may not function as guides of 2′-O-methylation of rRNA but may perform some other functions, such as being the precursors of miRNAs in Giardia. GlsR2 is a homologue of snoRNA U14 in yeast and vertebrates but without the necessary A domain for the chaperon function of U14 [35]. Since the predicted site of methylation by GlsR2 was not verified by primer extension [19], it may also be an orphan snoRNA functioning as a precursor of miR5. However, it should not imply that snoRNAs constitute the only source of miRNAs in Giardia. Even in the apparent absence of Drosha/Pasha and Exportin5, there could be other potential precursor RNAs matured into miRNAs through digestion by Dicer and other yet un-identified enzyme(s) in Giardia [21].
The GlDcr protein has been found to localize to the cytoplasm of Giardia trophozoites [36], which raises the question on how a cytoplasmic enzyme can digest snoRNAs presumably localize to the nucleolus. In mammalian cells, box C/D snoRNAs are in the nucleolar snoRNP consisting of phosphorylated export adaptor PHAX, the cap binding complex Ran and the exportin CRM1, suggesting a cytoplasmic phase during the maturation of snoRNP [37]. This suggestion was supported by the observation that U8 and U22 snoRNAs injected into the cytoplasm were imported into the nuclei of Xenopus oocytes [38], and that U8 pre-snoRNPs had a distinct distribution in the nucleoplasm and cytoplasm with an association with both nuclear import and export factors during maturation [39]. snoRNAs thus appear to have a cytoplasmic phase during their maturation. Ran and CRM1 homologues have been identified in the Giardia genome database ([40], data unpublished). It is possible that the Giardia box C/D snoRNAs can be exported to the cytoplasm by the exportin CRM1 complex during their biogenesis and subject to processings to produce miRNAs in the cytoplasm. More recently, CRM1 was found to mediate nuclear-cytoplasmic shuttling of miRNAs and co-immunoprecipitate with Argonaute in mammalian cells [41]. It raises the interesting possibility that the same CRM1 complex could transport miRNAs and their snoRNA precursors, thus constituting a primitive pathway of post-transcriptional regulation in Giardia.
For the 7 annotated ORFs that carry potential target sites for miR5, three of them encode VSPs. These VSPs could be clustered in a subfamily according to their sequence similarity. They are different from the 22 VSP genes shown to carry the putative targets of miR2 [18], suggesting that, like miR2 [18], miR5 may also play a role in regulating the expression of VSPs in Giardia. Thus, a miRNA machinery may constitute an important mechanism of VSP gene regulation.
Most of the RNAs processed by Dicer result in a product with a 5′-monophosphate (5′-monoP) and a 3′-hydroxyl (3′-OH) structure which could be identified using the standard small RNA cloning techniques [13], [42]. However, certain small RNAs carry modified 5′ and/or 3′ termini. In C. elegans, secondary siRNAs contain a 5′-triphosphate group and are exclusively loaded into the secondary Argonautes (SAGOs) that can amplify the gene silencing effect [43]. These small RNAs with 5′-polyphosphate ends were also identified in the single-celled anaerobic eukaryote Entamoeba histolytica [44]. They were mainly mapped to the antisense genes and found associated with an E. histolytica Piwi-related protein [44]. In addition, miRNAs and siRNAs in plants, Piwi-interacting RNAs in animals and some miRNAs and siRNAs in Drosophila have 2′-O-methylation (2′OMe) on the 3′ terminal nucleotide [45]. This 3′-end modification can protect the small RNAs from 3′-end uridylation and 3′-to 5′-exonuclease-mediated degradation [46], [47]. However, the standard small RNA cloning method, which was used by us in identifying miR2 [18] and miR5 in Giardia, will not efficiently capture the small RNAs with a 5′-triphosphate structure [44]. The 2′-O-methylation at the 3′-end of RNA reduces also the efficiency of ligation by T4 RNA ligase by the standard cloning technique [48]. Recently, it was shown that small RNAs which can be co-immunoprecipitated with GlAgo have a 5′-monophosphate and a 3′-OH [21]. These facts, plus the experimental data that chemically synthesized miR2 and miR5 with 5′-monophosphates and 3′-OH's can be introduced into Giardia trophozoites and effectively repress the translation of mRNAs with the corresponding targeting sites at the 3′-UTRs provide a strong indication that these miRNAs have 5′-monophosphates and 3′-hydroxyls.
MicroRNAs constitute one of the more abundant classes of gene-regulatory molecules in animals [49]. Bioinformatic analysis and some high throughput experimental approaches showed that each miRNA has hundreds of evolutionarily conserved targets and a number of non-conserved targets [25], [50]–[52]. Recently, it was demonstrated that the majority of human genes (>60%) are under the control of miRNAs [49]. Moreover, Wu et al. showed that 28 miRNAs can modulate the expression of CDKN1A (p21Cip1/Waf1) by directly binding to its 3′-UTR [53], which confirmed that multiple miRNAs can target a single gene transcript. This suggests a complex network of interactions between miRNAs and mRNAs. We showed in this study that, experimentally, at least four genes and, bioinformatically, at least 17 additional genes in Giardia could be regulated by miR5. All the four experimentally demonstrated miR5 binding sites start in the ORF region and are located near or covering the stop codon. This might be a unique feature of miRNA-mediated gene regulation in Giardia due to the extremely short 3′-UTR of its mRNAs. Because the criteria we used for predicting targets are relatively stringent [52], we can expect that there are more miR5 targets in Giardia. Further experimental analysis of the remaining potential miR5 binding sites will eventually lead us to identify the complete spectrum of miR5 binding sites in Giardia.
The miRNA pathway in Giardia demonstrated by us is not well suited for defending the cells against the invading foreign dsRNA carried by the Giardiavirus. Those organisms equipped with the RNAi machinery such as Trypanosoma brucei [54], Tetrahymena pyriformis [55] or Caenorhabiditis elegans [56] have been known to be dsRNA virus free. miRNA-mediated gene repression is likely an antiquated and evolutionarily well conserved pathway of post-transcriptional gene regulation. The utilization of small hairpin structured RNAs, such as the snoRNAs, as the precursors of miRNAs suggests that the miRNA pathway could have evolved prior to the addition of Drosha/Pasha and Exportin5 into the pathway.
There was, however, one recent publication claiming the presence of RNAi in Giardia [36]. Although there was no actual RNAi experiment presented in the report, Prucca et al. showed that small pieces of VSP RNAs could be produced in Giardia extracts, and that multiple VSPs could be expressed on the cell surface when GlDcr or RNA-dependent RNA polymerase (RdRP) was partially knocked down using antisense RNA. These results led to the conclusion that VSP dsRNAs are synthesized by RdRP and processed by GlDcr for an RNAi mechanism of regulating VSP expression in Giardia. There was, however, no evidence showing that the small RNAs in Giardia extracts could be generated by GlDcr. In fact, there are so many RNases released in a Giardia extract, it is virtually not possible to claim a small RNA band being the product from Dicer action. There is also no slicer activity found in GlAgo, a key requirement for RNAi, and GlDcr is apparently incapable of digesting long dsRNAs (see above). There are thus still many uncertainties concerning the presence of a RNAi machinery in Giardia.
|
10.1371/journal.ppat.1005868 | Natural Killer Cell Evasion Is Essential for Infection by Rhesus Cytomegalovirus | The natural killer cell receptor NKG2D activates NK cells by engaging one of several ligands (NKG2DLs) belonging to either the MIC or ULBP families. Human cytomegalovirus (HCMV) UL16 and UL142 counteract this activation by retaining NKG2DLs and US18 and US20 act via lysomal degradation but the importance of NK cell evasion for infection is unknown. Since NKG2DLs are highly conserved in rhesus macaques, we characterized how NKG2DL interception by rhesus cytomegalovirus (RhCMV) impacts infection in vivo. Interestingly, RhCMV lacks homologs of UL16 and UL142 but instead employs Rh159, the homolog of UL148, to prevent NKG2DL surface expression. Rh159 resides in the endoplasmic reticulum and retains several NKG2DLs whereas UL148 does not interfere with NKG2DL expression. Deletion of Rh159 releases human and rhesus MIC proteins, but not ULBPs, from retention while increasing NK cell stimulation by infected cells. Importantly, RhCMV lacking Rh159 cannot infect CMV-naïve animals unless CD8+ cells, including NK cells, are depleted. However, infection can be rescued by replacing Rh159 with HCMV UL16 suggesting that Rh159 and UL16 perform similar functions in vivo. We therefore conclude that cytomegaloviral interference with NK cell activation is essential to establish but not to maintain chronic infection.
| Natural killer (NK) cells are an important subset of the innate immune system that rapidly responds to cellular transformation and infection. The importance of NK cell control of viral infection is dramatically illustrated by our results revealing that cytomegalovirus (CMV) is unable to establish infections in healthy individuals unless NK cell responses are subverted. By studying infection of rhesus macaques with rhesus CMV, a highly representative animal model for human CMV, we identified a key viral factor that allows RhCMV to limit NK cell activation by preventing NK cell activating ligands from trafficking to the cell surface. Importantly, we observed that this avoidance of NK cell activation is essential to establish infection in vivo because RhCMV lacking the NK cell evasion factor was unable to infect animals unless NK cells were depleted. By unmasking such viral stealth strategies it might be possible to harness innate immunity to prevent viral infection, the primary goal of CMV vaccine development.
| NK cells are a significant component of innate immunity against viruses and NK cell-deficient individuals are highly susceptible to herpesvirus infections [1]. Herpesviruses, particularly cytomegalovirus (CMV), encode proteins that either inhibit or activate NK cells and elegant studies with murine CMV (MCMV) revealed a complex relationship between NK cell stimulation and MCMV evasion during infection [2]. NK cell activation is controlled by inhibitory and activating receptors with inhibitory receptors, such as the KIR and CD94/NKG2 that recognize MHC-I, generally overriding positive signals [3]. Destruction of MHC-I by CMVs generates a “missing self” situation that reduces inhibitory signals [4].
A major activating receptor on NK cells is NKG2D, which is also expressed on γδ T cells, some CD4+ T cells, all αβ CD8+ T cell in humans, and activated and memory αβ CD8+ T cells in mice [5]. NKG2D interacts with multiple ligands: MHC-I related molecules (MICA and MICB) and the UL16-binding proteins (ULBP1-6) in humans, and the H60, MULT-1 and RAE-1 proteins in mice (reviewed in [5, 6]); all of which are induced upon cell stress including viral infection. Both human CMV (HCMV) and MCMV devote multiple gene products to prevent the surface expression of NKG2DL, presumably because the induction of any one can activate NKG2D [7]. In HCMV, UL16 retains ULBP1, 2, 6 and MICB in the endoplasmic reticulum (ER) [8–12] and MICB is additionally targeted by micro-RNA UL112 [13]. In addition, ULBP3 is retained by UL142 [14], whereas MICA is downregulated by UL142 [15], as well as US18 and US20 [16]. The fact that both virus and host devote multiple gene products to modulating NKG2D activation suggests an evolutionary arms race [17] that is exemplified by the observation that a recently evolved truncated allele of MICA (MICA*008) is counteracted by HCMV US9 [18].
The impact of NKG2DL interception by HCMV on primary infection and persistence, as well as on reinfection of seropositive individuals is not known. Studies in mice indicate that NKG2DL-inhibitory MCMV gene products are not required for infection but reduced viremia is observed in their absence [19]. Interestingly, replacing the NKG2DL-inhibitor m152 with RAE-1γ increased CD8+ T cell responses, both short and long term, despite viral attenuation [20]. Thus, increased NKG2D activation might improve the immunogenicity of CMV-based vectors while increasing safety.
CMV-based vectors are currently being developed for HIV/AIDS based on findings obtained with RhCMV-vectors in rhesus macaque (RM) models of HIV [21–23]. These studies revealed an unprecedented level of protection by RhCMV-based vaccines against Simian immunodeficiency virus (SIV) along with an unexpected ability of RhCMV to control T cell epitope targeting [24, 25]. The close evolutionary relationship between human and RM host also extends to the CMVs with the vast majority of HCMV genes conserved in RhCMV [26]. Interestingly however, while gene products involved in T cell control are largely conserved between RhCMV and HCMV [27, 28], the NKG2DL-inhibitory HCMV gene products, UL16 and UL142, are notably absent from RhCMV. In contrast, most NKG2DL are highly conserved in RM: MIC1 and MIC2 are closely related to MICA and MICB, respectively, whereas MIC3 is a chimera of MIC1 and MIC2 [29] and the RM genome encodes for ULBP1-4 are also highly conserved compared to humans [30] (S1 Fig). Given the conservation of the ligands but not of the viral NKG2DL-inhibitors, we examined whether RhCMV evolved unique NKG2DL-inhibitors. Using a panel of cell lines expressing human and rhesus NKG2DLs we demonstrate that RhCMV inhibits surface expression of all NKG2DLs tested and we identify Rh159, the homologue of HCMV UL148, as a major gene product responsible for retention of NKG2DLs. Similar to UL16, Rh159 prevents surface expression of multiple NKG2DLs. In contrast, UL148 did not retain NKG2DLs consistent with divergent evolution of protein function from a common ancestor. In vitro, deletion of Rh159 increased human and RM MIC protein expression and NK cell stimulation by RhCMV-infected cells. In vivo, RhCMV was unable to establish infection in either RhCMV sero-positive or sero-negative animals when Rh159 was deleted. However, primary infection occurred when the CD8+ cell population, that includes NK cells, was depleted. Moreover, infection of RM occurred when R159 was replaced with UL16 consistent with functional but not sequence homology between these two proteins. Our results suggest that NK cell evasion by NKG2DL downregulation is essential for infection by RhCMV and, given the conservation of host NKG2DL and the fact that UL16 can functionally replace Rh159, most likely by HCMV as well.
Since both MCMV and HCMV interfere with expression and intracellular transport of NKG2DLs, we hypothesized that RhCMV would similarly affect NKG2DL expression. Given the high degree of homology between human and RM NKG2DLs (S1 Fig), we took advantage of a panel of established U373 cell lines stably expressing human ULBP1-3, MICA, or MICB [U373-NKG2DL] [31]. RhCMV can productively infect U373 cells (S2 Fig). Therefore, we infected U373-NKG2DLs with RhCMV and determined whether RhCMV interferes with expression of these human NKG2DLs by monitoring the cell surface expression levels of each NKG2DL by flow cytometry (“RhCMV” refers to BAC-cloned RhCMV 68–1 [26] unless otherwise noted). Cell surface expression of NKG2DLs was assessed on infected cells by co-staining for IE2+ cells. At 48 hours post-infection (hpi), a substantial decrease in cell surface levels of each of the five human NKG2DLs was observed in RhCMV-infected cells (Fig 1). In contrast, RhCMV infection had a minimal impact on surface expression of transferrin receptor (TfR) suggesting that RhCMV specifically targets NKG2DLs. Since RhCMV interfered with expression of each of the human NKG2DLs and RM NKG2DLs (see below) examined, these data suggest that RhCMV targets the full panel of NKG2D-activating ligands presumably to prevent NKG2D-dependent activation of NK cells.
We hypothesized that RhCMV targeted NKG2DL by a post-translational mechanism since transcription of these genes is controlled by heterologous transcriptional and translational control elements in U373-NKG2DL cells [31]. Therefore, we monitored glycosylation of MICB over 24 h of infection by collecting cells at various hpi and digesting cell lysates with Endoglycosidase H (EndoH) or Peptide N-Glycosidase F (PNGaseF) prior to electrophoretic separation and immunoblot with MICB-specific antibodies. Over time, increasing amounts of EndoH-sensitive compared to EndoH-resistant MICB protein accumulated in RhCMV-infected cells (Fig 2A). Since EndoH specifically removes high mannose oligosaccharides generated in the ER, EndoH sensitivity suggests that RhCMV retained MICB in the ER/cis-Golgi. The decrease of EndoH-resistant MICB over time mirrors the turnover of MICB seen in uninfected cells and thus was likely due to natural turnover [32]. To determine whether RhCMV specifically targeted newly synthesized MICB, we immunoprecipitated MICB from metabolically pulse/chase labeled U373-MICB cells at 24 hpi. In uninfected cells, the majority of MICB was EndoH resistant at 1 h post-chase (Fig 2B, left panel). In contrast, MICB did not attain EndoH-resistance in RhCMV-infected cells even at 3 h (Fig 2B, right panel), which was consistent with RhCMV preventing MICB maturation. Interestingly, an additional protein species (indicated by *) was observed in MICB-immunoprecipitations in RhCMV-infected, but not in uninfected samples at the 1 h and 3 h time point (Fig 2B). We hypothesized that the co-precipitating protein, which seemed to be ER-resident since it remained EndoH-sensitive throughout the chase period, was of viral origin. To identify this glycoprotein, we performed preparative MICB-immunoprecipitations from RhCMV-infected U373-MICB cells for analysis by mass-spectrometry (MS). When the immunoprecipitated proteins were separated by SDS-PAGE and visualized by Coomassie-staining, only MICB was observed in samples from uninfected cells whereas approximately equal amounts of MICB and the unknown ~40kDa species were co-precipitated by MICB-specific antibodies from RhCMV-infected cells (Fig 2C). The ~40kDa protein was excised from the gel, digested with trypsin and analyzed by liquid-chromatography tandem MS (LC-MS/MS). Multiple peptide-species were identified that corresponded in mass to predicted peptides of Rh159 (Fig 2D), a putative RhCMV glycoprotein with a predicted MW of 36.8 kDa (34.1 kDa without the putative signal peptide) that is similar to the observed MW of the (de-glycosylated) viral protein co-immunoprecipitating with MICB (Fig 2B). Rh159 is predicted to contain two N-linked glycosylation sites and to display a type1b transmembrane protein topology with a cleavable signal sequence. A very short cytoplasmic tail of seven amino acids encompasses a putative RXR ER retrieval motif (Fig 2D). Rh159 was thus a strong candidate for an ER-resident viral glycoprotein that associates with and retains newly synthesized NKG2DL by hijacking the cellular ER-retrieval mechanism.
To determine whether Rh159 was sufficient to retain MICB, we inserted codon-optimized Rh159 (including a C-terminal FLAG-epitope tag) into a replication-deficient adenovector under control of the tetracycline-regulated promoter. U373-MICB cells were co-transduced with the corresponding construct (AdRh159FL) and an adenovector expressing the tetracycline-regulated transactivator (AdtTa). For control we used an adenovector expressing GFP (AdGFP). At 24 hpi, MICB was immunoprecipitated from Ad-transduced U373-MICB cells at 0, 1, and 3 h post- chase. In U373-MICB cells transduced with AdGFP, the majority of newly synthesized MICB attained EndoH resistance within 1h post-chase (Fig 3A). In contrast, the majority of MICB remained EndoH-sensitive throughout the chase period in AdRh159FL-transduced cells (Fig 3A). We further observed that MICB-specific antibodies co-immunoprecipitated an EndoH-sensitive protein of ~40KDa from AdRh159FL-transduced cells but not from AdGFP-transduced cells (Fig 3A). The ~40kDa band was identified as Rh159 by immunoblot with anti-FLAG antibodies (Fig 3B) whereas the band was not observed in U373-ULBP3 cells transduced with AdRh159FL or when an adenovector expressing C-terminal FLAG tagged Simian varicella virus (SVV) open reading frame (ORF) 61 was used [33]. Interestingly, steady state protein levels of MICB were strongly reduced and were mostly EndoH sensitive at 48 hpi with AdRh159FL, suggesting that prolonged retention ultimately leads to degradation of MICB (Fig 3C). Taken together, these data demonstrate that Rh159 associates with and retains MICB in the ER.
To determine whether Rh159 also targets other NKG2DLs, we co-transduced the U373-NKG2DL panel with AdtTA and AdRh159FL or AdGFP. Expression of Rh159 was confirmed by immunoblot (Fig 3D) and NKG2DL-surface expression was monitored by flow cytometry (Fig 3E). Mean fluorescence intensity (MFI) of MICB was reduced by more than 2 orders of magnitude upon transduction with AdRh159FL compared to AdGFP whereas cell surface levels of TfR were not affected (Fig 3E). Similarly, Rh159-expression reduced surface levels of MICA, ULBP1, and ULBP2 whereas expression of ULBP3 was not affected (Fig 3E). Thus, Rh159 impairs the surface expression of most, but not all, NKG2DLs.
RhCMV Rh159 shares 30% sequence identity with HCMV UL148 [34], comparable to other RhCMV proteins demonstrated to be functional HCMV protein homologues [26, 27] (Fig 4A). To determine whether HCMV UL148 would target NKG2DL expression, we transduced each of the U373-NKG2DLs with a previously described adenovector expressing UL148 including a C-terminal V5-epitope tag [16]. However, in contrast to Rh159, MICB maturation was not affected by UL148 even at high MOI (Fig 4B) and MICB surface expression remained unchanged after transduction with Ad148 (Fig 4C). Similarly, cell surface levels of MICA, ULBP1, ULBP2 and ULBP3 were not affected by UL148 (Fig 4C). Since UL148 expression was verified by immunoblot (Fig 4D), we conclude that despite sequence and positional homology, UL148 and Rh159 diverge in their ability to target NKG2DL. While we cannot rule out that UL148 targets a NKG2DL not tested here, previous reports showed increased NKG2DL surface expression upon infection with HCMV lacking UL16 but containing functional UL148 [16]. In addition, no decrease in NK cytotoxicity was found when cells transduced with Ad148 were co-incubated with human NK cells when compared to control [35] consistent with our conclusion that HCMV UL148 is not an NK cell evasion factor despite common ancestry with Rh159.
To further demonstrate that Rh159 is responsible for ER-retention of MICB, we created a Rh159 knock out virus (ΔRh159) by replacing Rh159 with SIVgag using BAC mutagenesis (S2A and S2B Fig). Replacement of Rh159 with SIVgag had only a modest impact on viral growth in tissue culture (S2C Fig) and the viral genome was stable upon multiple passages as shown by full genome sequencing (S2D Fig). To determine the contribution of Rh159 to NKG2DL downregulation by RhCMV we compared NKG2DL surface levels in U373-NKG2DL cells infected with RhCMV or ΔRh159. Interestingly, both MICA and MICB surface levels were significantly higher in cells infected with ΔRh159 compared to RhCMV (Fig 5A). In contrast, only a minor increase of ULBP2 surface levels was observed in the absence of Rh159 and, as expected, ULBP3 levels were not increased (Fig 5A). Although exogenous expression of Rh159 reduced ULBP1 surface levels (Fig 3E), infection with RhCMV in the absence of Rh159 did not lead to increased ULBP1 surface levels compared to RhCMV (Fig 5A). Since MICA, MICB, and ULBP2 surface levels were significantly higher in cells infected with ΔRh159 compared to RhCMV, we wanted to specifically address the impact of Rh159 on the maturation of these NKG2DLs when newly synthesized during infection. Therefore, we examined NKG2DL maturation by pulse-chase labeling and immunoprecipitation upon infection with ΔRh159. In contrast to RhCMV-infected U373-MICB cells, MICB acquired EndoH resistance upon infection with ΔRh159 at the same rate as observed for uninfected cells (Fig 5B) suggesting that intracellular transport of MICB was no longer inhibited. Also note that the co-precipitating viral protein seen in Figs 2B and 5B is absent in ΔRh159-infected cells. Similarly, MICA maturation was unimpeded upon infection with ΔRh159 (Fig 5B) consistent with the increased MICA surface expression compared to RhCMV (Fig 5A). However, ULBP2 was still retained by ΔRh159 despite the fact that AdRh159FL downregulated these proteins (Fig 3E). The slight increase of ULBP2 in ΔRh159-infected cells (Fig 5A) was thus not due to lack of ULBP2 retention in the absence of Rh159. Taken together, these results suggest that RhCMV encodes additional proteins targeting ULBPs. In contrast, Rh159 seems to be predominantly, if not solely, responsible for the retention of MICA and MICB by RhCMV.
To verify that deletion of Rh159 would similarly impact the expression of RM NKG2DLs we generated telomerized RM fibroblasts (TRF) transduced with lentivectors expressing RM MIC1, RM MIC2, RM ULBP1, RM ULBP2, and RM ULBP3 (Fig 6A). Similar to the human ligands, we observed that all RM NKG2DLs were downregulated by RhCMV and that expression of RM MIC2, the equivalent of MICB, was restored upon Rh159 deletion (Fig 6A). Similar to human ULBP2, we observed an increase of RM ULBP2 whereas the MICA equivalent, MIC1, surface levels remained low in the absence of Rh159 and was only marginally increased compared to RhCMV (Fig 6A). Thus, for both human and RM NKG2DLs deletion of Rh159 predominantly increased MICB/MIC2 consistent with Rh159 being predominantly responsible for MIC2 retention whereas additional RhCMV proteins likely prevent MIC1 and ULBPs surface expression. Consistent with MIC2 upregulation, ΔRh159 also increased expression of endogenous rhesus MICs compared to RhCMV (Fig 6B). Moreover, we observed increased binding of soluble human NKG2D receptor to ΔRh159-infected versus RhCMV-infected TRF (Fig 6C). To directly determine whether increased MIC2 expression by ΔRh159 impacted NK cell stimulation, we infected primary fibroblasts derived from three individual macaques with RhCMV or ΔRh159 and monitored stimulation of purified, autologous NK cells at 48 hpi. Macaque NK cells were magnetically sorted from PBMC using antibodies to NKG2A with the majority of NKG2A+ NK cells expected to co-express NKG2D [36]. Following co-incubation with autologous fibroblasts for 4h, NK cell activation was measured by CD107a (lysosomal-associated membrane protein 1), a surrogate marker for degranulation [37]. NK cells efficiently recognized the MHC-I-deficient human cell line K562, which was used as a positive control (Fig 6D). In contrast, NK cells responded less to autologous fibroblasts infected with RhCMV when compared to uninfected K562 cells or to fibroblasts infected with ΔRh159 (Fig 6D). Differential NK cell stimulation was not due to different percentages of cells infected as shown by IE staining (Fig 6D). These observations demonstrate that Rh159 significantly limits NK cell stimulation by RhCMV-infected cells.
To determine whether decreased NK cell evasion impedes the ability of ΔRh159 to overcome pre-existing immunity during super-infection, as we reported for T cell evasion [28], we inoculated a CMV-positive animal with ΔRh159. Since the SIVgag protein was used to replace Rh159 in this construct, we monitored the T cell responses to SIVgag as a surrogate marker for infection. As shown previously, as little as 100 PFU of RhCMV encoding SIVgag are sufficient to elicit SIVgag-specific T cell responses in RhCMV-positive animals [28]. Surprisingly, however, even at a dose of 5x106 PFU, ΔRh159 did not induce SIVgag-specific T cell responses upon sub-cutaneous inoculation (Fig 7A). Since T cell responses are a sensitive measure of infection, these results suggested that Rh159 is either required to overcome pre-existing immunity (as shown previously for the MHC-I downregulating genes Rh182-189 (US2-11) [28]) or Rh159 is required for infection regardless of the immune status of the recipient. To determine whether Rh159 was required for primary infection we inoculated two CMV-naïve RM with the same dose of ΔRh159 and monitored the T cell response to SIVgag as well as to RhCMV IE. However, neither SIVgag nor IE-specific T cells were detected in either of the two CMV-naive RM at any time post-infection (Fig 7B). In contrast, wild type RhCMV never failed to induce SIV-specific T cell responses at comparable doses in hundreds of RM inoculated to date [21, 23, 25, 28]. Furthermore, SIVgag-specific T cells are observed regardless of the promoter driving gag expression, e.g. replacing Rh189, which results in SIVgag expression levels that are lower than that of Rh159 in vitro (S2 Fig) resulted in robust SIVgag-specific T cell responses in vivo [25]. These results thus suggest that RhCMV was unable to establish infection in the absence of Rh159. To determine whether increased clearance by NK cells prevented infection by ΔRh159, we wanted to monitor infection under conditions that temporarily eliminate NK cells. Since in RM all NK cells express CD8, depletion of CD8+ cells eliminates both CD8+ T cells and NK cells [38]. During the first days of infection, CMV-naïve animals lack CMV-specific CD8+ T cells and, for this reason, evasion of CD8+ T cells is not required for primary infection [28]. Therefore, CD8-depletion was used to temporarily eliminate NK cells and T cells on day 63 after the initial inoculation with ΔRh159 (Fig 7C). Re-inoculation of both RM with ΔRh159 resulted in SIVgag-specific and IE-specific CD4+ and (with some delay) CD8+ T cell responses (Fig 7B). These responses remain significantly above background levels to date and are likely to persist for the life of the 2 RM. Since depletion of non-specific CD8+ T cells was unlikely to impact infection by ΔRh159, these data strongly suggested that elimination of NK cells permitted infection by ΔRh159; and, therefore, that NK cell evasion by Rh159 is essential for primary infection.
We recently reported that RhCMV (strain 68–1) elicits unconventional CD8+ T cell responses in RM that recognize peptides presented by MHC-E and MHC-II and includes “supertopes”, i.e. peptides recognized by CD8+ T cells from MHC-disparate animals [24, 25]. To determine whether deletion of Rh159 would affect CD8+ T cell specificity we monitored the CD8+ T cell response to supertopes (presented by either MHC-E or MHC-II). Upon recovery of CD8+ T cells each of the SIVgag-derived supertope peptides stimulated CD8+ T cells from ΔRh159-infected RM suggesting that ΔRh159 maintained the unconventional T cell targeting phenotype of the parental RhCMV strain (Fig 7B).
Both Rh159 and HCMV UL16 share the ability to downregulate multiple NKG2DLs, including MICB. Since UL148 does not target NKG2DLs and since RhCMV does not encode a sequence homolog of UL16, we hypothesized that, although Rh159 is the sequence homologue of UL148, it is functionally more closely related to UL16. To test this hypothesis we replaced the coding region of Rh159 with that of UL16 by BAC mutagenesis of a recombinant RhCMV that contains an SIVgag expression cassette in Rh211 [21] thus generating ΔRh159/UL16R (S3 Fig). The BAC was fully sequenced and virus recovered in RM fibroblasts expressed UL16 (S3 Fig). To determine whether the HCMV protein was able to downregulate RM NKG2DLs in the context of RhCMV, we infected RM fibroblasts with ΔRh159/UL16R and monitored the surface expression of MIC2. As shown in Fig 8A, ΔRh159/UL16R prevented the upregulation of endogenous MIC proteins and reduced surface levels of transfected MIC upon infection, whereas endogenous MIC was upregulated and transfected MIC2 was no longer downregulated in ΔRh159-infected cells compared to RhCMV (Fig 6B). To examine whether ΔRh159/UL16R would be able to infect RM despite the lack of Rh159 we inoculated two RhCMV seropositive RM with a low dose 3x104 PFU of ΔRh159/UL16R and monitored the T cell response to SIVgag. Remarkably, both animals developed robust T cell responses to SIVgag suggesting that UL16 can substitute the immune evasion function of Rh159 (Fig 8B). Furthermore, using UL16-specific peptides previously used to map HCMV-ORF-specific T cell responses [39] we were able to detect a UL16-specific CD8+ T cell response when examining one of the RMs at 70dpi (Fig 8C). Although there is no sequence homology between UL16 and Rh159, the ability of both proteins to prevent the induction of NKG2DL, particularly that of MICB and MIC2, seems to promote CMV infection. These data thus further support our conclusion that preventing the induction of NKG2DL is essential for infection and strongly suggest that it is the NK cell evasion function of Rh159 that is required for infection.
Our data indicate that limiting NKG2DL expression on infected cells is essential for RhCMV infection. The conservation of most NKG2DLs between human and RM, together with the fact that UL16 can substitute for Rh159, renders it highly likely that HCMV will similarly depend on NK cell evasion to establish infection. Since each NKG2DL can engage NKG2D such stringent NK cell control creates an enormous selective pressure to counteract each activating ligand. Consequently, multiple NKG2DL-inhibitory gene products are expressed by HCMV, with UL16 being the quintessential NKG2DL-inhibitor acting on multiple NKG2DLs. In contrast to its host targets however, the NKG2DL-inhibitor UL16 is not conserved between HCMV and RhCMV. We identified Rh159, the homologue of HCMV UL148, as a broadly acting NKG2DL inhibitor. Similar to UL16, Rh159 is an ER-resident type I glycoprotein that retains MICA, MICB, ULBP1, and ULBP2 in the ER but not ULBP3. A tyrosine-based motif in the cytoplasmic domain of UL16 was implicated in ER-retrieval [9] whereas the luminal domain is thought to interact with NKG2DLs [11]. However, so far it has not been possible to demonstrate the formation of a stable complex between NKG2DL and UL16, or any other cytomegaloviral NKG2DL-inhibitor. In contrast, Rh159 was identified in co-immunoprecipitations with MICB consistent with a remarkably stable complex. The predicted type I topology of Rh159 suggests that its ectodomain interacts with NKG2DLs whereas the short, carboxyterminus (KRSREAH) is predicted to expose an RXR ER-retrieval motif [40] responsible for ER-retention of Rh159 and consequently NKG2DLs. However, the role of individual protein domains in Rh159 function has yet to be confirmed experimentally.
Interestingly, HCMV UL148 does not seem to interfere with NKG2DL expression despite clear homology to Rh159 suggesting that the two genes diverge functionally despite common ancestry. Homologues of UL148 are found in all primate CMVs but not in rodent CMVs, whereas UL16 and UL142 are only found in human and chimpanzee CMVs. Since the NKG2DL system predates primate evolution we speculate that the UL148 ancestor most likely evolved to counteract NKG2DL, a function that is conserved in non-human primate CMVs, but substituted with UL16 in human and ape CMVs. In HCMV, US18 and US20 complement UL16 by targeting MICA as well as UL142, which in addition targets ULBP3. Similarly Rh159 did not interfere with ULBP3 expression suggesting that additional RhCMV proteins are responsible for the reduction of ULBP3 surface expression observed in RhCMV-infected cells. Moreover, only human MICA and MICB as well as RM MIC2 were robustly upregulated in ΔRh159-infected cells suggesting that Rh159 is the only gene involved in MIC2 retention whereas Rh159 likely shares the ability to downregulate other NKG2DLs with additional RhCMV proteins. Possible candidate genes for additional MIC1 targeting are Rh199 and Rh201, the RhCMV homologs of HCMV US18 and US20 that were shown to degrade MICA via the lysosome [16]. In addition, RhCMV encodes about 20 ORFs not conserved in HCMV [26] and it is possible that some of these genes encode additional NKG2DL inhibitors. Further work will be required to identify these additional NKG2DL inhibitors and it will be interesting to determine whether upregulation of NKGDLs in their absence would lead to activation of NK cells in vitro and stringent control of CMV infection by NK cells in vivo as observed for ΔRh159.
The absolute requirement for MIC2 downregulation by Rh159 to establish infection upon sub-cutaneous inoculation was unexpected because deletion of the NKG2DL-inhibitors m138, m145, m152 or m155 from MCMV reduces viral titers, but does not prevent infection of CMV-naïve mice [2]. In fact, replacement of m152 with the murine NKG2DL Rae-1 increased CD8+ T cell responses to MCMV [19, 20]. This is in stark contrast to the complete absence of T cell responses observed in CMV-naïve animals inoculated with ΔRh159. We consider it unlikely that ΔRh159 established a low level infection that would not have elicited a T cell responses since CMV-specific T cell responses do not require viral dissemination or spreading from initially infected cells [41, 42]. Thus, the most likely explanation for the lack of SIVgag-specific and RhCMV IE-specific T cell responses in RM infected with ΔRh159 is that NK cells rapidly eliminated infected cells due to MIC2 induction. This conclusion is supported by the induction of SIVgag-specific and RhCMV-specific T cells by ΔRh159 upon treatment with CD8-specific antibodies or upon replacement of Rh159 with UL16, which restored MIC downregulation and the ability of ΔRh159 to infect RM. Treatment with anti-CD8 monoclonal antibody cM-T807 effectively removes both CD8+ T cells and NK cells from circulation [38]. However, it seems unlikely that T cell depletion permitted infection with ΔRh159, since CMV-naïve animals lack CMV-specific CD8+ T cells at the time of cM-T807 treatment. Moreover, we reported previously that evasion of CD8+ T cells is only required to establish secondary infections in CMV-immune animals, but not for primary infection [28]. Thus, although cM-T807 eliminates both T cells and NK cells, CD8+ NK cells were most likely responsible for the suppression of ΔRh159 infection. Interestingly, the magnitude and duration of the T cell response elicited by ΔRh159 did not diminish upon recovery of CD8+ cells suggesting that neither NK cells nor T cells were able to eliminate ΔRh159 once a persistent infection has been established. Similarly, we previously reported that RhCMV lacking MHC-I evasion genes infected CMV-positive animals upon CD8+ cell depletion and maintained a persistent infection [28]. Taken together these data suggest that both NK and T cells prevent viral spreading from the initial sites of infection, but neither innate nor adaptive cellular responses can eliminate latent viral reservoirs and clear persistent infection, even when NK or T cell immune evasion is incomplete. We thus expect to observe life long effector memory T cell responses in the anti-CD8 treated animals despite full recovery of NK and T cell responses. A possible explanation for this finding is that latent virus escapes NK and T cell clearance either by being immunologically silent or by using immune evasion mechanisms that differ from the ones used during lytic infection and viral dissemination.
In addition to inhibiting NKG2DL surface expression demonstrated here, Rh159 was previously shown to enhance infection of epithelial cells (EC) [43]. Specifically, it was demonstrated that disruption of the Rh159 gene by transposon insertion reduced growth of RhCMV (strain 68–1) in RM retinal ECs by an unknown mechanism. Thus, Rh159 seemed to mediate EC growth independently of the major determinant of EC entry, the pentameric complex gH/gL/UL128/130/131, because 68–1 lacks the UL128 and UL130 subunits [44]. Interestingly, it was reported that HCMV UL148 affects the ratio of the trimeric (gH/gL/gO) and the pentameric complex in viral particles [45]. In the absence of UL148, maturation of gH/gL/gO was impaired thus enhancing the amount of pentameric complexes and infection of ECs. Whether Rh159 similarly affects viral glycoprotein maturation is currently unknown, but it is conceivable that Rh159 might impact the maturation not only of host glycoproteins but also of viral glycoproteins. We recently reported that the spontaneous deletion of the homologs of UL128 and UL130 in RhCMV 68–1 results in the exclusive induction of unconventional CD8+ T cells recognizing peptides in the context of MHC-II or MHC-E [24, 25]. This unexpected function of pentameric complex components in modulating CD8+ T cell responses could thus potentially be further modified by Rh159. However, in CD8+-depleted animals infected with ΔRh159, the CD8+ T cell targeting phenotype was the same as that of the Rh159-intact parental strain 68–1, i.e. CD8+ T cells were restricted by MHC-E and MHC-II, but not by MHC-I, once CD8+ T cells recovered from depletion. We have not yet determined whether deletion of Rh159 would impact conventional CD8+ T cell responses by UL128-130-intact RhCMV. Thus, Rh159 might play multiple roles in vivo that range from NK cell inhibition during primary infection to affecting cell tropism to modulating adaptive T cell responses during persistent infection.
All cell lines were cultured in Dulbecco's modified Eagle medium (DMEM) supplemented with 10% fetal bovine serum (FBS), 100 IU of penicillin/ml and 100 μg of streptomycin/ml and incubated at 37C and 5% CO2 except where noted. K562 cells were obtained from ATCC. 293A cells were obtained from Invitrogen. 293-CRE expressing cells were a kind gift from Ashley Moses (Oregon Health and Science University). Telomerized RM fibroblasts (TRFs), primary cynomologus fibroblasts and NK cells were obtained from animals housed at Oregon National Primate Research Center. U373-NKG2DLs were previously described [31]. TRF-RM NKG2DLs were generated using the pLVX lentivector system (Clontech) by cotransfecting pLVX, along with vectors encoding vesicular stomatitis virus G (VSV-G) (pMD2.G; Addgene 12259) and Gag/Pol (psPAX2; Addgene 12260), into TRF cells using Lipofectamine LTX (Life Technologies) according to the manufacturer’s protocol. Supernatant containing lentivirus was harvested from the transfected cells 48 h posttransfection, and used to transduce TRF-RM NKG2DLs in the presence of 5 μg/ml Polybrene (hexadimethrine bromide; Sigma-Aldrich). This process was repeated 24h later, and the resulting cell lines were grown in the presence of 3 μg/ml puromycin to select for cells that expressed the viral genes. The RM NKG2DLs were synthesized (GenScript) based upon the following accession numbers: AF055387 (MIC1), AF055388 (MIC2), XM_001082270.2 (RM ULBP1), XM_001082656.1 (RM ULBP2), and XM_001083203.2 (RM ULBP3; since the deposited RM sequence lacked a stop codon we inserted nine nucleotides (5’ GGCAGATGA 3’) at the appropriate position based on their conservation in multiple non-human primate ULBP3 sequences). Synthetic genes were cloned into pcDNA3.1(-), and subcloned into pLVX containing an eF1α promotor. All constructs were synthesized to encode an HA epitope tag on the amino terminal end, after the predicted signal sequence (SignalP 4.1).
The following primary antibodies were used for immunoprecipitation, flow cytometric and/or immunoblot analysis of U373-NKG2DLs: mouse anti-MICA (AMO1), -MICB (BMO2), -ULBP1 (AUMO2), -ULBP2 (BUMO1), and -ULBP3 (CUMO3) (Axxora), -TfR (M-A712, BD Pharmingen), -GAPDH (6C5), -CD44 (DF1485) and -V5 (E10) (Santa Cruz), -RM MIC (a kind gift from Thomas Spies[46]), -FLAG (M2) and polyclonal rabbit anti-FLAG (Sigma), and polyclonal goat anti-ULBP3 (AF1517), -ULBP2 (AF1298) and -ULBP1 (AF1380) (R&D systems). For isotype controls, mouse IgG2a (02–6200) or IgG1 (MG100) (Life Technologies) were used. Chicken anti-mouse Alexa Fluor 647 (Life Technologies) and streptavidin PE-Cy7 (eBioscience) were used as secondary antibodies in FACs experiments. Goat anti-mouse IgG-HRP and donkey anti-goat IgG-HRP (Santa Cruz) were used for immunoblotting. Total surface NKG2DLs in TRFs were measured using human NKD2D-Fc chimera (R&D Systems) followed by anti-human IgG-PE (eBioscience). Human Fc-G1 (BioX-Cell) was used as isotype control. Streptavidin-APC (eBioscience) was used for intracellular secondary staining. For the NK activation assay NKG2A-PE (Z199, Beckman-Coulter) and anti-PE beads (Miltenyi Biotech) were used to magnetically sort NKG2A+ cells from PBMC, and subsequently stained after co-culture with autologous fibroblasts using anti-CD107a-FITC (H4A3), CD8-APC-Cy7 (SK1), IFN-γ-APC (B27) (BD Biosciences), yellow Live/Dead fixable stain (Invitrogen) and NKG2A-PE.
RhCMVΔRh159gag (ΔRh159) was created by BAC-recombineering [47] replacing Rh159 with an expression cassette for SIVgag in RhCMV 68–1 [48]. Briefly, primers containing 50 bp homology to regions flanking Rh159 (forward primer 5’GGTCGTTTGGTTGTTTCTCACCTATTGCTTGGTACTCTAGCT TCAGTAAG3′ and reverse primer 5’TAGTTTATAAACACACAATCACGTGGTGGT ACTGTGAACCCGCGTCGGTA-3′) were used to amplify SIV mac239 gag and a kanamycin resistance (KanR) cassette flanked by FRT sites from plasmid pCP015 and electroporated into EL250 bacteria containing the RhCMV 68–1 BAC for in vivo recombination. Upon removal of the KanR gene by FLP recombination, recombinant virus was reconstituted in primary RM fibroblasts. Self-excision of the loxP-flanked BAC-cassette in the resulting virus was confirmed by full genome sequencing.
In a two step process RhCMV ΔRh159/UL16R was generated by en passant mutagenesis [49] to precisely replace the RhCMV Rh159 coding region with that of HCMV UL16. In the first step we inserted a DNA fragment containing an I-SceI restriction site and a Kanamycin resistance cassette into UL16 of HCMV strain TR BAC [50] by homologous recombination using the following flanking primers:
Forward primer 5’CCAGCCGCATGGTCACTAATCTTACCGTGGGCCGTTATGACTGTTTACGC TGCGAGAACGGTACGATGAAAATAATCGAGCGCCTCCACGTCCGATTGGG TAGGGATAACAGGGTAATAAG-3’
Reverse primer 5’CGGAGGGGTGTTTGGCGAGCCCGGATCCGGGCGGTCTCGGATATAGCGAG CCCAATCGGACGTGGAGGCGCTCGATTATTTTCATCGTACCGTTCTCGCA AGAGCGCTTTTGAAGCTGG-3’.
Next, we amplified the UL16 gene containing the I-SceI site and Kanamycin resistance cassette by PCR with primers containing 50 bp homology sequences flanking Rh159:
Forward primer: 5’GGTCGTTTGGTTGTTTCTCACCTATTGCTTGGTACTCTAGCTTCAGTAAG ATGGAGCGTCGCCGAGGTACG-3’
Reverse primer: 5’GTTTATAAACACACAATCACGTGGTGGTACTGTGAACCCGCGTCGGTATC AGTCCTCGGTGCGTAACC3-’).
In the second step, the PCR fragment was transformed into E.coli strain GS1783 [49] harboring RhCMV 68-1/gag and Rh159 was replaced with the UL16/I-SceI/KanR fragment by homologous recombination and Kanamycin selection. Next, we removed the I-SceI/KanR cassette by inducing red recombinases and the I-SceI restriction enzyme. Cleavage of the BAC-DNA by the I-SceI enzyme followed by red-dependent recombination between two homologous sequences flanking the I-SceI/KanR fragment (shown in bold in above primer set) results in the precise excision of the I-SceI/KanR cassette and the restoration of the intact HCMV UL16 ORF. The resulting ΔRh159/UL16R BAC was characterized by next generation sequencing of the entire BAC.
To generate AdRh159FL, Rh159 was amplified from viral DNA isolated from TRFs infected with RhCMV using the following primers: 5' CGGGATCCCGCCACCATGGCCTACAACAG-3' and 5’GGAATTCCTTACTTATCGTCGTCATCCTTGTAGTCATGAGCTTCACGACTGCGTT-3’. The PCR fragment was cloned into pADTet7, and transfected into 293-CRE cells with psi5 helper virus [51]. RhCMVΔRh189gag and RhCMV68-1/gag were previously described [25]. Adenovirus expressing UL148 and empty vector control were a kind gift from Richard Stanton (Cardiff University) [52].
U373-NKG2DL cells were lysed in PBS 1%NP-40 and HALT protease inhibitor (Thermo Fisher) followed by SDS-PAGE protein separation and transfer to PVDF. Immunoblotted proteins were detected with ECL2 (Thermo Scientific). For Pulse-chase experiments, U373-NKG2DL cells were starved for 1 h in DMEM minus cysteine (cys) and methionine (met) supplemented with 10% FBS (pulse media). After starve period, EXPRE35S35S Protein Labeling Mix (PerkinElmer) was added to pulse media at 150 μCi/106 cells for 30 min. Cells were chased for the indicated times then lysed, digested, and immunoprecipitates were separated by SDS-PAGE and detected by autoradiography. Digestions using EndoH (Roche) or PNGaseF (NEB) were performed according to manufacturer’s instructions.
Cells were harvested, washed in 3% FBS/PBS, incubated with primary antibody followed by fluorophore-conjugated secondary antibody. Cells were then fixed in 1% paraformaldehyde, permeabilized in 0.1% triton-X100/PBS, washed and incubated with biotinylated anti-IE2 antibody (DBX biotin labeling kit, Molecular probes) followed by fluorophore-conjugated streptavidin. Flow cytometry data were acquired on LSR II (BD Biosciences) and analyzed with FlowJo X (v.10.0.7, Tree Star).
NKG2A+ cells sorted from PBMC were plated overnight in RPMI-1640, 15% FBS, 100 IU/ml of IL-2, and 10 ng/ml of IL-15 (NK media). NKG2A+ cells (2.5 x 106/ml) were incubated with autologous fibroblasts (1.25 x 105 IE-2 expressing cells/ml) for 30 min at 37°C in the presence of anti-CD107/FITC. Brefeldin A (10 μg/ml) and GolgiStop (1 μl/ml; BD Biosciences) were added after 30 min and samples were incubated at 37°C for 8 hrs. Cells were surface stained, fixed with 2% paraformaldehyde, and then permeabilized for intracellular IFN-γ staining with 1x PBS containing 10% FBS and saponin (1 g/L).
Proteins excised from Coomassie gels were trypsinized and peptides were analyzed by LC/MS-MS using an LTQ Velos Pro linear ion trap (Thermo Scientific) to collect data-dependent MS/MS data [53]. Sequest (version 28, revision 12) was used to search MS2 Spectra against a March 2012 version of the Sprot human FASTA protein database, with added sequences from the Uniprot Rhesus Cytomegalovirus and concatenated sequence-reversed entries to estimate error thresholds and 179 common contaminant sequences and their reversed forms. Database processing was performed with python scripts (ProteomicAnalysisWorkbench.com). SEQUEST results were filtered to strict peptide and protein false discovery rates (FDRs), estimated from the number of matches to sequence-reversed peptides, using PAW software [54]. Independent FDR control was performed, resulting in a 4.3% FDR for protein discovery (43 identifications to forward sequence proteins and 2 to reverse-sequenced proteins). 9 MS/MS spectra assigned to forward sequence peptides and none to reverse-sequence peptides from any entry in the database.
Five male, purpose-bred, Indian-origin RM (Macaca mulatta) were used: three animals being RhCMV-positive and two were specific-pathogen free, including RhCMV. ΔRh159 or ΔRh159/UL16R was inoculated subcutaneously at 5 × 106 PFU or 3 x 104 PFU, respectively. For CD8+ cell depletion, RM were treated with 10, 5, 5 and 5 mg/kg of the anti CD8 mab cM-T807 one day prior to inoculation with 5 × 106 PFU ΔRh159 and on days 2, 6, and 9 post inoculation, respectively. Antigen-specific CD4+ and CD8+ T cell responses were measured in mononuclear cell preparations from blood by ICCS [21]. Briefly, sequential 15-mer peptides (overlapping by 11 amino acids) comprising SIVMAC239gag or RhCMV IE or four core (optimal) SIVgag supertope peptides (2 each MHC-E- and MHC-II-restricted) [25]) were used in the presence of co-stimulatory CD28 and CD49d monoclonal antibodies (BD Biosciences). To detect UL16-specific responses we used a previously described overlapping peptide pool [39]. Cells were incubated with pooled peptides and co-stimulatory molecules for 1h, followed by addition of Brefeldin A (Sigma-Aldrich) for an additional 8h. Co-stimulation without antigen served as a background control. Cells were then stained with fluorochrome-conjugated monoclonal antibodies, flow cytometric data collected on a LSR II (BD Biosciences) and data analyzed using the FlowJo software program (version 8.8.7; Tree Star). Responses frequencies (CD69+/TNFα+ and/or CD69+/IFNγ+) were first determined in the overall CD4+ and CD8+ T cell population and then memory corrected (with memory T cell subset populations delineated on the basis of CD28 and CD95 expression).
IACUC Project Number 0691, Protocol number IS00002845, and started 05/01/09 The Five purpose-bred male RM (Macaca mulatta) of Indian genetic background were used with consent of the Oregon National Primate Research Center Animal Care and Use Committee, under the standards of the US National Institutes of Health Guide for the Care and Use of Laboratory Animals.
|
10.1371/journal.pgen.1002173 | Molecular Mechanisms Generating and Stabilizing Terminal 22q13 Deletions in 44 Subjects with Phelan/McDermid Syndrome | In this study, we used deletions at 22q13, which represent a substantial source of human pathology (Phelan/McDermid syndrome), as a model for investigating the molecular mechanisms of terminal deletions that are currently poorly understood. We characterized at the molecular level the genomic rearrangement in 44 unrelated patients with 22q13 monosomy resulting from simple terminal deletions (72%), ring chromosomes (14%), and unbalanced translocations (7%). We also discovered interstitial deletions between 17–74 kb in 9% of the patients. Haploinsufficiency of the SHANK3 gene, confirmed in all rearrangements, is very likely the cause of the major neurological features associated with PMS. SHANK3 mutations can also result in language and/or social interaction disabilities. We determined the breakpoint junctions in 29 cases, providing a realistic snapshot of the variety of mechanisms driving non-recurrent deletion and repair at chromosome ends. De novo telomere synthesis and telomere capture are used to repair terminal deletions; non-homologous end-joining or microhomology-mediated break-induced replication is probably involved in ring 22 formation and translocations; non-homologous end-joining and fork stalling and template switching prevail in cases with interstitial 22q13.3. For the first time, we also demonstrated that distinct stabilizing events of the same terminal deletion can occur in different early embryonic cells, proving that terminal deletions can be repaired by multistep healing events and supporting the recent hypothesis that rare pathogenic germline rearrangements may have mitotic origin. Finally, the progressive clinical deterioration observed throughout the longitudinal medical history of three subjects over forty years supports the hypothesis of a role for SHANK3 haploinsufficiency in neurological deterioration, in addition to its involvement in the neurobehavioral phenotype of PMS.
| Terminal chromosome deletions are among the most commonly observed rearrangements detected by cytogenetics and may result in several well-known genetic syndromes. We used 22q13 deletions to study how these types of chromosome abnormalities arise. Children with Phelan/McDermid syndrome, caused by deletion of the terminal portion of chromosome 22, experience developmental delay, absent or severely delayed speech, and frequent behavioral problems. Lack of one copy of SHANK3, a key gene for the correct development and organization of brain synapses, is very likely the basis of the syndrome's major neurological features. Deletion of additional genes probably causes more complex phenotypes in subjects with larger deletions. We also studied patients who only lack a portion of SHANK3 and demonstrated that small, hard-to-detect deletions of this gene may cause substantial clinical problems. Until now, the 22q distal deletion had been only diagnosed in very young people. We studied a large group of patients of different ages and discovered that all adult patients face progressive cognitive decline. Our study demonstrates that deletion of the terminal portion of chromosome 22, a prototype for terminal deletions in human chromosomes, can occur in several ways. Mosaic deletions of different size can also form in early embryogenesis.
| Deletions involving the distal portion of chromosomes are among the most commonly observed rearrangements detected by cytogenetics [1] and result in several well-known genetic syndromes such as 1p36 monosomy (MIM: 607872), Cri-du-chat (5p-, MIM: 123450), Miller-Dieker (17p-, MIM: 247200), monosomy 18q (18q-, MIM: 6011808) monosomy 9p (MIM: 158171), Wolf-Hirschhorn (4p-, MIM: #194190), 9q34.3 microdeletion (MIM: 610253), monosomy 2q37 (MIM: 600430) and Phelan-McDermid (PMS, MIM: 606232) syndromes. Over the past 15 years, technological advances in the molecular cytogenetic diagnosis of mental retardation, such as subtelomere screening and high-resolution genome analysis, have strongly enhanced the detection rate of an increasing number of chromosome rearrangements involving subtelomeric regions associated with mental retardation.
Telomere loss caused by double-strand breaks (DSBs) can generate, if not properly repaired, chromosome instability, cell senescence, and/or apoptotic cell death. Terminal deletions can be repaired and stabilized through the synthesis of a new telomere (telomere healing), demonstrated through sequence analysis of terminal deletions that showed de novo telomeric repeats attached to the remaining chromosomal sequences [2]–[4]; by telomerase-independent recombination-based mechanisms [5], [6]; by obtaining a telomeric sequence from another chromosome (telomere capture) resulting in derivative chromosomes [7], [8]; finally, by chromosomal circularization, leading to the formation of a ring chromosome [9], [10]. However, in spite of their relatively frequent occurrence, the molecular bases for generating and stabilizing terminal chromosome deletions in humans are still poorly understood, since the breakpoints have been analyzed at the base-pair level in only few studies [11], [12]. Questions remain about the timing of breakpoint repair, the relative importance of the above-mentioned mechanisms in terminal deletions affecting specific chromosomes, the role of repetitive elements, long terminal repeats and other DNA elements in chromosome breakage and stabilization.
In this study, we used deletions of 22q13, which represent a substantial source of human pathology [13], [14], as a model for investigating the molecular mechanisms of terminal deletions.
We characterized at the molecular level 40 new and 4 previously published subjects with 22q13 chromosome rearrangements [15], [16] aiming to identify the molecular mechanisms involved in stabilizing the deletions in patients with monosomy 22q13 and, more generally, to obtain new insight in the mechanisms underlying terminal deletions. Genotype-phenotype relationship, including the detailed clinical history of three adult patients that may help to define the lifelong outcome of PMS, is also discussed.
Patients included 26 females and 18 males, with ages ranging from birth to 47 years. Six patients (P25–29, P33) had a ring chromosome 22, five (P37–38, P42–44) had interstitial 22q13.3 deletions, three (P11, P15, P16) carried derivative chromosomes, while the remaining patients had terminal deletions (Table 1).
We excluded from the clinical analysis patients with a derivative chromosome 22 (P11, P15, P16) and subject P28 with a complex ring 22 rearrangement, since the additional duplicated regions could complicate the assessment.
The features observed in the 40 cases in our series were compared to the characteristic features of the 22q13 deletion syndrome [17] (Table S1).
The parental origin of the de novo 22q13 rearrangements was elucidated in 30 families (Table 3).
The majority of terminal (17/23) and interstitial (2/2) deletions for which parental origin was available had paternal origin. Three of five ring 22 cases (60%) were also of paternal origin, while two were maternal.
We collected 40 new unrelated patients with 22q13 deletions and re-analyzed four previously published cases (Table 1).
Nine subjects (P2, P8, P10, P14, P17, P20, P36, P40, P41) showed a 22q13 deletion on high- resolution G- banding karyotype (550 bands); in three of them, previous low resolution banding karyotype had missed the rearrangement. Six cases showed a ring 22 at karyotype analysis. One of them (P29) was a mosaic. In one subject (P31), the presence of a terminal 22q13.3 microdeletion was first suspected in a routine subtelomere screening by multiplex ligation-dependent probe amplification (MLPA, kit P036, MRC Holland) that showed a possible deletion at the RABL2B locus, and subsequently diagnosed by aCGH analysis (244k, Agilent). Subtelomeric FISH screening with cosmid clones covering the distal 22q-140 kb [18] (data not shown) further confirmed the terminal 22q13.3 microdeletion with breakpoint between exons 8–9 of the SHANK3 gene. The remaining twenty-four patients had normal karyotype results and were ascertained either through subtelomere-FISH or array-CGH analysis.
Whole-genome array-CGH using several available platforms (44k, 105k, 244k) was performed on all patients diagnosed through classical cytogenetic methods, except for subject P35, in order to determine the genomic size of the deletion and exclude any concurrent microdeletion/microduplication elsewhere in the genome.
This approach allowed the identification of 22q13 deletions, varying in size between 0.14 and 9.0 Mb, in 39 subjects (Figure 1). The breakpoints were scattered along the 22q13 region and no breakpoint grouping was observed. We precisely delineated the boundaries of each deletion by commercial high-resolution (244k, Agilent) or customized aCGH analysis. Further improvements in resolution, obtained with subject-specific qPCR amplification experiments, allowed the design of oligonucleotide primers to specifically amplify the junction fragments.
We attempted to clone the deletion breakpoints of all 33 patients with apparently terminal 22q13 deletions by postulating healing of the truncated 22q sequences through the addition of a new telomere sequence at the breakpoint. Forward primers were designed proximally to each breakpoint and used for nested PCR, together with telomere-specific primers. Using this strategy, we isolated twenty-two breakpoints from 20 cases with terminal deletions (P1, P3–P8, P12–P14, P20, P21, P30–P32, P34, P36) (Figure S2). Nineteen breakpoints from 17 subjects contain 3–48 copies of the GGTTAG hexamer. Alignment of the chromosome-specific sequences flanking the telomeric repeats with the human genome reference sequence revealed the immediate proximity of the repeats to the chromosome-specific sequences in 16 breakpoints. Three breakpoints (P20 BP3, P8, P7) contain 2, 14, and 20 additional bases not present in the reference sequence, respectively. Two subjects (P31, P32) carry recurrent 22q terminal deletions [18]. The junction fragment in subject P8 contains a perfect 7-bp inverted palindrome. Thirteen of the 19 breakpoints fall inside repetitive sequences (SINE, LINE, DNA-type, simple repeats) (Figure S2). One breakpoint junction (P2) contains a (GGTGAG)n repeat, fortuitously amplified because of its homology with the Tel-ACP primer, instead of the expected (GGTTAG)n. In a second junction (P39), the telomere sequence is preceded by (GGTCAG)6. A third breakpoint (P40) is joined to the terminal 450 bp of a Xp/Yp chromosome arm.
Interestingly, high-resolution aCGH analysis allowed the identification of a patient (P20) carrying a mosaic of at least three lines with 22q13.2 terminal deletions, each with a different breakpoint (Figure 2A). All breakpoints were located in a ∼400 kb interval. FISH analysis with clone RP11-141N8 (AQ388763 at 22q13.2), positioned between BP1 and BP2 (Figure 2A, 2B) confirmed the presence of a mosaic deletion in 30% of the metaphases analyzed (Figure 2B). We cloned all three identified breakpoints: the more proximal is located in intron 11 of the EFCAB6 gene; the intermediate falls in a MER5B repeat; the more distal in a Tigger5 repeat (Figure 2C). High-resolution aCGH profiling suggested the presence of at least two mosaic breakpoints in a second patient (P12) (Figure S3A, S3B), but we were only able to clone one of them (Figure S3C).
Our series also includes five patients with interstitial 22q13.3 deletions disrupting the SHANK3 gene (P37, P38, P42–44) (Table 2, Figure 3A, 3B). The region distal to the deletions in P38 and P42 lies in a paralogous sequence containing the RABL2B gene, with almost complete identity with the chromosome 2 region containing RABL2A, and only one 180k/244k (Agilent) aCGH probe covers it. Subtelomere FISH screening with cosmid probes spanning the terminal 100 kb of distal 22q (data not shown) and qPCR experiments confirmed the findings. Specific amplification of the junctions by long-range PCR followed by sequencing analysis precisely defined each rearrangement's structure (Figure 3C, Figure S2).
The 74 kb interstitial deletion in P37 encompasses exons 1–17 of SHANK3; the 44 kb deletion in P38 covers exons 19–23 of SHANK3 and the whole ACR gene; the 18 kb deletion in P42 includes exon 23 of SHANK3 and exons 1–3 of ACR; the 27 kb deletion in P43 includes exons 20–23 of SHANK3 and exons 1–3 of ACR; the 34 kb deletion in P44 overlaps exons 1–9 of SHANK3 (Figure 3B). We found no homology between any proximal and distal breakpoint region. Repeated sequences (LTR and LINE) are present in three breakpoints; ten additional nucleotides were inserted at the junction of P37 (Figure S2). In P42, the breakpoint junction contains 23–29 bps identical to the reverse complement of a sequence in the middle of the deleted region; this sequence shows 4 and 2 bp microhomologies with the proximal and distal breakpoints, respectively (Figure 3C).
Six subjects (P25–P29, P33) carry a 22q13 terminal deletion associated to ring 22 chromosome; one of them (P29) shows a mosaic deletion in 30% of the cells (not shown). We also identified a complex ring 22 rearrangement consisting of a 240 kb terminal 22q deletion, concurrent with two additional, non-contiguous, ∼18 Mb and ∼4.2 Mb 22q duplications at 22q11–q12.3 and 22q12.3–q13.2, respectively, in subject P28 (Figure S4).
The only breakpoint we were able to identify in a patient with ring 22 (P26) (Figure S5A) was cloned using inverse PCR and shows a junction between an Alu sequence on 22q and a repeated sequence with homology to pericentromeric and subtelomeric regions on several chromosomes (Figure S5B). There is no homology between the two breakpoints. In this case, as well as in cases P27 and P29, we verified the absence of interstitial pan-telomeric sequences with a PNA probe (Figure S5C). Unfortunately, FISH analysis could not be performed on the remaining three cases (P25, P28, P33) due to the lack of archival material.
Three patients (P11 and brother/sister pair P15–16) (Cases 1, 2, and 3, respectively, in [15]) carry a derivative chromosome 22 inherited from a parent carrier of a balanced translocation.
In case P11, aCGH analysis identified the loss of a 5 Mb segment of distal 22q13.31–q13.3 and the gain of a 5.7 Mb region of chromosome 12q24.32–q24.33 (Figure S6A); the proband's father carries a balanced 12q;22q translocation. We amplified the junction between 12q24.32 and 22q13.31 by long-range PCR using a forward primer (22F) from the der(22) undeleted flanking region and a reverse primer (12R) corresponding to the 12q duplicated region. The same fragment was amplified from the carrier father but not from the mother or other control DNAs (not shown). Sequencing of the junction fragments revealed that the two breakpoints share only a 4 bp microhomology (Figure S6B). Similarly, sibs P15 and P16 both inherited the der(22) chromosome from their mother who carries a balanced 12q;22q translocation. The two sibs carry a 4.3 Mb 22q13 deletion and a 0.5 Mb 12q24.33–q24.33 duplication (Figure S6C). In these patients, the rearrangement is between an Alu repeat on chromosome 22q and a (TGAG)n simple repeat on chromosome 12q. The two breakpoints share only a 5-bp microhomology (Figure S6D).
The constitutional 22q13 deletion is a fairly recently described genomic disorder that results in global developmental delay, delayed/absent speech, hypotonia and minor dysmorphic features. In spite of the fact that to date more than 100 cases (excluding ring 22s) have been detected by different molecular methods, when and how terminal deletions arise is still poorly understood. We have characterized from the clinical and molecular points of view 44 subjects with PMS resulting from simple 22q13 deletions (30 subjects, 72%), ring chromosomes (six subjects, 14%), unbalanced translocations (three subjects, 7%) and interstitial deletions (five subjects, 9%); all rearrangements result in haploinsufficiency of the SHANK3 gene (Table 1). We have also determined the breakpoint junction sequences of twenty subjects with terminal deletions, five with interstitial deletions, one with ring 22 and three with unbalanced translocations.
Although in our cases age at diagnosis ranged from birth to 41 years, no specific clinical phenotype diagnostic for 22q13 deletion could be identified at any age (Table S1), as already noted by Phelan et al. [13]. Thus, successful diagnosis of this syndrome depends almost exclusively on the use of molecular diagnostic tools, mainly subtelomeric FISH and high-resolution genome-wide array-CGH. The latter is also suitable to identify cryptic interstitial deletions involving only the SHANK3 gene, that are associated with an even less specific phenotype, as observed in our five patients (P37, P38, P42–P44). Their phenotype consisted mainly of developmental and language delay. PMS-suggestive facial dysmorphisms and hypotonia (limited to abdominal muscles) were observed only in one patient (P37), while no other physical abnormalities were noted in any of the patients. (Table 2). In patient P43, a defect in the abdominal wall with gut protrusion was detected by ultrasound during pregnancy and surgically corrected immediately after birth.
Owing to its emerging role in neuropsychiatric disorders and to the phenotypic overlap between autism and PMS, SHANK3 has become a target for mutation screening in patients with autistic spectrum disorders (ASD) and several studies [19]–[21] have discovered de novo mutations in such patients. Mutations in SHANK3 have also been found in schizophrenia [22] and non-syndromic intellectual disability [23]. The contribution of additional genes to the 22q13 deletion phenotype has also been debated. Very recently it has been proposed that deletion of the IB2 gene (also named MAPK8IP2 or JP2), mapping 70 kb proximal to SHANK3, may play a relevant role in PMS-associated ASD [24].
Two of our patients with interstitial microdeletions disrupting SHANK3 and ACR only (P38, P43) (Figure 3B) fulfill the clinical criteria for a diagnosis of autism, while the others (P37, P42, P44), do not (Table 2). Our findings emphasize the incomplete penetrance of the ASD phenotype in PMS, while confirming a role for SHANK3 in ASD. Additional deleted genes may contribute more strongly to accessory features, such as dysmorphisms and hypotonia, than to developmental and language delay.
Longitudinal clinical data on adult patients were collected in three subjects aged 40, 41 and 47 years. The severe progressive neurological deterioration reported in adult patients P10 (starting when she was 39 years old) and P30 (aged 40 years) was also described by Anderlid in a 30-year-old patient [25]. The minimal overlapping 22q13 region deleted in these three cases contains only SHANK3, RABL2B and ACR. In addition, subject P37 carrying an interstitial deletion involving only SHANK3 experienced tremors and tics starting at age 23 (Text S1). Shank proteins, that organize glutamate receptors at excitatory synapses, are dramatically altered in Alzheimer disease [26]. In turn, disruption of glutamate receptors at the postsynaptic platform had been reported to contribute to the destruction of the postsynaptic density underlying mental deterioration in Alzheimer disease [27]. According to our results, SHANK3 defects might indeed be responsible for progressive neurodegeneration, in addition to causing the neurobehavioral phenotype of the 22q13 syndrome.
Previous studies on a large cohort of patients with ring 22 demonstrated that there is considerable molecular and phenotypic overlap between individuals with ring 22 and those with del 22q13 [28]–[29]. All six subjects reported here showed features commonly found in 22q13.3 deletion syndrome, including accelerated growth in two of them (P26, P27), whereas one (P25) had slightly delayed growth.
Parental origin was determined in 30/44 cases. We observed a larger proportion of 22q13 deletions of paternal (22/30, 73%), compared to maternal (8/30, 27%) origin, in agreement with a previous large study in which 69% of the deletions were of paternal origin [14]. There was no deletion size bias. Interestingly, we observed that both recurrent deletions (P31, P32), as all previously reported cases [18], [19], [30], were of paternal origin. Furthermore, the two interstitial deletions (P37, P38) we characterized were also paternal. In other terminal deletion cohorts, the majority of patients carry small 1p36 deletions on the maternal chromosome, while larger deletions are predominantly paternal [31]. In contrast, de novo simple small terminal 9q34.3 deletions are predominantly paternal, whereas larger terminal deletions, interstitial deletions, complex rearrangements and unbalanced translocations are frequently maternal in origin [12].
Broken chromosome ends can be stabilized through at least three mechanisms: de novo telomere addition mediated by telomerase; telomere capture resulting in a derivative chromosome; stabilization by break-fusion-break (BFB) cycles, generating terminal deletions and proximal inverted duplications. The first two mechanisms have been identified in this study. Almost 60% of our patients carried apparently simple terminal deletions. Nineteen of the twenty-two breakpoints we cloned, including all three breakpoints in mosaic subject P20, show evidence of telomere healing. Fourteen breakpoints contain 1–5 base microhomologies with the canonical GGTTAG sequence at the fusion point of genomic and telomeric sequences (Figure S2), possibly reflecting the template-driven mechanism that telomerase uses to replicate chromosome ends [32], [33]. The same mechanism applies to terminal 4p deletions [34] where microhomology with telomere repeats was found in all analyzed subjects. Human telomeres contain large blocks of 100–300 kb TAR sequences, located just proximally to the (TTAGGG)n tandem repeats, providing significant sequence homology between non-homologous chromosome ends [35]. Two terminal deletions in our cohort (P2, P39) were healed by telomere capture of TAR sequences (Figure S2). One deletion (P40) was repaired by the capture of the distal portion of Xp/Yp (Figure S2).
Failure to identify the nine remaining breakpoints may be due to the presence of regions containing large repetitive sequences or other complex sequences that would hinder amplification. Alternatively, some of the deletions may lack a telomere repeat at the breakpoint because the deletion may have been repaired by alternative mechanisms.
The presence of short repetitive elements may play a role in generating or stabilizing terminal deletions. In this study, we have determined 42 breakpoint junctions within the 22q13 region. Repetitive sequences, such as Alu, LINE, SINE, LTR and simple repeats were often, but not always, present at or near the breakpoints (Table S2). These repetitive elements are susceptible to DSBs due to replication errors or to the formation of unusual secondary structures, including cruciforms, hairpins, and tetraplexes [36]. On the other hand, there is no hard proof that the breakpoints of terminal deletions are the actual site of the original DSB, rather than the site where telomerase was able to synthesize a new telomere sequence.
Analysis of case P20 revealed a mosaic of at least three cell lines carrying different terminal 22q13 deletions. Their breakpoints were located approximately 100 Kb from each other. This mosaicism may be due to distinct stabilizing events, occurring in different cells of the early embryo, of the same unstable terminal deletion. Our results demonstrate that primary terminal deletion breakpoints and repair sites are not necessarily coincident and can actually be far apart. We had already shown, in an exceptional case of mosaicism for maternal 22q13.2-qter deletion (45% of cells) and 22q13.2-qter paternal segmental isodisomy (55% of cells) that complex mosaicism can also arise from a postzygotic or early embryonic recombination event [16]. These data suggest that terminal deletions can be repaired by multistep healing events. Cryptic mosaics may also render genotype-phenotype relationship in deletions more complex than expected.
Deletion sizes in patients with monosomy 1p36 [31] and 9p21–p24 [37] vary widely, up to 20 Mb, while 9q34.3 deletions [12] do not exceed 3.5–4 Mb. The size of 22q13 deletions is highly variable, ranging from 100 kb to 9 Mb [14]. No single common breakpoint has been discovered in deletions of 1p36 [31] and 9q34.3 [12], both studied in great detail. In contrast, 9 cases with terminal 140 kb deletion and a breakpoint occurring in a short GC-rich simple repeat in intron 8 of the SHANK3 gene have been reported [18], [19], [25], [30], [38], [39]. In this study, we detected two new unrelated cases (P31, P32) with the same recurrent terminal deletion healed by de novo telomere addition. Computational analysis [40] predicts that this repeat would be able to form a secondary structure that may predispose to DNA double strand breaks, stabilize the broken chromosome end, or recruit telomerase more efficiently [36]. Subject P39 has a slightly larger deletion repaired by the capture of a TAR sequence.
Interstitial deletions affecting the 22q13 region have previously been described in three cases, one disrupting the SHANK3 gene [41] and two more proximal [42]; none of them has been finely characterized at the molecular level. We characterized five additional de novo interstitial deletions between 17 and 74 Kb in size and sequenced their breakpoints: three of the deletions (P37, P44) disrupt exclusively SHANK3, the others (P38, P42, P43) both SHANK3 and ACR (Figure 3B).
The interstitial deletions in four patients (P37,38, P42, P44) are compatible with NHEJ repair. P42 carries a more complex rearrangement where 40–47 bp from the deleted region are inserted in opposite orientation in the middle of the breakpoint junction (Figure 3C). A DNA replication model named FoSTeS [43], later generalized to the microhomology-mediated break-replication (MMBIR) model [44], has been proposed to explain complex rearrangements associated with several diseases. The rearrangement in P42 can indeed be explained by the FoSTeS/MMBIR mechanism (Figure 3D).
Apart from the cases described in this report, we have no information on the percentage of defects in SHANK3 caused by deletions/duplications involving only one or a few exons. The small size of these rearrangements poses substantial problems for their identification, at least with current commercial aCGH platforms having necessarily limited coverage of the SHANK3 gene. Arrays designed for the detection of clinically relevant exonic CNVs [45] may offer a solution.
NHEJ is the most likely repair mechanism leading to ring 22 formation in case P26. As this is the only ring 22 breakpoint we were able to clone, we cannot be sure that the same mechanism will apply to all cases with ring 22. Our inability to capture more breakpoints of ring 22 deletions may stem from the occurrence of the 22p breakpoints within highly repetitive sequences. Generation of breakpoints at both arms of the same chromosome, followed by circularization, has been usually assumed to be the basis of ring chromosome formation. Alternatively, telomere healing through circularization after the occurrence of a simple distal deletion, as it seems the be the case for ring chromosomes with concurrent deletion and duplication at one end [10], cannot be excluded. Thus, distal deletions and ring chromosomes might share the same initial event.
We also demonstrated that the complex phenotype in one ring 22 patient (P28) can be explained by the presence of further chromosome duplications at 22q11–12.3 and 22q12.3–13.2, undetectable with conventional cytogenetic analysis, in addition to the 22q13.3 deletion. The identification of the complex ring 22 rearrangement in this patient directly stems from the whole-genome aCGH analysis required by our protocol in order to exclude additional genomic aberrations.
All unbalanced translocations we analyzed (P11, P15, P16) were inherited from a parent carrying a balanced translocation. The microhomology found at all breakpoints points to NHEJ or MMBIR as the most likely mechanisms for these rearrangements; therefore they should be considered mechanistically different from all previously discussed chromosome 22 rearrangements.
All adult patients with 22q13 deletion showed progressive clinical deterioration, supporting the hypothesis of a role for SHANK3 haploinsufficiency in neurological deterioration. All patients with interstitial deletions involving only SHANK3 showed a neurological and behavioral phenotype, demonstrating once again the specific role of the gene in this syndrome.
The study of breakpoints in subjects with 22q13 deletion provides a realistic snapshot of the variety of mechanisms driving non-recurrent deletion and repair at chromosome ends, including de novo telomere synthesis, telomere capture and circularization. Distinct stabilizing events of the same terminal deletion can also occur in different early embryonic cells. These data are in agreement with those demonstrating that mosaic structural chromosome abnormalities are common in early IVF embryos [46] and that chromosomally unbalanced zygotes are submitted, during first mitotic divisions, to intense genomic reshuffling eventually leading to different situations, all compatible with survival [47]. As recently suggested, the burst of DNA replication that accompanies the rapid cell division required to go from a single post-zygotic cell to an embryo and then a fetus is a time in the human life cycle when more new mutations may occur than was previously appreciated. Depending on the timing, many such events may be difficult, if not impossible, to identify at the DNA level [48].
The study was approved by the Ethical Committee at the “Eugenio Medea” Scientific Institute.
Blood samples were obtained from probands and their parents after informed consent. All patients were referred for genetic evaluation to different medical centers because of developmental delay, delayed/absent language and dysmorphic features. Physical examination and review of medical and family history records were performed on each patient. The diagnosis of terminal 22q13 deletion syndrome had not been proposed in any of the patients before identification of the deletion by cytogenetic or molecular diagnostic analysis. Cytogenetic and molecular diagnosis had been obtained by conventional karyotyping, subtelomere FISH, 22q13 MLPA analysis, or oligonucleotide-based aCGH (44k, 105k, 180k or 244k Agilent platforms)(Table 1).
A very high-resolution 22q13 custom array was designed using the eArray software (http://earray.chem.agilent.com/); probes were selected among those available in the Agilent database (UCSC hg18, http://genome.ucsc.edu). A total of 24624 probes were selected within the distal 9.4 Mb region of 22q13 (chr22: 40269203–49565875), and 8660 probes within the distal ∼3.2 Mb of chromosome 12q (chr12: 129000012–132289374); the latter set was used to identify the breakpoint interval in cases with a derivative chromosome 22 associated with a 12q genomic segment (P11, P15, P16), and for quality control/normalization. The probes provided an average resolution of 400 bp. Genomic DNA was isolated from blood samples using the GenElute-Blood kit (Sigma). Gender-matched genomic DNAs were obtained from individuals NA10851 (male) and NA15510 (female) (Coriell). The quality of each DNA was evaluated by conventional absorbance measurements (NanoDrop 1000, Thermo Scientific) and electrophoretic gel mobility assays. Quality of experiments was assessed using Feature Extraction QC Metric v10.1.1 (Agilent). The derivative log ratio spread (DLR) value was calculated using the Agilent Genomics Workbench software. Only experiments having a DLR spread value <0.30 were taken into consideration.
Metaphase chromosomes and interphase nuclei were obtained from all patients and their parents from PHA-stimulated blood lymphocyte cultures. G-banding karyotypes at 400–550 bands resolution were performed using standard high-resolution techniques. FISH experiments with 22q13.3 subtelomeric cosmids n66c4 (AC000050), n85a3 (AC000036), n94h12 (AC0020556) and n1g3 (AC002055) [18] were performed to confirm the aCGH results in cases where the 22q13.3 deletion disrupted the SHANK3 gene (P31–32 P37–38, P42–43). FISH analysis with BAC clones, labeled with biotin-dUTP (Vector laboratories, Burligame, CA) using a nick translation kit (Roche), or probes for all subtelomeric regions (TelVysion kit, VYSIS) were performed on selected cases. The pan-telomeric peptide nucleic acid (PNA) probe (PNA FISH kit/Cy3, Dako Denmark A/S) which recognizes the consensus sequence (TTAGGG)n of human pan-telomeres was hybridized according to the manufacturer's instructions.
Hybridizations were analyzed with an Olympus BX61 epifluorescence microscope and images were captured with the Power Gene FISH System (PSI, Newcastle-upon-Tyne, UK).
Genotyping of polymorphic sequence-tagged sites (STS) was performed by amplification with primers labeled with fluorescent probes followed by analysis on an ABI 310 Genetic Analyzer (Applied Biosystems). In cases where STS analysis was not informative, SNP genotyping was performed by PCR amplification followed by sequencing. All amplifications were performed with AmpliTaq Gold (Applied Biosystems) using standard protocols.
Chromosome-specific target sequences for quantitative PCR analysis were selected within non- repeated sequences using Primer Express 2.0 software (Applied Biosystems) as described in Bonaglia et al. [18]. The annotated genomic sequence of chromosome 22 (March 2006 assembly, hg18) is available through the UCSC Human Genome Browser (http://genome.ucsc.edu/). Multiplex Ligation-dependent Probe Amplification analysis (MLPA) of the 22q13 region was performed with the SALSA MLPA kit P188 22q13 (MRC-Holland, Amsterdam).
Amplification of 22q13 deletions repaired by chromosome healing was performed as in Bonaglia et al [18]. PCR products were both directly sequenced and cloned with a TOPO TA cloning kit (Invitrogen), followed by sequencing of individual clones. Inverse PCR was performed on Sau3A-cut, ligated (in 1 ml volume to facilitate self-ligation of individual fragments) genomic DNA, using nested sets of primers. Long-range PCRs were performed with JumpStart Red ACCUTaq LA DNA polymerase (Sigma) and the following protocol: 30 sec at 96°C, 35 cycles of 15 sec at 94°C/20 sec at 58°C/15 min at 68°C, 15 min final elongation time. Sequencing reactions were performed with a Big Dye Terminator Cycle Sequencing kit (Applied Biosystems) and run on an ABI Prism 3130xl Genetic Analyzer. Primer sequences are available in Table S2.
The accession number and URLs for data presented herein are as follow: UCSC Human Genome Browser, http://genome.ucsc.edu/; Online Mendelian Inheritance in Man (OMIM), http://www.ncbi.nlm.nih.gov/omim.
|
10.1371/journal.pgen.1003729 | The Relative Contribution of Proximal 5′ Flanking Sequence and Microsatellite Variation on Brain Vasopressin 1a Receptor (Avpr1a) Gene Expression and Behavior | Certain genes exhibit notable diversity in their expression patterns both within and between species. One such gene is the vasopressin receptor 1a gene (Avpr1a), which exhibits striking differences in neural expression patterns that are responsible for mediating differences in vasopressin-mediated social behaviors. The genomic mechanisms that contribute to these remarkable differences in expression are not well understood. Previous work has suggested that both the proximal 5′ flanking region and a polymorphic microsatellite element within that region of the vole Avpr1a gene are associated with variation in V1a receptor (V1aR) distribution and behavior, but neither has been causally linked. Using homologous recombination in mice, we reveal the modest contribution of proximal 5′ flanking sequences to species differences in V1aR distribution, and confirm that variation in V1aR distribution impacts stress-coping in the forced swim test. We also demonstrate that the vole Avpr1a microsatellite structure contributes to Avpr1a expression in the amygdala, thalamus, and hippocampus, mirroring a subset of the inter- and intra-species differences observed in central V1aR patterns in voles. This is the first direct evidence that polymorphic microsatellite elements near behaviorally relevant genes can contribute to diversity in brain gene expression profiles, providing a mechanism for generating behavioral diversity both at the individual and species level. However, our results suggest that many features of species-specific expression patterns are mediated by elements outside of the immediate 5′ flanking region of the gene.
| DNA sequence variation underlies many differences both within and between species. In this paper, we investigate a specific DNA sequence that is thought to influence expression of a gene that modulates behavior, the vasopressin V1a receptor gene (Avpr1a). Specifically, differences in the expression of V1a receptor in the brain have been causally tied to social behavior differences, but the genetic basis of these differences is not understood. Using transgenic mice, we investigate the role of DNA sequences upstream of this gene in generating species-specific and individual variation in Avpr1a expression. We find that, contrary to our expectation, this region has only a modest influence on differences in expression patterns across rodent species. This indicates that DNA elements outside of this region play a larger role in species-level differences in expression. We confirm that variation in Avpr1a expression mediated by this upstream region translates to differences in behavior. We also find that variable DNA sequences associated with repetitive motifs within this region subtly influence gene expression. Together these findings highlight the complexity of genetic mechanisms that influence diversity in brain receptor patterns and support the idea that variable repetitive elements can influence both species and individual differences in gene expression patterns.
| The genomic mechanisms that give rise to phenotypic diversity across species or among individuals within a species are not well understood. Behavior is a trait that is particularly well suited for exploring genetic mechanisms underlying phenotypic plasticity, as it is an evolutionarily labile trait. Social behaviors, in particular, can be markedly variable among closely related species, and often display significant individual variability within a species [1]–[6]. Genomic mechanisms that give rise to diversity in behavior fall into two categories; those that alter protein structure and function (e.g. coding region mutations) and those that alter the expression of genes [7], [8]. In Caenorhabditis elegans, for example, variation in a single nucleotide of npr-1, which alters the neuropeptide receptor protein structure, has been shown to be responsible for strain differences in social feeding behavior [9]. However, it is likely that a significant portion of phenotypic diversity is derived from mutations that alter gene expression [10]–[13]. Sequences in the 5′ flanking region of genes regulate tissue-specific expression in many cases, and are thus likely candidates for contributing to species-specific expression patterns. In addition, unstable, polymorphic repetitive elements surrounding genes have been proposed as a mechanism to enhance evolvability of traits by increasing diversity in gene expression [14], [15]. The vasopressin 1a receptor gene (Avpr1a) provides an excellent opportunity to explore both of these potential mechanisms of gene expression divergence [16], [17].
Arginine vasopressin (AVP) is an evolutionarily conserved neuropeptide that modulates a wide range of behaviors including stress coping, territorial aggression, mate-guarding, pair bonding and paternal care [18]–[21]. The vasopressin 1a receptor (V1aR) is a G-protein coupled receptor that mediates many of the behavioral effects of AVP [22]. While the structure and brain distribution of AVP are highly conserved among mammals, the behavioral effects of this peptide, and the neural distribution of V1aR vary markedly across species [23]–[25]. Among voles, for example, AVP facilitates affiliative behavior and selective aggression related to pair bonding in monogamous prairie voles (Microtus ochrogaster), but not in the closely related, non-monogamous montane voles (M. montanus) [24], [26]. Accompanying these species differences in behavioral response to AVP are remarkable species differences in V1aR distributions in the brain. For example, monogamous prairie voles have higher densities of V1aR in the ventral pallidum, central amygdala, and dentate gyrus than nonmonogamous montane or meadow voles (M. pennsylvanicus) [1], [27], [28]. These differences in V1aR distribution are due to species differences in the expression of the Avpr1a gene [26]. Furthermore, pharmacologically blocking ventral pallidal V1aR prevents mating-induced partner preference formation in prairie voles [29]. These observations suggest that variation in the neural expression patterns of Avpr1a may underlie species differences in AVP-dependent behaviors.
There is substantial direct evidence supporting the hypothesis that diversity in expression of Avpr1a within the brain contributes to both intra-and inter-specific differences in behavior. For example, increasing V1aR density in the lateral septum using viral vector mediated gene transfer enhances social recognition memory in rats while increasing V1aR density in the ventral pallidum of prairie voles facilitates affiliation and pair-bond formation [30], [31]. In addition, increasing V1aR density in the anterior hypothalamus increases selective aggression in prairie voles [32]. Relatively subtle variation in expression can profoundly affect behavior since viral vector mediated RNA interference in the ventral pallidum, which results in a 30% reduction in V1aR binding, significantly reduces pair bonding behavior in prairie voles [33]. Remarkably, increasing V1aR in the ventral pallidum of meadow voles using a viral vector to mimic the distribution of V1aR in the prairie vole brain confers the ability to form a partner preference in this promiscuous species [34]. Likewise, transgenic mice carrying the prairie vole Avpr1a locus display a pattern of V1aR binding similar to that of prairie voles, and this difference in receptor patterns leads to increased affiliative behavior in response to AVP [24]. These experiments demonstrate conclusively that diversity in Avpr1a expression within the brain directly contributes to both inter and intra-species variability in AVP-mediated behaviors. Here we explore the contribution of both the 5′ flanking region and variation in hypermutable microsatellite sequences within this region in the generation of this variability in gene expression.
The 2.2 kb of sequence upstream of the Avpr1a transcription start site have been hypothesized to contain regulatory sequences that contribute to the prairie vole-like patterns of V1aR [24]. Transgenic mice carrying a randomly inserted prairie vole Avpr1a transgene, comprised of 2.2 kb of 5′ flanking sequence, exons, introns and some downstream sequences from the prairie vole, displayed a receptor pattern more similar to that of a prairie vole than a mouse [24]. However, this prairie vole-like pattern was found in only one of four independently derived transgenic mouse lines carrying identical transgenes, suggesting that the integration site within the genome had a strong impact on expression pattern and raising a question as to the extent to which this region is actually responsible for species-specific expression patterns. Within this 5′ flanking region is a series of variable nucleotide tandem repeats (VNTRs) interspersed with non-repetitive DNA, known as the Avpr1a microsatellite. This microsatellite lies ∼760 bp upstream of the Avpr1a transcription start site and displays polymorphisms in repeat numbers and sequence content across and within vole species, and thus represents a “hot spot” for mutations (Figure 1) [17], [35].
Both species and individual differences in the Avpr1a microsatellite are sufficient to drive differences in gene expression in a cell-type specific manner in vitro, suggesting that this genetic region may represent a source of both inter- and intraspecies receptor expression variation [17], [35]. This hypothesis is further supported by associations between microsatellite length, receptor patterns, and social behavior in prairie voles [17], [36]. Specifically, in a laboratory setting, male prairie voles selectively bred to have long Avpr1a microsatellites were more likely to form partner preferences than males with short microsatellites [17]. Furthermore, long and short Avpr1a microsatellite prairie voles have different patterns of V1aR distribution in the brain [35]–[37]. However, this selection experiment cannot distinguish between the contribution of the microsatellite and other linked functional genetic variants that affect expression. Remarkably, similar polymorphisms in AVPR1A microsatellites have been associated with gene expression, brain activation, and social behavior in humans and chimpanzees [38]–[41]. Thus the vole Avpr1a is an ideal model locus for exploring the genomic mechanisms contributing to diversity in brain gene expression and behavior that has relevance to human behavior. More specifically, Avpr1a provides an opportunity to explore the relative contribution of species-specific regulatory elements in the 5′ flanking region and of polymorphic repetitive elements in that region in generating species-specific patterns and individual variation in brain gene expression.
We hypothesized that replacing 3.4 kb of the 5′ flanking region of the mouse Avpr1a gene with the prairie vole homologue would result in a prairie vole-like pattern of V1aR binding. We further hypothesized that variation in the microsatellite element in this region would confer variation in receptor distribution. To test these hypotheses, we used homologous recombination to create three lines of knock-in mice in which 3.4 kb of the mouse 5′ flanking region was replaced with the corresponding prairie vole sequence. Each line differed only with regard to the microsatellite element – either from meadow vole or the prairie vole long or prairie vole short variants that had previously been associated with individual variation in V1aR distribution. Using receptor autoradiography, we assessed the contribution of the 5′ flanking region for determining species-specific V1aR expression patterns and microsatellite variability in generating variation in V1aR expression patterns in vivo in these three lines of mice. Our results demonstrate that both inter- and intra-species variability in the microsatellite confers differences in receptor binding in the thalamus, amygdala, and dentate gyrus, mirroring naturally-occurring differences observed both between and within vole species. However, the 5′ flanking region is not sufficient to confer species-typical binding patterns, but is sufficient to quantitatively change expression levels in a direction consistent with species differences in binding. Based on these results, we determined whether the observed differences in expression led to behavioral differences, and found that alterations in receptor binding are associated with differences in coping strategy in the forced swim test but not with differences in learning and memory in the novel object recognition task.
We used recombinant transgenic technology to replace the 5′ flanking region of the mouse Avpr1a gene with corresponding sequence from the prairie vole. We chose to replace 3.4 kb because this was larger than the 2.2 kb previously used to generate a traditional transgenic mouse through pronuclear injection [24] and contained a high density of low sequence homology (53.6% identity between mouse and prairie vole) while still being small enough for efficient homologous recombination (Figure 1a). We hypothesized that this sequence divergence would contain the elements that confer species-specific expression patterns. In order to also investigate the hypothesis that microsatellite variation within this region may contribute to species and individual differences in expression patterns, we generated 3 lines that differed only in the content of one of three Avpr1a microsatellites – either the meadow version, or a long or short version from the prairie vole (Figure 1b) - within the same prairie vole 5′ flanking region. The meadow microsatellite cassette was 175 bp long, and the short and long prairie alleles were 608 and 623 bp long, respectively.
The targeting strategy is illustrated in Figure 2. For the meadow line, we screened 288 ES cell clones and identified 1 recombinant. For the prairie short line, 192 clones yielded 2 correct recombinants, and for the prairie long line, 288 clones yielded 2 correct recombinants. This corresponds with an overall recombination efficiency of 0.6% (5 of 768). The floxed PGK-NeoR cassette was successfully removed via breeding to a ubiquitously expressing EIIa-Cre recombinase line as confirmed by PCR and Southern Blot (Figure 2). Because the Acc651 site used to screen for recombinant stem cells was located within the floxed region, excision of the NeoR also resulted in the recombinant allele yielding a ∼9.5 kb band when detected with the external probe. The three resulting recombinant alleles, prairie vole long (pvKI-long), prairie vole short (pvKI-short), or meadow vole (mvKI) were identical in sequence except for the composition of the microsatellite element. All three lines were backcrossed to a C57Bl/6J background for at least 5 generations prior to neuroanatomical and behavioral experiments.
Because we performed recombination in hybrid B6/129 ES cells, there was a possibility that recombination could occur at either the C57Bl6/J or the 129SvEv locus. In order to determine the integration site for our three lines, we genotyped rs13480799, which is located outside of the 5′ homology arm upstream of the Avpr1a gene. This SNP is a G in most lines examined, including C57-related lines, but is a C in 129-related lines [42]. Sequencing revealed that the targeting construct recombined in the C57Bl6/J allele in the mvKI line, and into the 129SvEv alleles in both pvKI lines. While this represents a potential confound that should be taken into account when considering our results, C57Bl6/J and 129SvEv strains differ very little at this locus. When comparing C57-related (C57Bl6/J and C57L/J) and 129-related (129S1/SvImJ and 129X1/SvJ) strains within 100 kb surrounding the Avpr1a locus (Chr10:121850000–121950000; NCBI37/mm9), only 5 known SNP differences (rs29315655, rs29348001, rs13480799, rs29342115, rs633704) and 1 unresolved potential difference (rs29379744) have been described in the JAX Mouse Genome Informatics SNP database [42], [43]. 28 SNPs have been described across all strains for this region. Thus while it is possible that our line differences could be attributable to the strain origin of the locus of recombination, it is unlikely because these mouse strains are so similar in this region.
Previous work had suggested that some of the elements integral to species-specific neural V1aR patterns existed within the 2.2 kb upstream of the transcription start site of the Avpr1a gene [24]. However, independently derived lines of transgenic mice carrying this region displayed different expression patterns due to differences in chromosomal integration of the transgene. Furthermore, those transgenes also contained coding regions, introns and 3′ flanking sequences. In order to more precisely explore the role of the 5′flanking region in guiding species-specific V1aR patterns, we compared V1aR binding (as a proxy for Avpr1a expression) in wildtype (WT) and pvKI-long littermates at post-natal day (PND) 60–70. Because the endogenous mouse 5′ flanking region was replaced with prairie vole sequence, this technique was not subject to random integration effects, as occurs in traditional pronuclear injection transgenics.
We used V1aR autoradiography as a proxy for Avpr1a gene expression since this technique is much more quantitative and sensitive than in situ hybridization, provides greater anatomical resolution than qPCR, and accurately reflects Avpr1a mRNA patterns (Young 1997). Furthermore, since the replaced region lies upstream of the transcription start site, variation in that sequence should not affect post-transcriptional processing. The greater signal to noise ratio of this technique allows us to detect relatively subtle differences in V1aR protein binding. Replacement of the 5′ flanking region of the murine Avpr1a locus with the same region from prairie voles yielded qualitative patterns of V1aR with elements of both mouse and prairie vole expression (Figure 3). To initially explore the effects of our manipulation on V1aR levels, we performed an overall ANOVA with three factors: genotype (WT, mvKI, pvKI-short, and pvKI-long), brain region, and sex. We identified main effects of genotype (F(3, 137) = 35.7, p<0.001), brain region (F(4, 137) = 1035.1, p<0.001), and sex (F(1, 137) = 10748.9, p<0.001). In addition, there were genotype×brain region (F(12, 137) = 8.902, p<0.001) and genotype×sex interactions (F(3, 137) = 3.451, p = 0.02), but no brain region×sex (F(4, 137) = 2.0, p = 0.09) or genotype×brain region×sex (F(12, 137) = 1.425, p = 0.16) interactions. Therefore, we did not analyze sex differences for each brain region in each of the lines. The main effect of sex appeared to be driven by the fact that females tend to have slightly higher levels of V1aR binding in some brain regions. However, since there are equal numbers of males and females across groups, and our focus was on the impact of promoter elements on expression, we collapsed males and females into a single group. We then performed three separate ANOVAs to test the a priori hypotheses regarding 1) the role of the 5′ flanking region, 2) species differences in the microsatellite, and 3) intraspecies differences in the microsatellite. To address the first of these, we compared pvKI-long mice homozygous for this region with WT mice. pvKI-long mice showed an increase in V1aR binding in the ventral pallidum (VP), central amygdala (CeA), paraventricular nucleus of the thalamus (PVthal), and dentate gyrus (DG) of the hippocampus, consistent with the distribution in the prairie vole. Specifically, we compared V1aR binding in pvKI-long mice to WT mice, and identified a significant effect of both brain region (F(4, 100) = 511.6, p<0.001) and genotype (F(4, 100) = 79.2, p<0.001) on V1aR levels. In addition, there was a significant interaction between brain region and genotype (F(4, 100) = 10.6, p<0.001), and simple main effects with Sidak-adjusted α showed that pvKI-long animals had significantly higher levels of V1aR than WT mice in the VP (p<0.001), CeA (p<0.001), PVThal (p<0.001), and DG (p<0.001), but not in the lateral septum (LS; p = 0.24; Figure 3). While it is intriguing that there is a significant difference in binding in the VP of these two lines, the actual difference is quite modest (on average, 1.15 fold higher). In addition, it should be noted that the binding in the dentate gyrus was distributed differently between prairie voles compared to pvKI-long mice, potentially due to different effects of the mouse versus vole coding sequences on receptor trafficking. However, the pvKI-long mice did not show the prairie vole specific binding in the laterodorsal thalamus (LDthal) or medial amygdala (MeA) (Figure 3). The overall similarities in binding pattern between the pvKI-long and WT mice demonstrate that elements outside of the replaced element (e.g. distal 5′ flanking regions, introns, and other surrounding elements) contribute significantly to species-typical expression, perhaps more so than the sequences in the 3.4 kb regions that we tested. However, the quantitative differences between these two lines of mice demonstrate conclusively that the proximal 3.4 kb of the 5′ flanking region also contributes to species-specific patterns of V1aR distribution in the brain.
Having established that replacement of the 5′ flanking region contributes to differences in V1aR levels in the thalamus, amygdala, ventral pallidum, and hippocampus, we next investigated whether differences in the composition of the Avpr1a microsatellite might mediate species differences in V1aR binding within any of these regions. Specifically we compared V1aR binding in KI mice homozygous for either the prairie long (pvKI-long) or meadow vole (mvKI) Avpr1a microsatellite, and identified a significant effect of both brain region (F(4, 77) = 114.7, p<0.001) and genotype (F(4, 77) = 160.9, p<0.001) on V1aR levels. In addition, there was a significant interaction between brain region and genotype (F(4, 77) = 74.2, p<0.001), and simple main effects with Sidak-adjusted α showed that pvKI-long animals had significantly higher levels of V1aR than mvKI animals in the CeA (p<0.001), PVThal (p = 0.002), and DG (p<0.001), but not in the lateral septum (LS; p = 0.68; Figure 4E) or ventral pallidum (VP; p = 0.75; Figure 4E).
Although data are not available directly comparing V1aR in these brain regions in meadow and prairie voles, meadow voles have an expression pattern that is similar to that of montane voles, and that comparison has previously been examined [27]. Table 1 shows the ratio of expression calculated for prairie∶montane voles derived from the binding values reported in table 1 from Wang et al. [44] compared with the binding ratio of the same regions in pvKI-long∶mvKI mice. These studies are independent and warrant care in drawing parallels, but overall, these ratios indicate that the binding differences between vole species are broadly mirrored in the thalamus, CeA, and DG but not the VP or LS in these mouse lines.
Allelic variation in Avpr1a has also been tied to intra-species variation in V1aR patterns in prairie voles [17], [36], [37]. However, these studies could not distinguish direct effects of the microsatellite from the possibility that the microsatellite is in linkage disequilibrium with other functional elements. In order to determine the direct contribution of the microsatellite to individual differences in neural V1aR distributions, we compared V1aR binding patterns in mice homozygous for either the long (pvKI-long) or the short version (pvKI-short) of the prairie vole microsatellite in the VP, LS, CeA, PVthal, and DG. The prairie long and short form of the microsatellite are substantially more similar to each other than to the meadow microsatellite. As such, we predicted that the potential differences conferred by this region would be relatively subtle. A 2-way ANOVA revealed a main effect of both brain region (F(4, 80) = 165.0, p<0.001) and genotype (F(4, 80) = 12.1, p = 0.001) on V1aR levels (Figure 5E). In addition, there was a significant interaction between brain region and genotype (F(4, 80) = 7.8, p<0.001), and simple main effects analysis with Sidak-adjusted α showed that pvKI-long mice had higher V1aR levels in the DG (p<0.01) but not in the CeA (p = 0.42), PVThal (p = 0.96), LS (p = 0.93) or VP (p = 0.77; Figure 5). Although significant for the DG, the differences observed between these mice are less profound than reported for prairie voles with different microsatellite lengths, suggesting that while individual differences in microsatellite structure do directly impact expression in the brain, other linked polymorphisms may account for the larger number of regional differences found in prairie voles.
While there is considerable evidence in mice and voles that variation in Avpr1a expression has behavioral consequences [24], [31], [32], [34], [45], we wanted to determine whether the variation in V1aR distribution in our KI lines contributed to variation in behavior. V1aR activation modulates a wide array of behaviors, and we used the existing literature on the role of V1aR and our binding data to guide our behavioral investigation. While variation in V1aR distribution in voles has been studied extensively with respect to social behavior, the brain regions showing line differences in our mice have not been implicated in regulating AVP-dependent social behaviors. Instead, we focused on changes in the hippocampus and the CeA — regions in which AVP and V1aR function has previously been studied in rats and mice [46]–[49].
Novel object recognition is a hippocampus-dependent task, and performance on this task is tied to differences in excitability of the dentate gyrus [50], [51]. Thus, we hypothesized that activation of V1aR in the hippocampus, which leads to increased firing rates [52], [53], might impact novel object recognition. All groups showed normal locomoter habituation upon repeated exposure to the novel object chamber (Figure 6A). A repeated measures ANOVA with the Greenhouse-Geisser F-test revealed a main effect of trial (F(3.27, 244.5) = 196.2; p<0.001) but no interaction between trial and genotype (F(9.65, 244.5) = 1.64; p = 0.097). In addition, all groups showed a preference of the novel object during the probe trial (Figure 6B), and no group differences were observed in the percent time spent investigating the novel object (one way ANOVA; F(3, 486.4) = 1.29; p = 0.30). A repeated measures ANOVA with the Greenhouse-Geisser F-test revealed a main effect of object (F(1, 22197.6) = 55.15; p<0.001) but no interaction between object and genotype (F(3, 196.4) = 0.49; p = 0.69). Post-hoc paired T-tests with Bonferroni correction indicate that all groups preferred the novel object (WT: t(37) = 5.27, p<0.001; mvKI: t(9) = 6.77, p<0.001); pvKI-short: t(11) = 3.5, p = 0.005; pvKI-Long: t(19) = 3.67, p = 0.002).
In addition, V1aR in the CeA has been shown to modulate stress coping behavior in rats and mice. In particular, swim stress elicits release of AVP into this region, and localized V1aR receptor blockade increases the amount of time rodents spend struggling in the forced swim test [46], [47]. Thus in a separate cohort of mice, we tested stress coping behavior and hypothesized that pvKI mice (both long and short), which have higher levels of V1aR in the CeA than WT and mvKI mice, would show lower levels of active coping in the forced swim test. While struggling did not differ across WT mice from the three lines (F (42) = 0.13, p = 0.88), we found that pvKI-long and pvKI-short mice struggled less in the forced swim test than did their WT littermates (Figure 6C; t = −2.35, p = 0.02), consistent with what would be expected based on pharmacological studies.
Microsatellite sequences have been hypothesized to act as evolutionary “tuning knobs,” because they mutate at faster rates than other parts of the genome, potentially due to “slippage” of the DNA polymerase while copying these highly repetitive regions. To investigate the rates of mutation in the Avpr1a microsatellite, we compared the sequence of the microsatellite region in 6th and 7th generation animals to those of the founder animals (n = 17 pvKI-long, 15 pvKI-short, and 10 mvKI microsatellite alleles from individuals born to different parents). No spontaneous mutations occurred in the intervening generations. While mutation rates are species specific, this suggests that changes in the microsatellite sequence do not occur every generation, but rather on a longer evolutionary scale, which is in accordance with previously reported mutation rates of 10−2 to 10−5 mutations per locus per generation [54].
The mechanisms underlying microsatellite-mediated differences in Avpr1a expression are not known. Transcriptional differences may ultimately depend on a combination of differences in DNA secondary structure, epigenetic characteristics, and/or differential binding of transcriptional enhancers within the microsatellite-containing region [15]. In order to gain insight into the latter, we used the transcription factor prediction software, MatInspector, to investigate the sequences shown in Figure 1, which include both short tandem repeats and interspersed non-repetitive DNA, (Genomatix, AnnArbor, MI) [55], [56]. We used Matrix Family Library Version 8.4 to match a database containing potential binding sites of 7018 vertebrate transcription factors to our sequences. MatInspector identified 21 potential TF binding sites in the meadow microsatellite, and 160 and 141 sites in the long and short allele, respectively. These sites corresponded with 21 different transcription factors potentially capable of binding to the meadow allele, and 60 for the long allele and 59 for the short allele. Comparison of these lists indicated that 4 transcription factors putatively bind the meadow microsatellite region but not the prairie alleles. In addition, we identified 5 factors that would uniquely bind to the prairie long allele, and 2 to the short allele. In order to further focus these lists, we examined their expression profiles using the Allen Brain Atlas [57], reasoning that any transcription factor responsible for differences in Avpr1a expression would need to be expressed within the brain. While most transcription factors showed at least moderate levels of expression in a few brain regions, one factor that putatively binds uniquely in the long microsatellite, Ras-responsive element binding protein 1 (Rreb1), was particularly notable because it is highly expressed within the dentate gyrus (Figure 1a). This corresponds with the differences in V1aR levels in the DG of pvKI-long versus pvKI-short mice. It should be noted that the Ras-responsive element is located in a non-repetitive sequence and is due to a G/A single nucleotide polymorphism rather than a VNTR polymorphism.
In addition, we hypothesized that differences in the number of transcription binding sites may also be important for modulating V1aR levels. We compared the number of predicted binding sites identified in the long and short allele. Among transcription factors that putatively bind both the long and short allele, 11 factors had more potential binding sites in the short allele and 8 in the long allele. The most notable differences in the number of putative binding sites were attributable to variation in length of repetitive sequences. For instance, expansion of a GAGA tetra-nucleotide repeat in the long allele generates up to 23 additional opportunities for binding of the GAGA-binding factor, cKrox/th-POK while expansion of a TATA repeat in the short allele resulted in 6 additional binding opportunities for TATA-binding factors (Figure 1a). These analyses provide potential new avenues of research to better understand the transcription-factor based mechanisms that may underlie microsatellite-mediated differences in Avpr1a expression.
Changes in transcriptional regulation are a primary driver of phenotypic evolution [58]. Here we demonstrate that the proximal 3.4 kb of the 5′ flanking region of the rodent Avpr1a gene has only a modest impact on species-specific expression patterns, indicating that elements outside of this region are important for many expression differences. Further studies using targeting vectors incorporating elements downstream of that used here, including coding region, intron, and 3′ untranslated region would be useful to determine whether the species specific patterns seen in our previous transgenic mouse study were conferred by downstream elements. Studies examining the genetic regulation of oxytocin and AVP gene expression have revealed the important role of intronic or 3′ flanking regions for cell-type specific expression [59]–[63]. Alternatively, more distal 5′ flanking regions, or even chromosomal landscape may play an important role in determining species-specific expression patterns [64]. However, our data do confirm that both species differences and intra-species variation in microsatellite structure contribute to variation in gene expression. To our knowledge, this is the first demonstration that species differences and individual variation in microsatellite structure has a direct impact on the expression pattern of a behaviorally relevant gene.
Our findings support the hypothesis that the instability of genetic elements proximal to genes may act as “evolutionary tuning knobs” to enhance the evolvability of traits through alteration of gene expression [14], [65], [66]. Microsatellite sequences typically mutate at faster rates than non-repetitive DNA [67], and unlike other forms of mutations, such as SNPs and indels, expansion or contraction of a microsatellite sequence is reversible [68]. Further, addition or subtraction of repeat units can exert small, quantitative effects on gene expression levels, such as those seen in the DG of pvKI-long and pvKI-short mice, leading to high gene expression divergence in a population of individuals carrying different microsatellite alleles. Repeat variation can alter gene expression via multiple mechanisms, including differential recruitment of transcriptional enhancers, altered secondary structure (e.g. bendability) of the DNA strand, and differences in epigenetic modifications that affect nucleosome binding [15]. Our results indicate that the Avpr1a microsatellite affects gene expression in multiple, but not all, brain regions. Because different mechanisms may underlie microsatellite-mediated expression changes in different brain regions, the inherent flexibility of phenotype conferred by repetitive sequences may be enhanced in the complex cellular environment of the brain, as compared to a single cell organism or a more homogenous tissue. It should be noted, though, that the differences in each of our comparisons between the lines created less divergence in V1aR binding than we anticipated, suggesting that other linked polymorphisms found outside of the microsatellite, as well as outside of the 3.4 Kb 5′ flanking region are contributing to the more robust differences reported in the vole studies.
Avpr1a is a particularly interesting locus for understanding the genomic mechanisms of phenotypic diversity, as it has been implicated in modulating social behavior in multiple species, including humans. An initial study reported that monogamous prairie and pine voles had longer Avpr1a microsatellites than nonmonogamous meadow and montane voles, suggesting that the presence of the microsatellite may have contributed to the evolution of the monogamous mating strategy in voles [24]. However, a subsequent survey of the Avpr1a locus in several other vole species and, more recently, in Peromyscus species did not support the hypothesis that the presence or absence of the microsatellite element was associated with monogamy [64], [69]. Nevertheless, more subtle differences in microsatellite structure may result in inter- and intraspecies differences in receptor expression, which could contribute to species differences in the expression of behaviors associated with monogamy [17], [70].
There is conclusive evidence that variation in Avpr1a expression contributes to variation in social behavior [17], [32], [34]. The present findings cannot confirm that variation in the microsatellite structure contributes to variation in social behavior in mice. Indeed, it is unlikely that social behaviors are significantly affected in our knock-in mice since the greatest alteration in V1aR expression were found in regions that have not been implicated in AVP-dependent social behavior. However, our results do support the more general hypothesis that variation in the Avpr1a microsatellite structure directly contributes to variation in V1aR density in a brain region specific manner.
Similar VNTRs are found proximal to the primate AVPR1A gene, and differences in the presence and composition of these regions exists both within and between species [71]–[73]. In humans, at least 16 alleles exist for a complex microsatellite located upstream of AVPR1A, known as RS3 [72]. It is worth noting that the specific sequences and location of the human microsatellite are different from that found in voles, but this region represents an analogous genetic region with putatively enhanced mutation rates. Variation in the length of this region has been associated with differences in V1aR mRNA levels in post-mortem human hippocampus, similar to our findings in the prairie long and short KI lines [74]. In addition, RS3 allelic variation predicts amygdala reactivity in response to face presentation, a highly salient social stimulus for humans [75]. Genetic studies have suggested a role for variation in RS3 and other AVPR1A microsatellites in multiple aspects of human social behavior, including male pair bonding and relationship quality, and altruism [39], [40], [74], [76]–[78]. In addition, nominal associations between RS3 variants and autism, a disorder characterized by deficits in social behavior, have been reported [79]–[81].
Chimpanzees are polymorphic for an indel that includes RS3 [71], and the presence or absence of this VNTR-containing region is associated with differences in a variety of personality traits. In particular, males carrying the RS3-containing allele demonstrated higher levels of dominance traits and lower levels of conscientiousness than males that lacked RS3 [41]. Together, these studies suggest that microsatellite diversity affecting Avpr1a expression may be a general mechanism for generating behavioral diversity in primates as well as rodents.
Our results suggest that variation in the microsatellite structure of Avpr1a can impact expression in the brain, but only to a modest extent, at least in mice. While we did not see an effect of the microsatellite on expression in regions associated with social behavior in our mice, it is conceivable that in the context the vole or human genome, similar microsatellite variation could have a larger impact on expression in regions involved in modulating social behavior, and thus could generate variation in the expression of behavior. Our results do suggest that the regulatory elements contributing to species-specific expression patterns are not confined to the proximal 5′ flanking sequence, and the regulation of species-specific expression patterns for this gene is more complex than we originally hypothesized. Future studies replacing larger stretches of the 5′ flanking region, exons and introns, or utilizing BAC transgenics may be able to further elucidate how species-specific patterns of gene expression in the brain are achieved.
All animal protocols were approved by the Columbia University Internal Animal Care and Use Committee and were conducted in accordance with the National Institutes of Health Guide for Care and Use of Laboratory Animals.
KI mice were generated using a targeting construct illustrated in Figure 2. The homology arms were amplified from a bacteria artificial chromosome (BAC) containing the C57Bl6/J Avpr1a locus using an enzyme mixture that includes both taq polymerase and a proof-reading polymerase (Epicentre Biotechnologies, Madison, WI). The homology arms were sequenced and the same homology arms were used in all three targeting constructs. Three versions of the prairie vole Avpr1a 5′ flanking region containing the meadow and prairie long and short microsatellite versions were isolated from previous expression constructs [82]. Specifically, because the three versions of the microsatellite were independently cloned into the same vector containing the prairie vole 5′ flanking region, this region for each construct was identical except for the structure of the microsatellite. This was confirmed by direct sequencing. A floxed PGK-Neo cassette was inserted upstream of the prairie 5′ flanking region and an HSV-tk cassette was placed downstream of the 3′ homology arm. The construct was linearized via digestion with Sbf1.
The linearized construct was sent to Ingeneious Targeting (Stonybrook, NY) where it was electroporated into hybrid C57Bl/6J/129SV embryonic stem cells. DNA from neomycin resistant/gancyclovir-sensitive clones were screened via southern blot. Specifically, genomic DNA was digested with Acc651 and evidence of recombination was detected using a probe located upstream of the 5′ homology arm (Figure 2). This yielded a 9.5 kb band in WT and a 5.1 kb band in correctly targeted recombinant alleles. Positive recombinants were further verified using two internal southern probes, PCR, and sequencing.
Correctly targeted recombinant stem cells were injected into blastocysts by Ingenious Targeting. Offspring of the chimeras carrying the targeted allele were crossed with mice expressing EIIa-Cre recombinase on a C57Bl6/J background. Because Cre-mediated recombination in this line is not 100% efficient, offspring were screened for deletion of the PGK-Neo cassette via PCR. All three lines were then bred to C57Bl/6J background for at least 5 generations. Animals were genotyped using the following primers: 5′ TACAAGTGAGTGGGCCTTTCCTGT and 5′ GAGCCTCGCGGGAAACTCAT for the WT allele (754 bp) and 5′ AGCTCTCTTCCATGCATTCGACCA and 5′ ACAGAAGCAACAGTGACCTTCCCT for the KI allele (334 bp) (Figure 2).
Mice were housed in groups of 3–5 animals with mixed genotypes, had ad libitum access to food and water, and were maintained on a 12∶12 light∶dark cycle. Mouse lines were maintained separately and WT and KI experimental animals were derived from heterozygous breeding pairs in each line (pvKI-long+/−, pvKI-short+/−, mvKI+/−).
N5 and N6 generation mice were euthanized between PND 60–70 via cervical dislocation followed by decapitation. Receptor autoradiography was performed as previously described [83]. Slide mounted sections at 100 mM intervals were thawed at room temperature for 1 hour, briefly fixed on 0.1% paraformaldehyde for 2 minutes, rinsed twice with 50 mM Tris buffer (pH 7.4), and incubated with 50 pM 125I-linear-AVP ligand [Phenylacetyl-DTyr(Me)-Phe-Gln-Asn-Arg-Pro-Arg-Tyr-NH2; Perkin Elmer, Waltham, MA] in buffer containing 50 mM Tris (pH = 7.4), 10 mM MgCl, and 0.1% BSA for 1 hour. The slides were then washed 4×5 min in 50 mM Tris buffer with 0.2% MgCl at 4°C followed by a final 30 minute rinse in the same buffer at room temperature with agitation. Slides were rinsed briefly in double distilled water and allowed to dry overnight before exposure to BioMax MR film along with an ARC146-F 14C standard. Multiple exposures, ranging from 18 to 72 hours, were performed to ensure all regions of interest could be evaluated within the linear range of the film. All slides were processed simultaneously.
Receptor densities were quantified by densitometry using MCID software as previously described [33]. Quantification was performed blind to genotype. Diagrammatic representative brain sections from Paxinos and Franklin (2008) were used to define anatomical regions. Briefly, for each region quantified, 3 serial sections were sampled bilaterally. Non-specific binding was calculated by selecting a background region not expressing V1aR for each section to account for potential section to section variation. Optical density was converted to pCi/region using the standard curve calculated from the co-exposed standard. Non-specific binding as subtracted from total binding to yield values for specific binding. Specific binding values were normalized to fold change relative to WT levels. Four WT animals for each line (n = 12 total) were pooled to generate a single WT group, derived from 9 independent litters from 7 breeder pairs. Eight knockin mice from each KI line were used, originating as follows: mvKI mice – 5 litters from 3 breeder pairs, pvKI-short mice – 5 litters from 4 breeder pairs, and pvKI-long mice – 6 litters from 3 breeder pairs. In each case, the groups were half male and half female. One mvKI individual was dropped from analysis of the CeA, PVThal, and DG V1aR levels due to tissue damage.
All statistical calculations are presented as mean ± SEM, and were performed in SPSS version 19. We tested for line differences by comparing WT littermates of all 3 lines (pvWT-long, pvWT-short, and mvWT; n = 4/line) using a 2-way ANOVA with line and brain region (CeA, PVthal, DG, LS, VP) as factors. We found a significant effect of brain region (F (48) = 122.2; p<0.001), but no significant effect of line (pv-long, pv-short, mv) (F (48) = 1.095; p = 0.35), and no evidence of interaction between the two (F (48) = 0.823, p = 0.56). Based on these results, WT littermates from all three lines were grouped together in subsequent analyses as the WT comparison group. To compare V1aR density in the brains of mice with different KI genotypes, we again used 2-way ANOVAs with genotype and brain region as factors. When significant main effects of genotype or interactions were observed, we conducted a simple effects analysis for genotype using a Sidak corrected α to account for multiple comparisons.
|
10.1371/journal.pcbi.1006595 | STRFs in primary auditory cortex emerge from masking-based statistics of natural sounds | We investigate how the neural processing in auditory cortex is shaped by the statistics of natural sounds. Hypothesising that auditory cortex (A1) represents the structural primitives out of which sounds are composed, we employ a statistical model to extract such components. The input to the model are cochleagrams which approximate the non-linear transformations a sound undergoes from the outer ear, through the cochlea to the auditory nerve. Cochleagram components do not superimpose linearly, but rather according to a rule which can be approximated using the max function. This is a consequence of the compression inherent in the cochleagram and the sparsity of natural sounds. Furthermore, cochleagrams do not have negative values. Cochleagrams are therefore not matched well by the assumptions of standard linear approaches such as sparse coding or ICA. We therefore consider a new encoding approach for natural sounds, which combines a model of early auditory processing with maximal causes analysis (MCA), a sparse coding model which captures both the non-linear combination rule and non-negativity of the data. An efficient truncated EM algorithm is used to fit the MCA model to cochleagram data. We characterize the generative fields (GFs) inferred by MCA with respect to in vivo neural responses in A1 by applying reverse correlation to estimate spectro-temporal receptive fields (STRFs) implied by the learned GFs. Despite the GFs being non-negative, the STRF estimates are found to contain both positive and negative subfields, where the negative subfields can be attributed to explaining away effects as captured by the applied inference method. A direct comparison with ferret A1 shows many similar forms, and the spectral and temporal modulation tuning of both ferret and model STRFs show similar ranges over the population. In summary, our model represents an alternative to linear approaches for biological auditory encoding while it captures salient data properties and links inhibitory subfields to explaining away effects.
| The information carried by natural sounds enters the cortex of mammals in a specific format: the cochleagram. Instead of representing the original pressure waveforms, the inner ear represents how the energy in a sound is distributed across frequency bands and how the energy distribution evolves over time. The generation of cochleagrams is highly non-linear resulting in the dominance of one sound source per time-frequency bin under natural conditions (masking). Auditory cortex is believed to decompose cochleagrams into structural primitives, i.e., reappearing regular spectro-temporal subpatterns that make up cochleagram patterns (similar to edges in images). However, such a decomposition has so far only been modeled without considering masking and non-negativity. Here we apply a novel non-linear sparse coding model that can capture masking non-linearities and non-negativities. When trained on cochleagrams of natural sounds, the model gives rise to an encoding primarily based-on spectro-temporally localized components. If stimulated by a sound, the encoding units compete to explain its contents. The competition is a direct consequence of the statistical sound model, and it results in neural responses being best described by spectro-temporal receptive fields (STRFs) with positive and negative subfields. The emerging STRFs show a higher similarity to experimentally measured STRFs than a model without masking, which provides evidence for cortical encoding being consistent with the masking based sound statistics of cochleagrams. Furthermore, and more generally, our study suggests for the first time that negative subfields of STRFs may be direct evidence for explaining away effects resulting from performing inference in an underlying statistical model.
| The goal of this paper is to understand the computational principles which underpin neural processing in auditory cortex. In particular, we investigate the hypothesis that neural processing is shaped by the statistics of natural sounds, the physical rules governing how those sounds combine, and the form of the initial processing performed by the ear.
It is well known that the outer, middle and inner ear transform an incoming sound pressure waveform into a representation at the auditory nerve which can be approximately described by a filtering stage (in which the sound is broken into subbands), followed by an envelope extraction and compression stage. This approximation to the auditory nerve’s representation of a sound is called a cochleagram and intuitively it can be thought of as revealing the spectro-temporal variations in the energy of the input waveform. It is believed that subsequent stages of auditory processing might decompose this representation into basic “structural primitives”, i.e., components or building blocks from which natural sounds are composed. Such a representation would provide a basis to support more complex computation at higher levels in the system (compare, e.g., [1]). The idea of representations in terms of primitives is supported to some extent by in vivo recordings in the primary auditory cortex of mammals which suggests that neurons are most sensitive to structures that are localized in time and frequency [2–6], but the hypothesis still lacks convincing evidence.
One way of investigating the hypothesis that auditory cortex is representing the components of natural sounds is to learn their form from a corpus of natural sounds. A particularly popular approach, which has been used for great success for visual data [7] and subsequently for audio data [8, 9], is based on the idea that the stimulus is formed by a linear combination of components which are sparsely activated. However, for auditory stimuli, this “sparse coding” approach is arguably not the most natural one to take for three main reasons. First, a linear mixture of sound pressure waveforms (formed either from multiple sources in the environment or from a single source comprising a linear mixture of primitive components) results in a non-linear mixture at the level of the auditory nerve and it seems likely that downstream processing would respect this fact. Second, the cochleagram is non-negative which is not reflected by the standard form of the sparse coding model. Third, sparse coding (or ICA) operates most effectively on whitened data (although this might be due to current algorithmic limitations, rather than a general feature of the approach).
In the visual system it has been argued that the lateral geniculate nucleus (LGN) performs such a whitening step [10] but the initial transformations employed in the auditory system are quite different, making this sort of preprocessing harder to justify. Whitening for cochleagrams would essentially mean that neural activities do not encode energies in frequency bands but deviations from a mean energy relative to energy variances. Adaptation effects to mean and variances over time are well known for regions upstream of the cortex such as the auditory nerve and inferior colliculus [11–14]. However, this adaptation should not be equated with whitening. If it was this would imply that the absence of any signal energy should lead to (on average) equally strong responses as energies above the mean. If we do not assume a whitening stage for cochleagrams or a similar preprocessing to obtain mean-free stimuli, then we are confronted with the question: How do measured STRFs with their positive and negative subfields emerge? In vision, after an assumed whitening stage, stimuli contain positive and negative parts which directly result in components extracted by sparse coding to have negative and positive subfields. For the non-negative energy representation of cochleagrams it is so far unclear how negative subfields can emerge without a whitening stage. Statistical data models not requiring whitening suggest alternative mechanisms commonly referred to as “explaining away effects” which have so far not been linked to negative subfields of neural response properties. As an example for “explaining away” consider the situation of sitting in a park. It is a nice warm day, you have your eyes closed, and are just listening to the sounds around you. There is a small orchestra somewhere with musicians practicing for a concert, and there are birds in the trees. If you now perceive a very short melodic sequence, it may have been generated by a bird or by a musician’s flute. As you are too far away from any of the sources, and as the perceived sequence is too short and unspecific, it is not possible for you to say for sure which of the potential sources may have generated the sound. But you do know that a high probability for one source, e.g. the flute, would mean a low probability for the other. This dependency between the probabilities for the two potential sources given a sound is called “explaining away”. If you were more certain that it was the flute playing (e.g., by getting additional visual input), the flute would “explain away” the alternative explanation of the sound having been generated by a bird. The statistical models investigated here will have similar explaining away effects but on a lower level of sound processing (Fig 6 will give a low level example later on). The primary statistical model investigated here assumes the data to be non-negative (and not whitened), and it assumes the structural primitives to combine non-linearly. More concretely, we assume structural primitives to combine such that the maximal energy in each time-frequency interval determines the superimposed signal (Fig 1 shows an illustration).
To summarize our goal, instead of using the dominating approach of standard sparse coding as statistical model to study neural representation in auditory cortex [8, 9], we investigate for the first time a non-linear and non-negative alternative. Our approach is motivated by the observation that alternatives to the assumptions of linear superposition and whitening may be more natural for acoustic data, and it offers an alternative explanation for the inhibitory subfields of STRFs which were previously closely linked to signal whitening.
We will now describe how we change the previously used assumptions of statistical models as discussed above. Engineers have known for a long time that representations such as the cochleagrams result from a non-linear interaction of primitive auditory components. Such non-linear interactions give rise to psychoacoustic masking effects, which have been successfully exploited in technical applications such as source separation (e.g., [15–17]). Underlying such masking effects are that natural sound energies tend to be sparsely distributed across frequencies and time, and that high energies dominate low energies in any spectro-temporal interval of a cochleagram. In practice this property is exploited by assigning each time-frequency interval to the one sound source or component that exhibits maximal energy [15–17], a procedure sometimes referred to as log-max approximation. This assumption is widely used in probabilistic models for auditory data processing [15, 16, 18] and finds application in denoising and source separation problems. Here we will also assume a combination rule of this form. Unfortunately, the audio-processing models mentioned above can only handle a small number of components (typically fewer than 10, compare [16]). In contrast, we expect the number of structural primitives required to explain natural sounds to be much larger (similar to a large number of edge-like components required to explain natural images). Therefore, we use instead the relatively novel model of Maximal Causes Analysis (MCA; [19]) that can be scaled to handle hundreds or up to a few thousands of components [20–22]. Not only does this model incorporate the non-linear max combination rule, it also comprises non-negative components much like a non-linear version of non-negative matrix factorization. Importantly, the method performs effectively without need for whitening and so it can be applied directly to non-negative cochleagrams as computed by auditory preprocessing models. The MCA approach, hence, matches those salient features of natural sound statistics previously not captured, making it to a more sensible alternative model for auditory processing in mammals.
Animal experiments were done at the Department of Physiology, Anatomy, and Genetics, University of Oxford, performed under license from the United Kingdom Home Office and were approved by the ethical review committee of the University of Oxford. The electrophysiological recordings were made from an adult pigmented ferret under ketamine (5 mg/kg/h) and medetomidine (0.022 mg/kg/h) anesthesia. After recording, the animal was killed with 1ml/kg i.v. Pentoject.
In the inner ear, sound pressure waves are considered to be broken-down into their frequency components by the cochlea, which then also compresses the frequency response amplitudes to form log-spectrograms resembling cochleagram representations of the input signal. The cochleagrams are then further communicated via the auditory nerve for neural processing and as they arrive in higher brain areas such as the primary auditory cortex, the cochleagrams are believed to get decomposed into elementary components for higher-level processing.
We assume that a cochleagram representation y → ∈ R D can be composed as a combination of a (small) number of primitive auditory components W → h ∈ R D, which form elements of a large dictionary W = ( W → 1 , … , W → H ) of H components. For such a multi-component encoding scheme, classical modeling approaches such as standard sparse coding [7] or ICA [27, 28] assume a linear interaction of the components to define a data generation process:
y → = ∑ h s h W → h + η → ,
where s h ∈ R determines the mixing factors for components W → h and η → denotes added noise in the generative process (which usually is assumed to be zero for ICA). However, cochleagrams are a representation of a non-linear interaction between the auditory components, for which a more accurate generative process can be derived from the log-max approximation [15–17]. The log-max approximation implies that the cochleagram of a linear mixture of sound waves can be well approximated by taking the pointwise maximum of cochleagrams computed from the individual waveforms. Fig 1 illustrates the approximation based on the cochleagram model used in this study. The example shows a better match by the point-wise maximum than by a linear combination. Hence, based on the approximation, we can define the following probabilistic generative model for cochleagrams:
p ( s →|Θ ) = ∏ h π s h ( 1 - π ) 1 - s h m (Bernoulli) (2) p ( y →|s → , Θ ) = N ( y → ; max h { s h W → h } , σ 2 I ) , (3)
where the max operation is applied element-wise, i.e., (maxh{ x→h })d=maxh{ xdh }, and where I denotes the identity matrix. Here we assume the factors sh ∈ {0, 1} to be Bernoulli distributed, whereas the observed noise is assumed to be Gaussian. Eqs 2 and 3 are a version of the MCA generative model [19, 20]. Parameters of the model are: the frequency π with which a component is activated, the variance of the observation noise σ2, and the generative components or fields W → h, which we will later relate to STRFs. For notational convenience Θ = (π, σ, W) denotes the set of all these parameters.
As a control for later numerical experiments with the MCA model, we will also consider a model assuming a standard linear combination of structural primitives. More concretely, we use a model that shares preprocessing, prior, and noise assumption with the MCA model but uses a linear superposition model instead of the point-wise max:
p ( y →|s → , Θ ) = N ( y → ; ∑ h s h W → h , σ 2 I ) . (4)
Eq 4 has the standard form of linear sparse coding approaches [7], and is because of the prior (2) a form of Binary Sparse Coding (BSC; [21, 29, 30]).
Given a set of N cochleagrams { y → ( n ) } n = 1 , … , N computed as in Section Cochlear model and spectrogram generation., we now seek parameters Θ* that optimally fit the MCA model to the data. We use likelihood maximization to find the optimal parameters and apply an approximate version of expectation maximization (EM; [31]) for their efficient estimation.
The application of standard maximum a-posteriori (MAP) based approximations is prohibitively suboptimal for the MCA model because the non-linear interaction of components typically results in multi-modal posteriors. An efficient approximate EM approach which can capture multi-modal posterior structure is, however, provided by Expectation Truncation (ET; [20]). ET can be regarded as a variational EM approach, and it has successfully been applied to MCA [21, 22, 32] and many other generative models [33, 34]. ET approximates the computationally intractable full posterior p ( s →|y → , Θ ) by a truncated one [20]:
q ( n ) ( s → ; Θ ) ∼ p ( s →|y → ( n ) , Θ ) δ ( s → ∈ K n ) , (5)
where δ is an indicator function (i.e., δ ( s → ∈ K n ) = 1 if s → ∈ K n and zero otherwise). If K n is chosen to be small but such that it contains the states with most posterior probability mass, the computation of the expectations in Eq 5 becomes tractable while a high accuracy of the approximations can be maintained [20]. The set K n is, therefore, chosen to consider the subset of the H′ most relevant hidden units for a patch y → ( n ). Furthermore, at most γ of these H′ units are assumed to be active simultaneously | s → | ≤ γ. Please see Efficient Likelihood Optimization in Supporting Information for a formal definition of K n.
Parameter update equations for the MCA model have been derived earlier [19, 21, 32]. They are given by:
W d h new = ∑n A d h ρ ( s → , W ) q ( n ) y d ( n ) ∑n A d h ρ ( s → , W ) q ( n ) , A d h ρ ( s → , W ) = ( ∂ ∂ W d h W ¯ d ρ ( s → , W ) ) , ( 6 ) W ¯ d ρ ( s → , W ) = ( ∑ h ( s h W d h ) ρ ) 1 ρ , ( 7 )
σ new = 1 N D ∑ n ⟨∥ y → ( n ) - max h { s h W → h } ∥ 2 ⟩ q ( n ) , i i i i π new = 1 H N ∑ n ⟨ | s → | ⟩ q ( n ) , (8)
where the parameter ρ in Eq 7 is set to a large value (we used ρ = 20) and ‖⋅‖ in Eq 8 denotes the L2-norm. The learning algorithm for the MCA generative model is thus given by the equations above with expectation values computed w.r.t. the approximate posterior in Eq 5. The linear BSC model, Eqs 2 and 4 is trained analogously to the MCA model with parameter update equations as derived earlier (e.g., [30]). Please see “Efficient Likelihood Optimization” in Supporting Information for more details.
We applied our method to male and female anechoic speeches in English, Japanese, Italian, and German. The data also included recordings of natural sounds such as rustling leaves, clattering stones and breaking twigs. More details about the data acquisition procedure are given in Natural Sound Recordings in Supporting Information.
We cut the waveforms of the recordings sampled at 44.1 kHz into snippets of 160 ms with a 32 ms overlap. The snippets were then transformed to cochleagram representations following section “Cochlear Model and Spectrogram Generation”. For the gammatone preprocessing we used a 32−channel filterbank with center frequencies ranging between 1000 and 22050 Hz. In this work we used Slaney’s implementation [35] to apply a 4th order gammatone filter. The outputs of the filter were averaged over a 20 ms sliding window with a 10 ms step size. The averaged energies were then compressed through the logarithm (as described earlier) to generate 32 × 15 cochleagrams, that is the energy at 32 center frequencies over 15 consecutive time windows.
We applied the MCA learning algorithm using H = 1000 generative fields to a set of N = 72800 cochleagrams. Individual cochleagrams were normalized by the L2-norm of their energies. To find the maximum likelihood parameters Θ approximately, we performed 70 EM iterations of the ET based learning algorithm described in “Efficient Likelihood Optimization”. The truncation parameters H′ and γ were set to 10 and 6, respectively. We initialized each of the components in the W matrix with the mean of the data perturbed by standard Gaussian noise with zero mean and variance set to 1/4th of the variance of the data.
Parameter σ was initialized to the square root of the variance of the data and π was set to 30/H where H = 1000. To minimize the possibility of running into local optima, we applied deterministic simulated annealing [36, 37] for the first half of the EM iterations with a linearly decreasing temperature from 10 to 1 (compare [20]). As a control, we also trained the linear BSC model analogously to MCA, i.e., using the same data preprocessing and initialization details as for MCA.
Fig 2C shows 100 of the 1000 learned generative fields after the 70 EM iterations. As can be observed, most of the fields are very localized in time and frequency. The generative fields resulting from applying the BSC model are provided in Supplementary S2 Fig.
In order to relate the MCA encoding of cochleagrams to neurons in the auditory cortex, we estimate spectro-temporal receptive fields (STRFs) from the inference results of the trained MCA model on the natural sound data. In physiological studies, an STRF is the numerically computed estimation of the linear mapping from sound cochleagrams that best predicts a neuron’s response. Similarly we compute STRFs that we consider to be tuned to individual latent components that we learn. To estimate STRFs W ^ * for the MCA model, we seek parameters that minimize the following function:
f ( W ^ ) = 1 N ∑ n = 1 N ∑ s → ( n ) ∈ K n p ( s → ( n ) | y → ( n ) , Θ ) ∥ W ^ y → ( n ) - s → ( n ) ∥ 2 + λ ∥ W ^ ∥ 2 , (9)
where y → ( n ) is the nth stimuli, W ^ is the row-dominated matrix of predicted STRFs, and λ is the coefficient for L2 regularization. Here we assume that the neural response to a stimulus will be a sample from p ( s → ( n ) | y → ( n ) , Θ ), in which case the experimentally measured STRFs will minimize the squared error between W ^ y → ( n ) and s → ( n ). Our assumption is consistent with interpreting neural responses as posterior samples [38], and the regularization term corresponds to assuming a zero-mean Gaussian hyperprior for the weights (compare ridge regression, e.g., as discussed in [39]). The intractable posterior over the latent factors p ( s → ( n ) | y → ( n ) , Θ ) in Eq 9 is truncated to only cover the subspace K n, as defined by the variational approximation technique in Efficient likelihood optimization. By setting the derivative of the cost function (9) to zero, W ^ can be estimated as:
W ^ = ( ∑ n = 1 N ⟨ s → ( n ) ⟩ q n ( y → ( n ) ) T ) ( λ N I + ∑ n = 1 N y → ( n ) ( y → ( n ) ) T ) - 1 (10)
where I is the D × D identity matrix and where 〈 · 〉 q n denotes the expectation value w.r.t. the approximation q n ( s → ; Θ ) of the posterior p ( s → ( n ) | y → ( n ) , Θ ) of the MCA model. The additional term λNI results from a L2-regularization for W in the cost function. Without regularization, the eigenvalues of the data covariance matrix ∑ n y → ( n ) ( y → ( n ) ) T were frequently very close to zero causing numerical instabilities. For the regularization parameter λ, we empirically found that a value in the mid-range of the minimum and the maximum eigenvalues of the data covariance matrix was sufficient to resolve the numerical instability.
Corresponding to the generative fields shown in Fig 2C, Fig 2D illustrates the STRF estimates computed from (10). We will refer to these estimates as model STRFs from now on. Observe first that many of the model STRFs are localized in time and frequency, a very common feature of receptive fields in the A1 [3, 40, 41]. Receptive fields produced by earlier sparse coding models do not as extensively have this punctate character [9, 42, 43]. Observe also that many of the model STRFs show flanking inhibition both spectrally and temporally, which is likewise a common feature of A1 receptive fields. However, a difference is that receptive fields of auditory cortical neurons tend to show asymmetry in their temporally flanking inhibition, most inhibition being found in the past relative to the excitatory region.
In Fig 3 (left) let us first consider 9 exemplary model STRFs, that illustrate various features which are also seen in experimentally recorded A1 STRFs as illustrated on the right-hand-side of Fig 3 (for how the STRFs were recorded from ferret cortex and estimated see the Supplement). Reading the Fig 3 (left) from left to right, the first unit shows punctate high frequency excitation, the second two units show punctate mid frequency excitation, and the next two units show punctate low frequency excitation. This illustrates that the units’ spectral tuning are spread over the frequency range, as found in physiology, as shown in Fig 3 (right). The sixth unit illustrates an upward sweep in frequency, and the seventh a downward sweep. The eighth and ninth units illustrate receptive fields that are spread out over frequency and time respectively. Again these four types of STRF are found in A1, as show in Fig 3 (right).
To quantitatively compare the model STRFs and auditory cortical STRFs across the population we took 244 experimentally recorded STRFs from Ferret A1 and AAF (taken from [45], see Supporting Information: “Neural Recordings and Real STRFs”) and compared them to the most frequently used model STRFs (i.e., to those fields which were the most probable to be activated across all stimuli). For the comparison, a 2D-Fourier transform was applied to each model receptive field and STRF, this provided the modulation transfer function of each receptive field and STRF (3 STRFs were excluded as all their values were zero, see Methods). Then, for each of the 241 remaining real STRFs and model STRFs the frequency modulation and temporal modulation at which the highest value occurred was taken (the best scale and best rate, respectively). A histogram of distribution of best scale and rate is plotted for the real A1 STRFs in Fig 4A and 4B (left), and for the MCA model STRFs in Fig 4A and 4B (middle). The histogram for the BSC model STRFs is shown in Fig 4A and 4B (right). For Fig 4 we used (to match the number of neurons we recorded from) the 241 most frequently used model fields, which represent ≈80% of the overall posterior mass for the MCA model. For comparison, the same histogram but using the 600 most frequently used model fields is shown in the Supplementary S3 Fig (capturing 97% of the posterior mass for the MCA model). S3A Fig (middle) is similar to Fig 4A (middle) but with more model fields at rate zero. The additional fields of Fig S3 which make up the difference to Fig 4 are, however, four times less likely to be active, which makes Fig 4A (middle) more representative for a comparison, see Supplement “Generative Fields and Estimated Model STRFs” for details. In contrast, the histograms for the 241 and the 600 most frequently used BSC fields show comparable percentages of STRFs close to rate zero.
Considering Fig 4, observe that the real STRFs and the receptive fields of the MCA model span a similar range of temporal modulations (rates) and a similar range of spectral modulations (scales). Fields tuned to higher scales and fields with higher and lower magnitudes of rate are a bit more frequent for the MCA model than for the experimental data. For the BSC model, the difference of the histogram to the measured data is larger. Significantly more fields have the best rates around zero or at higher magnitudes than the experimental data. The better match of histogram for the MCA model compared to the linear BSC model can be quantified using a χ2 test (Fig 4C). In conclusion, the receptive fields of the MCA model and real STRFs span a similar range of temporal modulations (rates) and a similar range of spectral modulations (scales). The model STRFs of the BSC model also span similar ranges of temporal and spectral modulation but this similarity is less pronounced than for the masking-based MCA model.
We also examined the tuning width, over frequency and over time, of the excitatory and inhibitory fields of the real and model STRFs. We used the same most frequently active model fields as for Fig 4, and a tuning width measurement method modified from [46]. For the measurement of frequency tuning width of the excitatory fields, the negative values of the STRFs were set to zero, then the STRF was squared in an element-wise manner and then the STRF was summed over the time bins to give a weighting vector over frequency bands. The excitatory frequency tuning width was then measured as span of frequencies (in octaves) whose weighting was ≥ 50% of the highest weighted frequency channel. For the measurement of temporal tuning width of the excitatory fields, the negative values of the STRFs were set to zero, then the STRF was squared in an element-wise manner, and then the STRF was summed over frequencies, to give a weighting vector over time bins. The excitatory temporal tuning width was measured as the number of time bins that were ≥ 50% of the maximum value of the resulting vector, multiplied by the time bin size of 10 ms. The inhibitory frequency and temporal tuning widths were measured similarly but instead the positive values of the STRF were set to zero, rather than the negative values. For a visualization of how they are measured see Supplementary S4 Fig.
Observe that for frequency, for the inhibition and to a lesser extent the excitation, the tuning widths of the MCA model STRFs match relatively well the tuning widths of the STRFs of real neurons. For the temporal dimension we see more strongly diverging properties which may have been expected by considering the statistical modeling approach: Like sparse coding or ICA we do focus on the composition of the data points in terms of structural primitives. Our model itself does not contain statistical dependencies in time unlike hidden Markov models or linear dynamical systems would do. As acoustic data does contain such dependencies on multiple time scales, it is likely that neural processing reflects also these dependencies. The discrepancy of temporal modulation in contrast to frequency modulation may therefore be taken as evidence for the auditory cortex capturing the intricate statistical dependencies over time which neither sparse coding, ICA nor the here studied MCA model addresses. The control experiments using BSC support this interpretation. Also for BSC no asymmetry similar to the one of the measured ferret STRFs is observed. Histograms for BSC computed analogously to Fig 5 are given in the Supplementary S5 Fig. In contrast to the histograms of best modulation frequencies, no notable differences between MCA and BSC histograms were observed.
We have investigated a computational model of auditory processing of sound waveforms in mammals that respects three key constraints. First, that a linear mixture of waveform components results in a non-linear mixing of cochleagram components, which is well approximated by the log-max non-linearity [15, 16]. Second, that the components in the model are positive and sparse. Third, that the statistical model operates on a stimulus closely aligned with biologically processing (cochleagram representation). As such the here followed maximal causes analysis (MCA) approach is arguably a more sensible approach than that provided by linear sparse coding methods that have previously been related to neural STRFs (e.g., [1, 9, 47]), and also of non-negative matrix factorization (NMF; [48, 49]). Perhaps surprisingly, whilst frequently used for sound processing tasks, to the best of our knowledge NMF has not been related to STRF recordings. In fact a relatively recent contribution explicitly states that NMF “does not allow for STRFs with inhibitory subfields” due to the positivity constraint [49].
We have shown that the MCA model exhibits a close correspondence to some of the STRF properties of neurons in ferret primary auditory cortex. Like STRFs of the real neurons, the MCA model STRFs show one or a few excitatory regions that are often punctate, being restricted over frequency and often over time. The excitatory regions of the MCA model STRFs are also often flanked by inhibition in frequency and/or time, consistent with real STRFs. The real neurons of our dataset and another ferret cortical dataset [50] show diverse STRFs, likewise the MCA model captures a similar diversity of STRFs with some model STRF broadly tuned over frequency or time, some narrowly tuned, some complex with multiple excitatory regions and some directional with diagonally oriented fields. However, the model STRFs do not capture the fact that inhibitory regions that flank in time tend to occur predominately after excitatory regions, rather than on both sides. This is unsurprising as the MCA model does not have the capacity to reflect causal statistical dependencies in time. MCA shares this property with other ICA-like and sparse coding models (including BSC). It may be noteworthy at this point that already in short-time STRFs, such as we use or are often measured in physiology, the limits of approaches that do not explicitly model dependencies in time are apparent. Measurements and analysis of neural responses in the auditory forebrain of birds [51] suggest that short-time STRFs do represent regularities important for capturing sound regularities over time. There, different types of STRFs have been linked to the processing of different sound properties such as spectral-pitch, rhythm, timbre or periodicity-pitch. Notably, specific functional roles of broad-band STRFs, and of STRFs with inhibition after excitation as well as STRFs with excitation after inhibition have been discussed in this context [51]. Also, the ‘noisy’ type STRFs of Carlin et al [50] with very disordered field structure are not notable in the models here considered.
The control model (BSC) produces STRFs with many properties similar to the MCA model, and most quantitative differences are relatively small. A main difference is that whereas the MCA model reproduces fairly well the distribution of best spectral and temporal modulation frequencies of real neurons, albeit somewhat overestimating the span of rates and scales, the BSC model shows significantly greater overestimation. On other measures they are similar. The MCA model captures fairly well the frequency tuning widths of real neurons, if underestimating to a degree, however in this capacity it did not perform noticeably better than the BSC model. Curiously, although in ferret data and our models the distribution of frequency tuning widths appears unimodal, in bird auditory forebrain [51] the distribution of frequency tuning widths is bimodal, we speculate as a consequence of the statistics of birdsong. Regarding temporal tuning, birds [51], our ferret data, and our models all show apparent unimodal distributions of temporal tuning widths. Both the MCA model and the BSC model substantially overestimate the temporal tuning widths of the STRFs of real neurons, which is again unsurprising as neither model has the capacity to reflect causal statistical dependencies in time.
Furthermore it should be noted that STRFs are far from a complete description of the tuning properties of auditory cortical neurons. Firstly, auditory cortical neurons show many non-linear properties [52] such as conjunctive AND-gate-like behavior [46], or amplitude modulation phase invariance [53]. Secondly, neural tuning properties, including STRFs, can also depend to an extent on stimuli used to gather them [45, 54–59]. Finally, STRFs can also show rapid plasticity depending on the task performed by an awake animal [5].
More generally, it is important to acknowledge that comparing normative models such as MCA to real data is difficult and depends on a number of factors including: details of the training corpus, details of different models of preprocessing and details of the STRF estimation. Any of these factors has an influence on quantitative comparisons as those made in this study. For instance, the data used to optimize a statistical model is unlikely to perfectly match the acoustic statistics experienced by the animals used to obtain the experimental data. Or different STRF estimation techniques applied to meet the requirements of experimental recordings or of the used models will effect the quantitative properties of estimated STRFs. Likewise, different preprocessing models (which we have not explored) influence STRF properties (see [60] for a discussion), and have also affected previous work on this topic [8, 9]. Any preprocessing scheme will, however, agree on cochleagrams being representations of acoustic waveform energies in time-frequency intervals. While such representations may be computed by very complex functions, any energy representations will assume non-negative values. Also strong masking non-linearities of the combination of structural primitives within cochleagram representations are widely agreed on in the literature. Notably, although the generative model here considered incorporates the positivity constraint (which we believe is biologically important), the recognition model nevertheless exhibits inhibitory subfields that arise due to explaining away effects among the components. This result indicates, perhaps counter intuitively, that models with positive generative components can still show inhibitory subfields if STRFs for these components’ generative fields are estimated—a finding which has implications beyond the specific model studied here and beyond the auditory system. More precisely, our study shows that inhibitory subfields can be a direct consequence of the statistical model assumed for explaining the data. Even if the data is non-negative and if the used model assumes non-negative generative fields and non-negative latent activities, inhibitory subfields can emerge directly from explaining away effects, without any additional assumptions. Similar to the introductory example, “explaining away” refers to a dependency between alternative explanations for a given stimulus. For our statistical models, possible explanations of a given stimulus take the form of combinations of generative fields (which are typically localized in time and/or frequency). The co-activation of two similar fields is unlikely (because of sparsity) which means that a high probability for one field results in a low probability for the other (and visa versa). Fig 6 aims at providing an intuition why inhibitory subfields emerge because of “explaining away”. Note that the fact that inhibitory subfields do emerge is independent, e.g., of the combination rule assumed by the statistical model, i.e., inhibitory subfields can be obtained for non-linear models of generative field combinations (MCA but also, e.g., noisy-OR models [61]) as well as for linear models. For the linear BSC model, we verified such an emergence of negative subfields also for non-negative weights by running additional experiments. While the BSC model we used for controls showed essentially positive weights, negative entries close to zero of the W matrix could be obtained and were obtained in our experiments. To ensure that negative subfields of STRFs also emerge for non-negative weights, we artificially enforced all W entries for BSC to be non-negative in our additional numerical experiments. Also in that case STRF estimation by Eq 10 resulted in negative subfields (see Supplement “Efficient Likelihood Optimization” for details).
If measured inhibitory subfields are a consequence of explaining away, then their shapes and the predicted dependencies among hidden neurons change depending on the assumed statistical model. By providing strong evidence for inhibitory subfields to be solely obtainable as a consequence of explaining away, our study offers novel ways of neuro-physiologically evaluating statistical models of neural processing.
Here we have compared spectral and temporal modulation as well as temporal and frequency tuning in order to compare different statistical models with data. Comparison of models is made difficult due to the above discussed factors. Significant differences of predicted STRFs can, nevertheless, be obtained if directly comparing statistical models with and without masking non-linearity (e.g., Fig 4) while all other model properties, training, and preprocessing remained fixed. A step further in the direction of neural evaluation would be represented by a direct in vivo comparison of neural responses to specifically designed stimuli. Given a set of neurons with previously measured STRFs, their responses could be predicted based on different statistical models. These different models will predict different response distributions, and artificial stimuli could be designed to be maximally discriminative between any two statistical models. Based on the results of this study, we predict responses for neurons in A1 which compete to explain an acoustic stimulus to not show a linear anti-correlation (as predicted by linear models). Explaining away resulting from a masking-based model (such as MCA), in contrast, would predict that neurons explaining the same stimulus compete rather in a k-winner-take-all manner, i.e., small sets of neurons suppress activity in the other neurons with only the maximally active neuron being relevant. For a comparison of explaining away effects between linear models and MCA see e.g. [62], for k-winner-take-all neural circuits see e.g. [63, 64]. In this context, let us, furthermore, remark that any neural activity distribution predicated by a model will not only depend on the model for generative field combinations but also on assumed priors, noise model and on the applied approximate inference approach. Furthermore, it will be important which variables of the model are assumed to match any measured neural activity best. Progress in neural recordings, simultaneous recording and stimulus generation, and refined neural modeling may make a direct comparison of statistical models feasible in the intermediate future.
A number of normative approaches have been taken to understand auditory spectro-temporal receptive fields as a consequence of stimulus statistics (e.g. [1, 9, 42, 43, 47, 49, 50, 65, 66]). Before discussing similarities and differences in relation to the models used here, let us stress that the capturing of stimulus statistics is not the only constraint of importance governing the structure of the nervous system. Biophysical constraints such as energy costs or wiring length are also important, as well as other functional constraints such as the role of particular sounds in an animal’s behavior.
Among the stimulus-statistics-based models, the most common approach has been the encoding of spectrogram-like representations of natural sounds subject to a sparsity constraint on the activity of the encoding units. Some sparse normative models balance a constraint for sparsity (or temporal slowness, [50]) while forcing dispersal [43] or decorrelation [42, 50] between the unit responses, and then learn the encoding receptive fields. More relevant for our study are those models which demand sparsity of unit responses while also generatively estimating the spectrograms from the unit activity via learned generative fields [1, 9, 47, 49, 65]. All the above sparsity and slowness models show some capacity to capture certain characteristics of STRFs. We have made explicit comparison of our model to a linear sparse model in the results (Fig 4), as it is the standard leading normative model of sensory coding, and we indicate the particular strengths of our model. The model of Carlin et al. [50] is less directly comparable to our model as it does not involve an explicit generative model. While it does in some ways better explain auditory cortical STRFs than a sparse coding model, it is clear that the MCA model captures certain aspects of the neural data that the slowness model of Carlin et al. does not address. Notably, the Carlin et al. model shows a near uniform distribution of best scales up to 2.5 cycles/octave, this is in contrast to our neural data (and that of Carlin et al.) and the MCA model where the density decays as scale increases (Fig 4).
In general, masking-based non-linearities, i.e., the dominance of one source in any time-frequency bin, is a property of acoustic data that has frequently been used for acoustic data processing (e.g. [16, 67]). In contrast, however, for the task of generatively explaining acoustic data by statistically learned structural primitives, almost all contributions in the literature rely on standard linear models. This applies for studies with functional focus (e.g., NMF-like [68, 69]) as well as for studies explaining neural response properties [1, 9, 49, 65]. The main reason for this strong focus on linear models is presumably related to the challenge of scaling strongly non-linear models to the large sizes required for sensory data. While linear models, e.g. for visual data, are routinely used with hundreds of generative fields / basis functions since about two decades [27, 70–72], non-linear models have been trained at large scales only relatively recently [21, 22, 62]. Earlier non-linear models, e.g., based on a noisy-OR non-linearity [61] or the maximum [19], have not been sufficiently efficient for learning with large numbers of generative fields.
While the approach used here does model masking, we do (as discussed above) not employ a statistical model that captures regularities in time. Other approaches do consider this important aspect of neural processing [66, 73, 74] e.g., to model longer term amplitude modulation structure of acoustic signals [73, 74]. Moreover, incorporating additional temporal statistical regularities is clearly important for acoustic synthesis [75] and might therefore be expected to have a strong effect on the neural representation of sound.
Among the approaches using assumptions formulated in terms of a statistical model, recent work by Yildiz et al. [47] is closely related to the linear models used in our study. That study, like our approach, seeks to explain acoustic stimuli by combinations of structural primitives. The focus by Yildiz et al. is a specific neural circuit implementation for probabilistic inference and learning. The derivation of the neural circuit relies on a mean field approximation for efficient inference, an adaptive Markovian dynamics, and a divisive inhibitory interaction among neurons representing structural primitives. The interaction of these mechanisms are shown to result in a stimulus representation with the underlying goal of providing a Bayes optimal explanation using combinations of learned generative fields. While this goal is shared with our approach, the assumed linear combination of primitives is the crucial difference of Yildiz et al. 2016 to our non-linear approach, i.e., they do not model masking. The generative data model underlying Yildiz et al. consequently more closely corresponds to the Binary Sparse Coding (BSC) model which we used as a control (Eqs 2 and 4). However, while Yildiz et al. infer STRFs from the circuit approximation of probabilistic inference, the results of S2 Fig of our study are based on directly inferring model STRFs from the linear BSC model itself. This makes the emergence of inhibitory subfields a direct consequence of the used generative data model, while Yildiz et al. first motivate a divisive form of inhibition to implement approximate probabilistic inference by their suggested circuit. On the other hand, both the here presented study and the study by Yildiz et al., 2016, provide evidence for auditory STRFs emerging from probabilistic inference and learning. Also both studies may be regarded as providing evidence for inhibitory subfields being a consequence of explaining away effects, as first hypothesizes by preliminary results obtained for our study [76]. In terms of concrete neural circuits that may realize such inference and learning, the study by Yildiz et al. 2016 goes very significantly beyond the research questions addressed here. On the other hand, in terms of showing that inhibitory subfields are a direct consequence of probabilistic inference, and in terms of using such fields to discriminate between different statistical models, our study significantly goes beyond the work by Yildiz et al. 2016.
Finally, note further technical but potentially import differences of approximate probabilistic inference applied to our and related approaches. The dominating approach for learning representations in terms of structural primitives are maximum a posteriori (MAP) approximations [7], i.e., the stimulus is represented by the latent state (i.e., by the neuron activities) with the highest posterior probability (highest p ( s →|y → , Θ ) in our case). MAP approximations are both scalable and relatively straight-forward to apply, which makes them being very frequently used also for statistical models of acoustic data (e.g., [1, 9]). However, with only maintaining the most probable hidden state for inference, no rich posterior structure is represented: neither correlations, multiple-modes nor any other type of the here very important explaining away effects is captured. In contrast, for our study and for other recent approaches (e.g., [47]) richer posterior representations play an important role. The observation that no previous study using MAP approximations has related inhibitory subfields of STRFs to explaining away effects, indicates that richer posterior representations seem to be required. However, while Yildiz et al. [47] as well as the BSC model used here maintain non-trivial posterior structures, the types of approximations used are different. Yildiz et al. 2016 employ a fully factored variational approximation (i.e., mean field). Such an approximation essentially assumes a posteriori independence of neural units, which has (given a stimulus) a direct impact on the activity dependencies among the stimulus encoding neurons. In contrast, the BSC model (as well as the MCA model) uses a truncated EM approximation which does not assume a posteriori independence [20]. The a posteriori independence of mean field has been criticized for introducing biases during learning [77, 78] while approaches that use truncated EM instead have been favorably compared with mean field [34].
To summarize, we have here shown that statistical models reflecting challenging data properties such as masking-based combinations of structural primitives and non-negativity are applicable to complex sensory data such as cochleagrams. Furthermore, we have found that inhibitory subfields of estimated model STRFs can directly emerge from explaining away effects of the assumed statistical model. This observation may lead to novel tools for the investigation of assumptions underlying probabilistic inference in the auditory cortex, in other sensory areas, and beyond.
|
10.1371/journal.ppat.1007062 | Metabolic reprogramming of Kaposi’s sarcoma associated herpes virus infected B-cells in hypoxia | Kaposi’s sarcoma associated herpesvirus (KSHV) infection stabilizes hypoxia inducible factors (HIFs). The interaction between KSHV encoded factors and HIFs plays a critical role in KSHV latency, reactivation and associated disease phenotypes. Besides modulation of large-scale signaling, KSHV infection also reprograms the metabolic activity of infected cells. However, the mechanism and cellular pathways modulated during these changes are poorly understood. We performed comparative RNA sequencing analysis on cells with stabilized hypoxia inducible factor 1 alpha (HIF1α) of KSHV negative or positive background to identify changes in global and metabolic gene expression. Our results show that hypoxia induces glucose dependency of KSHV positive cells with high glucose uptake and high lactate release. We identified the KSHV-encoded vGPCR, as a novel target of HIF1α and one of the main viral antigens of this metabolic reprogramming. Bioinformatics analysis of vGPCR promoter identified 9 distinct hypoxia responsive elements which were activated by HIF1α in-vitro. Expression of vGPCR alone was sufficient for induction of changes in the metabolic phenotype similar to those induced by KSHV under hypoxic conditions. Silencing of HIF1α rescued the hypoxia associated phenotype of KSHV positive cells. Analysis of the host transcriptome identified several common targets of hypoxia as well as KSHV encoded factors and other synergistically activated genes belonging to cellular pathways. These include those involved in carbohydrate, lipid and amino acids metabolism. Further DNA methyltranferases, DNMT3A and DNMT3B were found to be regulated by either KSHV, hypoxia, or both synergistically at the transcript and protein levels. This study showed distinct and common, as well as synergistic effects of HIF1α and KSHV-encoded proteins on metabolic reprogramming of KSHV-infected cells in the hypoxia.
| Hypoxia inducible factors (HIFs) play a critical role in survival and growth of cancerous cells, in addition to modulating cellular metabolism. Kaposi’s sarcoma associated herpesvirus (KSHV) infection stabilizes HIFs. Several factors encoded by KSHV are known to interact with up or downstream targets of HIFs. However, the process by which KSHV infection leads to stabilized HIF1α and modulation of the cellular metabolism is not understood. Comparative RNA sequencing analysis on cells with stabilized hypoxia inducible factor 1 alpha (HIF1α), of KSHV negative or positive cells led to identification of changes in global and metabolic gene expression. Our results show that hypoxia induces glucose dependency of KSHV positive cells with high glucose uptake and high lactate release. KSHV-encoded vGPCR was identified as a novel target of HIF1α regulation and a major viral antigen involved in metabolic reprogramming. Silencing of HIF1α rescued the hypoxia associated phenotype of KSHV positive cells. Analysis of the host transcriptome identified several common targets of hypoxia and KSHV-encoded factors, as well as other synergistically activated genes belonging to cellular metabolic pathways. This study showed unique, common and the synergistic effects of both HIF1α and KSHV-encoded proteins on metabolic reprogramming of KSHV-infected cells in hypoxia.
| Kaposi’s sarcoma associated herpesvirus (KSHV) is the etiological agent of Kaposi sarcoma, primary effusion lymphoma and multicentric Castleman disease [1–3]. By altering the expression of core metabolic enzymes, KSHV infected cells acquire a metabolic strategy of aerobic glycolysis generally referred as to the Warburg effect where these cells drive a high rate of glycolysis even in the presence of molecular oxygen [4–8]. This alteration of host metabolism mimicking Warburg effect by KSHV is believed to be necessary for the maintenance of latently infected cells [4]. Similar to most cancer cells, the mitochondria is also an organelle targeted by KSHV in viral infected cells altering apoptotic pathways and metabolism, so necessitating up-regulation of glycolysis to compensate for the energy demands of rapidly growing cells [9–12].
Metabolite profiling of KSHV infected cells suggest a wide difference between metabolite pools of KSHV infected cells when compared to control cells, including those which are common to anabolic pathways of most cancer cells [5]. Interestingly, the metabolite changes are not limited to only carbohydrates, but also included fatty acids and amino acids where inhibition of key enzymes in this pathway led to apoptosis of infected cells [5,13]. KSHV infection-mediated elevation of metabolites pools are due to enhanced anabolic activity rather than degradation from respective macromolecules [5]. Previous attempts to identify the mechanism of such reprogramming confirm increased expression of host factors such as glucose transporters, as well as hypoxia inducible factor (HIF1α) which are prerequisites for such changes in KSHV infected cells. In addition, decreased mitochondrial copy number and down regulated EGLN2 and HSPA9 have been reported upon over-expression of KSHV coded microRNAs, and are believed to be among the many KSHV factors involved in metabolic changes [14]. Nevertheless, previous observations either do not support or was unable to determine the involvement of other KSHV-encoded factors involved in metabolic differences caused by KSHV infection.
Hypoxia and HIF1α play critical roles in pathogenesis of KSHV by modulating expression of KSHV genes as well as stabilizing several KSHV-encoded proteins [15,16]. KSHV infection alone can mimic several physiological and metabolic changes due to hypoxia and those common to cancer cells. Hypoxia on the other hand plays an important role in KSHV reactivation biology where HIF1α facilitates KSHV-encoded RTA-mediated reactivation by binding with LANA and up-regulating RTA expression [16,17]. Hypoxia is also reported to enhance viral reactivation potential associated with 12-O-tetradecanoylphorbol-13-acetate [18]. The role of hypoxia in maintenance of latency and KSHV associated pathogenesis is also crucial, where the promoter of the key latent gene cluster coding for LANA, vFLIP and vCyclin harbors hypoxia responsive elements and can be activated by HIF1α [15]. Among other KSHV factors affecting the HIF1α axis, is the constitutively active G protein-coupled receptor (vGPCR) encoded by KSHV [19,20]. vGPCR is a bonafide oncogenic protein and stimulates angiogenesis by increasing the secretion of vascular endothelial growth factor (VEGF), which is a key angiogenic stimulator and a critical mitogen for the development of Kaposi’s sarcoma [21,22]. KSHV-encoded vGPCR enhances the expression of VEGF by stimulating the activity of the transcription factor HIF1α, which activates transcription from a HRE within the 5'-flanking region of the VEGF promoter [23]. Stimulation of HIF1α by KSHV encoded vGPCR involves phosphorylation of its regulatory/inhibitory domain by p38 and mitogen-activated protein kinase (MAPK) signaling pathways, thereby enhancing its transcriptional activity [24]. Specific inhibitors of the p38 / MAPK pathways are able to inhibit the transactivating activity of HIF1α induced by the KSHV-encoded vGPCR, as well as the VEGF expression and secretion from cells expressing this receptor [24]. These findings suggest that the KSHV-encoded vGPCR oncogene subverts convergent physiological pathways leading to angiogenesis and provides the first insight into a mechanism whereby growth factors and oncogenes acting upstream of MAPK, as well as inflammatory cytokines and cellular stresses that activate p38, can interact with the hypoxia-dependent machinery of angiogenesis [24]. However, the role of vGPCR in modulating other physiological pathways is poorly explored.
In the present study, we investigated the role of stabilized HIF1α on the metabolic status of KSHV positive cells and compared the results with KSHV negative cells with the same genetic background under normoxic or hypoxic conditions. We present data for differentially expressed KSHV-encoded genes when HIF1α is stabilized. We then showed the changes in global transcription of cells growing in normoxia or hypoxia with HIF1α to identify the common targets of HIF1α and KSHV infection. Our results showed enhanced induction of a tumorigenic metabolic phenotype in KSHV-positive cells growing in hypoxia compared to KSHV-negative cells growing under the same condition. Further, we now identify a comprehensive list of metabolic genes differentially expressed on KSHV infection in the hypoxic environment. These results provide new insights into the role of KSHV factors, in cooperation with hypoxia on the global metabolic status of KSHV positive cells.
KSHV infection is known to stabilize HIFs and this stabilization provides the cells a mechanism to survive in a hypoxic environment by up-regulating several cellular pathways involved in metabolism, survival and angiogenesis. We wanted to determine how KSHV infected cells respond compared to their isogenic KSHV-negative counterparts in hypoxic environments, and their metabolic requirements. There is no available control cell line with the same isogenic background for comparative studies in B-cells. Therefore, we selected KSHV-negative BJAB cells and KSHV positive BJAB-KSHV stably infected with KSHV [25]. We first characterized and confirmed the presence of full length KSHV in BJAB-KSHV cells at the level of the viral genome and transcriptome to determine if gross genomic alterations had occurred. Amplification of 10 different KSHV genomic regions with KSHV specific primers confirmed the presence of a KSHV genome most likely intact in BJAB-KSHV cells (S1A and S1B Fig). The sequences of the primers used to characterize BJAB-KSHV cells are also provided in S1 Table. The BJAB-KSHV cells were further characterized at the level of transcripts by amplifying the KSHV-encoded latent gene vCyclin from the cDNA made from BJAB-KSHV cells (S1C Fig). The isogenic background and authenticity of these two cell lines were also examined by short tandem repeat (STR) profiling. The STR profiling results for these two cell lines were compared with each other as well as with BJAB cells STR profile obtained from ExPASy Bioinformatics Resource portal database. The STR profile results confirmed the same origin and isogenic background of these cells (S2 Table). To study the effects of hypoxia, we proceeded with two different approaches to induce hypoxia in cell culture. In the first approach, we treated the cells with CoCl2, a chemical inducer of hypoxia to induce stabilization of HIF1α with minimal effects on the growth rate of the cells [26]. In the second approach we grew cells in 1% O2 hypoxic condition. Puromycin was omitted from the media of BJAB-KSHV and control cells for the entire treatment period. HIF1α stabilization was confirmed by western blot using HIF1α specific antibody (Fig 1A and 1B). An estimation of glucose consumed by BJAB or BJAB-KSHV cells suggested that the bulk of glucose from medium was being consumed during the initial period of 24–48 hours, in which cells growing under normoxic conditions showed an exponential growth pattern (Fig 1C & 1E). The growth patterns of cells growing in either CoCl2 or 1% O2 were quite different and showed diminished proliferation rates. Growing the cells in the same partially depleted medium showed a retarded growth in both normoxic as well as hypoxic environments, though hypoxic induction due to low oxygen showed a more drastic adverse effects on cell survival (Fig 1C & 1E). The estimation of glucose consumed by BJAB and BJAB-KSHV cells grown in normoxia and hypoxia showed a large difference in the consumption of glucose between these cells. Within the initial 24 hours, BJAB-KSHV cells showed an almost 18% higher glucose consumption as compared to BJAB cells during the same time period (940.7mg compared to 773.8mg glucose per million cells) (Fig 1C & 1E). The BJAB-KSHV cells showed a similar increase in uptake of glucose throughout the time points of 48, 72 and 96 hours compared to BJAB cells. A time dependent enhancement in the glucose uptake was observed for BJAB-KSHV cells when compared to BJAB cells growing in normoxic condition. However, the diminished medium condition and hypoxia due to 1% oxygen led to a drastic reduction in cell survival and growth past 72 hours. To rule out the possibility of an effect of puromycin pretreatment on glucose uptake, glucose consumption was also measured in cells growing either in the presence of puromycin, or its absence for 48 hours. The results showed no significant difference in glucose consumption due to presence or absence of puromycin in culture of BJAB-KSHV cells (S1D Fig). The effect of puromycin due to hypoxic induction or its downstream target was also determined by measuring real time expression of HIF1α and VEGFA. The results showed no effects of puromycin on expression of HIF1α nor VEGFA (S1E Fig).
To estimate lactate released in medium by these cells a standard curve of lactate ranging from 0 to 10 nmol/μl was prepared followed by a pilot experiment to determine the range of lactate in the medium. Here, different volumes of fresh culture medium (1μl and 10 μl) and 1μl medium from growing cultures were used (S1F Fig). Based on the pilot experiment, 10 μl of a 10X diluted culture medium was used to estimate lactate released in medium by BJAB and BJAB-KSHV cells growing under normoxia or CoCl2/1%O2-induced hypoxia (Fig 1D and 1F). A pattern similar to glucose uptake was observed for lactate release in these cells under similar growth conditions suggesting a directly proportional relationship between glucose uptake and lactate release. We also investigated whether this metabolic phenotype was mimicked in primary infection to peripheral blood mononuclear cells, KSHV infection of PBMCs was monitored growing them in the presence of CoCl2 or 1%O2. The infection of PBMCs with KSHV was confirmed by immuno-staining for KSHV latent protein LANA and the induction of hypoxia was confirmed by western blot to detect HIF1α (Fig 1G and 1H). The percentage of cells infected with KSHV was empirically calculated by LANA immune-staining. The infection efficiency of PBMCs with KSHV was approximately 50%. Estimation of glucose uptake and lactate release by infected PBMCs grown under conditions of normoxia or in CoCl2 or 1%O2 at 48 hours post-infection showed an enhanced glucose dependency and lactate release similar to BJAB and BJAB-KSHV cells (Fig 1I and 1J).
As the 1% oxygen for induction of hypoxia showed highly adverse effects on cell survival, we performed RNA sequencing experiments on BJAB and BJAB-KSHV cells growing in normoxia or CoCl2-induced hypoxia showing a more relevant physiological response of HIF1α stabilization due to KSHV infection. Analysis of RNA sequencing data for differential gene expression of KSHV encoded genes identified 42 transcripts coded by KSHV (Fig 2A–2D). A histogram for genes across the KSHV genome is provided in Fig 2A. Statistical analysis revealed that expressions of 11 KSHV-encoded genes were significantly changed when grown under CoCl2–induced hypoxia compared to their normoxic counterpart. Among these 11 genes, the viral G-protein coupled receptor (vGPCR), which is a constitutively active homolog of human G-protein coupled receptor [24], was found to be up-regulated by 3.62 fold (Fig 2B). Three other genes up-regulated due to hypoxia were K1 (Immunoreceptor tyrosine-based activation motif containing signal transducing membrane protein), ORF2 (homolog of cellular Dihydrofolate reductase), and ORF4 (Complement binding protein) with a fold change of 1.58, 1.26 and 1.38, respectively (Fig 2B–2D). Among the down-regulated genes, K12, ORF 40, and vFLIP were heavily down-regulated with a fold change of -4.1, -3.58 and -2.68, respectively (Fig 2B–2D). The levels of LANA and vCyclin transcripts were induced but not statistically significant due to possible differential efficiency of sequencing through these templates (Fig 2B). However, they were clearly induced as shown by RT-PCR of cells grown in CoCl2 and 1% O2 (Fig 2B & 2C). RTA transcripts were moderately increased as detected by sequencing, but was clearly increased when validated by RT-PCR in CoCl2 and 1% O2 (Fig 2B & 2C). Interestingly, all the four KSHV encoding interferon regulatory factors (vIRFs) were down-regulated with a fold change of -3.19, -2.19, -1.7 and -2.69 for vIRF1, vIRF2, vIRF3 and vIRF4, respectively (Fig 2B–2D). To validate the results obtained from differential gene expression seen for KSHV-encoded genes by RNA-sequencing, real-time PCR was also performed for the individual genes using gene specific primers. The primers used for real-time PCR are provided in S3 Table. Similar results were obtained by real-time PCR assays where vGPCR showed the highest up-regulation and K12 as a greatest down-regulated gene (Fig 2B & 2C). Similarly, RTA was shown to be up-regulated by RT-PCR in CoCl2 and 1% O2, as expected (Fig 2B & 2C). To further corroborate the differential gene expression of KSHV-encoded genes, we wanted to determine if a similar pattern was observed in low oxygen environment. BJAB-KSHV cells grown in a hypoxic chamber with 1% oxygen were collected and real-time PCR analysis was performed on the KSHV-encoded genes. The results showed a similar pattern of expression for the genes analyzed. However, the magnitude of change was slightly lower for vGPCR while it was slightly higher for K1 as compared to their expression in CoCl2-induced hypoxia (Fig 2C & 2D). The expression of ORF2, ORF4, vFLIP, vCyclin, LANA and RTA was also observed with the same pattern as it was seen in CoCl2-induced hypoxia (Fig 2C & 2D). Interestingly, the expression of some of the vIRFs were slightly less than that observed in CoCl2-induced hypoxia suggesting that the expression of vIRFs are also dependent on the overall ATP pool (Fig 2D). To determine the physiological relevance of the differentially expressed KSHV-encoded genes in response to hypoxia, real time expression of vGPCR, K1, vFLIP, vCyclin, LANA and RTA were also analyzed in the primary effusion lymphoma (PEL) cell line BC3, grown in both CoCl2 as well as 1% O2 induced hypoxia. The results strongly supported a universal effect of hypoxia on the expression of these KSHV-encoded genes (Fig 2E & 2F). These results led to further analysis of other critical KSHV-encoded genes when HIF1α was stabilized in the naturally infected KSHV positive cell line, BC3. We analyzed expression of 27 candidate KSHV-genes from BC3 cells grown under normoxic and CoCl2 induced hypoxic condition. The primer sets used are included in S3 Table. The resulting data showed that ORF9 (DNA polymerase), ORF18 (involved in late gene regulation), ORF25 and ORF26 (major capsid protein), ORF27 (Glycoprotein), ORF28 (BDLF3 EBV homolog), ORF34, ORF40 (Helicase-primase), ORF57 (mRNA export/splicing), and ORFK14.1 were significantly up-regulated in BC3 cells grown under hypoxic conditions (S2A–S2C Fig). Similarly, ORF11 (predicted dUTPase), ORF31 (nuclear ad cytoplasmic protein), ORF32, ORF33 (tegument proteins), ORF44 (Helicase), ORF64 (Deubiquitinase) and ORFK14 (vOX2) were significantly down-regulated in BC3 cells grown under hypoxic condition (S2A–S2C Fig). The expression of ORF6 (ssDNA binding protein), ORF7 (virion protein), ORF8 (Glycoprotein B), ORF36 (serine protein kinase), ORF54 (dUTPase/Immunmodulator), ORF56 (involve in DNA replication), ORF69 (BRLF2 nuclear egress) and ORFK8.1 (Glycoprotein) showed little or no significant change (S2A–S2C Fig).
Based on the results showing HIF1α stabilization and up-regulated expression of vGPCR, we performed a bioinformatics analysis of the vGPCR promoter region for identification of possible hypoxia responsive elements (HREs) [16]. A search for HREs consensus (ASGT; where S = C/G) within the vGPCR promoter identified 9 different HREs (Fig 3A). To determine the role of these HREs in directly regulating transcription of vGPCR in a HIF1α dependent manner, luciferase based reporter assays were performed. In brief, 10 different clones from the promoter region of vGPCR were generated and the results from the luciferase activity showed that the HREs at the 3rd, 4th, 5th, 6th, and 7th positions were significantly responsive to HIF1α (although not equally responsive). The promoter region containing all 9 HREs showed the strongest response, followed by clone C6 containing the initial 5 HREs (Fig 3C). The primers used to generate the clones are provided in S4 Table. Next, we wanted to determine if HIF1α knockdown in KSHV-positive cells can rescue the hypoxia associated expression of KSHV-encoded genes. The ShControl and ShHIF1α-BC3 cells were generated by lentivirus based transduction. Knock-down of HIF1α transcripts was confirmed at the transcript levels by real time PCR (Fig 3D). We confirmed the expression of HIF1α at the protein level by HIF1α western blot of lysates from ShControl and ShHIF1α BC3 cells grown under CoCl2 or 1% O2 induced hypoxia (Fig 3E). Real-time PCR analysis to determine the vGPCR and vFLIP expression in CoCl2 treated cells. HIF1α knockdown cells showed a reversal of expression as treatment with CoCl2 did not show the effect in HIF1α competent cells (Fig 3F and S2D Fig). As vGPCR is a potent candidate for the activation of several proliferation pathways, we wanted to determine whether the metabolic phenotype observed for KSHV positive cells was only due to elevated HIF1α levels, or if vGPCR expression was sufficient to induce the metabolic changes. We transfected HEK293T cells with an expression plasmid coding for KSHV-encoded vGPCR and compared it to that of cells transfected with empty vector. We also compared the results with cells expressing HIF1α. The results suggest that hypoxia or vGPCR can modulate the metabolic phenotype (Fig 3G and 3H).
RNA sequencing on total RNA from BJAB and BJAB-KSHV cells grown under normoxic condition or CoCl2 induced hypoxic condition was analyzed to determine the differential gene expression profiles. To increase our confidence in the differential gene expression data for host genes, the fold-change difference for VEGFA (a known target of HIF1α) was first analyzed (Fig 4A). Real time PCR validation was also performed for VEGFA transcripts from CoCl2 treated cells (Fig 4B). A comparative analysis was performed between BJAB vs BJAB-CoCl2, BJAB vs BJAB-KSHV and BJAB vs BJAB-KSHV-CoCl2 cells. The comparative analysis between BJAB cells grown under normoxia and CoCl2 induced hypoxia revealed major transcriptional changes between the two conditions (Fig 4C). CoCl2 treatment resulted in up-regulation of 2,182 transcripts (p≤0.001; FC ≤2 or ≥ 2). Similarly 1882 transcripts were observed down-regulated due to CoCl2 treatment (p≤0.001; FC ≤2 or ≥ 2). A volcano plot for the differentially expressed genes in BJAB-CoCl2 cells compared to BJAB cells is also presented showing that expression of a large number of genes was clearly modulated (Fig 4C).
The top 10 up-regulated and top 10 down-regulated genes from the comparative group are provided (Fig 4D). Analysis of the RNA sequencing data for the differential gene expression in BJAB-KSHV cells compared to BJAB cells resulted in detection of 357 up-regulated transcripts (p≤0.001; FC ≤2 or ≥ 2). Similarly, 233 transcripts were observed down-regulated in BJAB-KSHV cells compared to BJAB cells (p≤0.001; FC ≤2 or ≥ 2). A volcano plot for the differentially expressed genes in BJAB-KSHV cells compared to BJAB cells is shown (Fig 5A). The top 10 up-regulated and top 10 down-regulated genes from the comparative group are provided (Fig 5B). To further corroborate the transcriptional profiles of CoCl2 induced hypoxia or the combinatorial effects of hypoxia and KSHV infection, a comparative analysis of RNA sequencing data between BJAB and BJAB-KSHV-CoCl2 was performed. The results revealed a more enhanced effect of CoCl2 on transcription of host genes. Compared to 2182 transcripts up-regulated in BJAB-CoCl2 cells, 2560 transcripts were up-regulated in BJAB-KSHV-CoCl2 cells (p≤0.001; FC ≤2 or ≥ 2). Similarly, an enhanced effect on down-regulation of transcripts was also observed in BJAB-KSHV-CoCl2 cells where a total of 2,143 transcripts were observed down-regulated (p≤0.001; FC ≤2 or ≥ 2), compared to only 1,882 genes in BJAB-CoCl2 cells. A volcano plot for the differential gene expression of BJAB vs. BJAB-KSHV-CoCl2 cells is shown in S3A Fig. The top 10 up-regulated and top 10 down-regulated genes from this comparative group is also provided in S3B Fig.
KSHV infection is known to stabilize HIF1α [16,20]. We wanted to determine which genes are common targets of KSHV infection, and CoCl2-induced hypoxia. We also analysed the data to identify synergistic activation or suppression activities linked to the combination of KSHV and hypoxia. A Venn diagram was prepared (using Partek software) for the differentially expressed genes (p ≤0.001; FC ≤2 or ≥ 2) in BJAB-CoCl2, BJAB-KSHV and BJAB-KSHV-CoCl2 cells compared to BJAB cells (Fig 6A). Among the 357 transcripts observed up-regulated in BJAB-KSHV cells and 2,182 transcripts up-regulated in BJAB-CoCl2 cells, 160 transcripts were common (Fig 6A). Similarly, among 233 down- regulated transcripts in BJAB-KSHV cell and 1,882 down-regulated transcripts in BJAB- CoCl2 cells, 60 transcripts were common (Fig 6A). Interestingly, 105 transcripts out of 357 up-regulated in BJAB-KSHV cells were specific for KSHV. These transcripts were also up-regulated in BJAB-KSHV-CoCl2 cells (Fig 6A). Among the 233 down-regulated genes in BJAB-KSHV cells, 59 were observed to be specific for KSHV. These genes were also down-regulated in BJAB-KSHV-CoCl2 cells (Fig 6A). Intensity maps of common up-regulated and down-regulated genes are provided in S4 Fig. We wanted to know if transcription regulatory genes were common targets of CoCl2-induced hypoxia, and KSHV infection. Our analysis showed that the DNMTs, mainly DNMT3A and DNMT3B were two common targets for down-regulation induced by both hypoxia and KSHV infection. Real-time PCR analysis for these DNMTs followed by western blot analysis for protein levels showed similar results as observed from our RNA sequencing, although DNMT3A was clearly more dramatic in its suppression (Fig 6C, 6D and 6E).
310 genes involved in glucose, fatty acids and amino acids metabolism were identified by reviewing genes involved in these processes (S6 Table). This list was used to identify differentially expressed genes in each group when compared to BJAB cells. To optimize the number of genes differentially expressed in BJAB-KSHV, BJAB–CoCl2 or BJAB-KSHV-CoCl2, the analysis stringency was maintained to allow for statistical significance (p<0.05). A Venn diagram was created for the common genes from the 310 metabolic regulated genes which were differentially expressed (up-regulated or down-regulated) in each group above. Compared to BJAB cells, a total of 16 metabolic genes were up-regulated in BJAB-KSHV cells (Fig 7A). These up-regulated genes predominantly belonged to either glycolysis or the pentose phosphate pathway (ALDOA, ENO1, ENO2, HK2, PDK3, PDP2, PFKL and PGK1, PRPS1, PRPS2 and RPE), which supports a direct role for KSHV-infection in elevation of glycolysis. In addition, a subset of TCA cycle regulated genes ACLY, IDH3B, MDH1 and PCK1 were also up-regulated (Fig 7C). Interestingly, these genes were exclusively restricted to the KSHV-positive background, and were not up-regulated in cells grown under CoCl2-induced hypoxic condition (Fig 7A & 7C). Interestingly, in cells with elevated HIF1α due to CoCl2 treatment, activation of a different subset of metabolic genes from the glycolysis and TCA cycle pathways (15 genes for BJAB-CoCl2 and 16 genes for BJAB-KSHV-CoCl2) were identified. The up-regulated genes due to CoCl2 treatment were ALDOA, ALDOB, BPGM, PDPR and PGM3 (glycolysis) and DLST and IDH1 (TCA cycle) (Fig 7A & 7C). In addition, a set of glycogen synthesis genes (GSK3A, GSK3B, PHKG1, PHKG2 and PYGM) were also observed to be induced in CoCl2 treated cells (Fig 7A & 7C). Interestingly, HIF1α stabilization due to CoCl2 treatment appeared to be dominant over KSHV infection. The expressions of these up-regulated genes were similar in both BJAB and BJAB-KSHV cells treated with CoCl2 (Fig 7A & 7C). A second set of evidence showing glycolytic up-regulation by KSHV infection, or CoCl2-induced HIF1α was visible from the set of down-regulated genes in the TCA cycle (IDH2, PDHB, SDHA and SUCLG2), and glycogen metabolism (AGL, GB1, PCK2 and PGM1) (Fig 7B). These genes were down-regulated in both KSHV alone, and CoCl2 alone, in addition to their combination. CoCl2 treatment in fact showed an additional set of genes which were down-regulated compared to BJAB or BJAB-KSHV cells (Fig 7B & 7C). The differentially expressed genes observed by RNA sequencing were further validated by real time PCR (Fig 8A), using gene specific primers (S5 Table). Among the metabolic genes, Transketolase (TKT) and Succinate dehydrogenase subunit A (SDHA) were down-regulated by either KSHV-infection or CoCl2 treatment. Interestingly, the expression of both TKT and SDHA were suppressed by KSHV infection and HIF1α stabilization (Fig 8A and 8B). As vGPCR over-expression was associated with global transcriptional regulation and generation of reactive oxygen species (ROS) [27], we hypothesized that expression of these genes can be a consequence of induced vGPCR in the hypoxic environment. To confirm the role of vGPCR in the down-regulation of TKT and SDHA, vGPCR knock down BJAB-KSHV cells were generated by lentivirus based transduction (Fig 8C). Real-time expression of TKT and SDHA was analyzed in Sh-vGPCR BJAB-KSHV cells grown under normoxia or CoCl2 induced hypoxic conditions. The results showed clear involvement of vGPCR in regulating expression of these genes. ShControl cells showed a significant down-regulation of both TKT and SDHA expression in the hypoxic environment, however, Sh-vGPCR cells did not show any significant down-regulation (Fig 8D & 8E). Importantly, we did not observe any strong up-regulation of TKT and SDHA in hypoxia. However, in Sh-vGPCR cells their expression was definitely increased compared to wild type KSHV indicating a role in transcription regulation under hypoxic conditions.
To further corroborate a role for KSHV-encoded vGPCR in metabolic changes observed in the hypoxic environment, we investigated whether vGPCR knockout cells showed a reversal of the metabolic phenotype observed under hypoxia. The KSHV-bacmid clone containing vGPCR-Frame shift knock out mutant (KSHV-vGPCR-FS-KO) or vGPCR-Frame shift mutant reversed (KSHV-vGPCR-FS-R) KSHV [28] were transfected into HEK293T cells to generate stable lines. The transfected GFP positive cells were selected using hygromycin (Fig 9A). To confirm that the frame shift mutation and revertant was maintained, genomic DNA from stable cells were isolated and the KSHV region encompassing the insertion site, PCR amplified and sequenced. The electropherogram showing the sequencing results confirmed the frame shift insertion and reversion to wild type (Fig 9B). KSHV reactivation of the vGPCR-Frame shift mutant or revertant was induced by treatment with TPA and Butyric acid followed by standard virus purification. As expected, the vGPCR-Frame shift mutant showed a substantial decrease in lytic replication with significantly less yield in genome copies. To determine if vGPCR was a critical factor for the observed metabolic changes in hypoxia, PBMCs were infected with the vGPCR knockout and revertant KSHV virus and cells were subjected to hypoxic induction by treating with CoCl2. KSHV infection was monitored by GFP signal and the induction of hypoxia was confirmed by western blot to detect HIF1α (Fig 9D). The percentage of cells infected with KSHV was empirically calculated by counts of GFP signals in the cell population. The infection efficiency of PBMCs with KSHV was approximately 50%, although we observed slightly weaker GFP signal intensity from cells infected with the KSHV-vGPCR-FS-KO (Fig 9C). Cells were grown in normoxic, or CoCl2 induced hypoxia for 48 hours followed by media collection and measurement of glucose uptake. The results suggest a clear role for KSHV-encoded vGPCR in contributing to the metabolic changes. The PBMCs infected with the vGPCR kockout KSHV showed a significantly lower glucose uptake compared to the revertant (Wild type) KSHV infected cells (Fig 9E). As vGPCR has been shown to modulate transcriptional changes through the global signaling molecule i.e reactive oxygen species (ROS), we wanted to determine if vGPCR mediated ROS had any role in the transcriptional regulation of genes which were differentially expressed in our study. We first determined the levels of reactive oxygen species in PBMCs infected with revertant (wild type) KSHV treated with or without ROS scavenger, superoxide dismutase (SOD) by DCFH-DA staining (Fig 9E and 9F). As expected, the results showed a high level of reactive oxygen species in PBMCs infected by KSHV. The level was significantly lower in infected cells treated with SOD (Fig 9F). RNA was isolated from these cells and reversed transcribed. The expression of TKT was determined by real-time PCR of cDNA. The results showed a reversal in the expression pattern of TKT upon SOD treatment, suggesting a possible role of vGPCR mediated ROS in transcriptional regulation of cellular genes.
Similar to infection with most of oncogenic viruses, KSHV infection leads to stabilization of HIFs in host cells by either preventing its degradation or by up-regulating its expression at the transcription level [24,29–33]. The stabilized HIF1α alone or in conjugation with host and viral factors modulates several physiologic pathways supporting survival and growth of the infected cells [8,24]. Further, stabilization of hypoxia inducible factors due to viral infection only partially mimic the in vitro experimental methods of inducing hypoxia in cell culture by growing the cells either under low oxygen or chemical induction by Cobalt Chloride (CoCl2)/Deferoxamine mesylate (DFO) [34,35]. The stabilization of hypoxia inducible factor due to viral infection activated the HIF1α dependent pathways, whereas hypoxia due to low oxygen led to activation of several other energy associated pathways such as the AMPK dependent pathways [36–38].
Independent of the HIF1α stabilization mechanism, the interaction of stabilized hypoxia inducible factors with KSHV factors resulted in modulation of several pathways with impact on the host cell as well as the virus biology [16,18,39]. Hypoxia induces expression of the latency associated nuclear antigen (LANA), the key viral factor responsible for attachment of viral episomal DNA to the host chromosome [15,40]. On the other hand hypoxia is known to induce lytic replication of KSHV as well as enhancing the reactivation potential of known chemical inducers TPA and Butyric acid [16,18].
LANA can be described as a bonafide oncogenic protein with the ability to degrade cellular tumor suppressors as well as activate oncogenes [41–43]. Therefore, we would predict an enhanced tumorigenic state of KSHV infected cells in the hypoxic environment. However, KSHV-encoded reactivation and transcriptional activator (RTA) is also a downstream target of HIF1α and so allows for the possibility that KSHV-positive cells grown in a hypoxic environment will be more susceptible to lytic reactivation. Independent of cell destiny, both the latent or lytic pathways require enhanced metabolic activities for generating their required macromolecular components. However, the exact contributors to induction of the hypoxic phenotype to KSHV-positive cells have not been fully explored. What are the differences in utilization of physiological pathways in KSHV-negative and KSHV-positive cells grown in hypoxia has not been completely explored, particularly in B-cell lineages.
A huge challenge to investigating these questions is the limitation of available KSHV negative cell line controls. The comparative studies done previously for KSHV modulated pathways were performed by infecting cells of epithelial or endothelial origin while taking the parental cells as control[44]. However, for studies in B-cells, peripheral blood mononuclear cells were used [45]. Furthermore, the efficiency of KSHV infection remains a critical determinant in these studies [46].
In our current study, we used BJAB and BJAB-KSHV cells with the same genetic background [25], for comparative analysis of differential metabolic signatures and associated mechanisms due to the change in transcription profiles. Characterization of BJAB-KSHV cells showed that it can be used as a model B-cell line for our comparative analysis. We first investigated the metabolic behavior of BJAB and BJAB-KSHV cells grown in 1% oxygen or CoCl2-induced hypoxia. The results suggested an enhanced glucose dependency and a cancer cell metabolic phenotype of high lactate release in BJAB-KSHV cells growing in hypoxia due to both low oxygen, as well as CoCl2 treatment as compared to their counter KSHV negative BJAB cells. The observed changes were not exclusive to long-term infected cells. The initial infection of PBMCs with KSHV can also induce a similar pattern of changes when grown under hypoxic conditions. Interestingly, hypoxia induction due to low oxygen concentration induced a much higher glucose dependency and lactate release compared to CoCl2-induced hypoxia. Also, hypoxia due to low oxygen did not allow long-term survival of cells compared to hypoxia induced due to CoCl2. A higher rate of cell death after 48 hours was observed in culture. As hypoxia induction due to CoCl2 treatment shared more physiological relevance with stabilized HIF1α in KSHV positive cells at least in terms of cell survival, the RNA sequencing experiment performed on BJAB-KSHV cells growing in normoxia or CoCl2 induced hypoxia led to identification of novel HIF1α targets encoded by KSHV. One such target, vGPCR, is a constitutive homolog of cellular GPCR [24] and has been implicated in up-regulation of pathways common to most cancer cells including MAP-kinase, and the angiogenic pathways [24]. Although, vGPCR is considered a lytic gene for KSHV reactivation [47], its role in tumorigenesis is also reported due to its activation of on MEK/ERK [24,47]. In addition, vGPCR is also known to inhibit transcription of other KSHV-encoded lytic genes [48,49]. It can modulate global expression of host genes as its induced expression in cells is associated with global changes in signaling due to increased levels of ROS [27]. The slightly higher expression of KSHV-encoded ORF2 (homolog of cellular Dihydrofolate reductase) suggest a shift towards the anabolic pathway for nitrogenous base synthesis, which is a pre-requisite for cell transformation and viral reactivation. The induced expression of vGPCR and its correlation with enhanced glucose uptake, as well as its suppression in ShHIF1α cells provided hints towards its role in metabolic changes in KSHV infected cells in hypoxia. The results showing differential expression of KSHV-encoded transcripts also provided information about the half life and stability of these KSHV-encoded transcripts. For example, despite being transcribed from the same regulatory region, the transcript levels of LANA, vCyclin and vFLIP showed differential abundance, as well as a different pattern of expression under varying conditions. Though, it is now a well known fact that the transcript abundance not only depends on its transcriptional generation but also on the rate of its decay as a consequence of either its half life or regulation by non-coding RNAs, it would be interesting to explore the mechanism behind the differential abundance of vCyclin and vFLIP specifically under hypoxic conditions.
The analysis of global transcription changes in KSHV positive background and its similarity to those in KSHV negative cells grown in hypoxia suggested that HIF1α plays a role in KSHV infection in its associated pathology. Recently, a study designed to explore common transcriptional signatures for hypoxia and KSHV infection was performed using cells of different origins under low oxygen conditions. A convergence of targets due to KSHV infection and hypoxia was observed [50]. However, this study identified a limited number of genes differentially expressed under these conditions [50]. In combination with our data, we now provide a global picture of the role of hypoxia, KSHV infection, and their combinatorial effect on transcription of cellular genes regulated in the background of KSHV-infected cells. Further, it is possible that the role of HIF1α in B-cells, and other cell lineages including epithelial or endothelial cells may have differences in terms of transcripts or translated products. In B-cells, the oxygen partial pressure resembles that of blood supply. This is likely very different in solid tumors originated from epithelial or endothelial cells. In KSHV infected B-cells, induced HIF1α levels is mainly due to signaling modulated by KSHV-encoded antigens, whereas in KSHV-infected endothelial cells, the stabilization of HIF1α is likely due to the combined effects of signaling modulated by KSHV-encoded antigens in addition to low oxygen. In KSHV-infected endothelial cells, the hypoxic conditions may lead to activation of AMPK pathways due to low oxygen supply, which eventually affects ATP production through mitochondrial respiration and accumulation of AMP. The effect of induced HIF1α in B-cells may not activate the AMPK pathways. This study also provided information on the differences due to hypoxia induced due to low oxygen compared to CoCl2, especially in terms of cell survival and growth. An analysis of the effects on transcription of metabolic genes shows the up-regulation of the glycolytic pathway due to KSHV infection, and that induced hypoxia can affect the same pathways but though different targets. However, shut-down of specific cellular genes was similar in hypoxia and in KSHV infection.
While investigating possible cellular factor(s) involved in transcriptional regulation, analysis of global modulators such as DNMTs showed a significant down-regulation in their expression. This was mainly for DNMT3A and 3B due to CoCl2 induced hypoxia (for both BJAB and BJAB-KSHV cells) as seen in our RNA sequencing data sets. The expression of both DNMT3A and 3B was low in KSHV infected cells, however, the differences between them were not significant. Validation of the RNA sequencing results of DNMTs expression by real-time PCR confirmed suppression by both hypoxia and KSHV infection on DNMT3A, and 3B expression. Further greater suppression of at least DNMT3A was seen when hypoxia and KSHV infection were combined. The analysis of DNMTs expression is likely not the only explanation for the changes in expression of these large set of genes. However, it would be interesting to investigate the role of other global modulators including small non-coding RNAs on the expression of large set of genes in response to hypoxia or/and KSHV. These studies are ongoing with potential for elucidating additional mechanistic insights into the hypoxia-KSHV axis.
Peripheral blood mononuclear cells (PBMCs) from undefined and healthy donors were obtained from the Human Immunology Core (HIC), University of Pennsylvania. The Human Immunology Core maintains approved protocols of Institutional Review Board (IRB) in which a Declaration of Helsinki protocols were followed and each donor/patient gave written, informed consent.
BJAB (KSHV-negative B-cell) cells [51] were obtained from Elliot Kieff (Harvard Medical School,Boston, MA) originally purchased from American Type Culture Collection (ATCC). KSHV-positive BJAB-KSHV cells [25] were obtained from Michael Lagunoff (University of Washington, Seattle, WA). The KSHV-positive lymphoma-derived BC3 cell line was obtained from the ATCC. BJAB, BJAB-KSHV and BC3 cells were grown at 37°C/5% CO2 in RPMI medium containing 7% bovine growth serum (BGS) and penicillin (100 units/ml)/ streptomycin (0.1mg/ml). BJAB-KSHV cells were maintained with additional selection using puromycin (2μg/ml). Human Embryonic Kidney cell line (HEK 293T) was obtained from Jon Aster (Brigham and Womens Hospital, Boston, MA) and grown in DMEM medium containing 5% BGS with antibiotics at the above concentration. Cobalt chloride stock (100mM) was prepared in water. Hypoxia was induced by adding CoCl2 at a final concentration of 100 μM or by growing the cells in a hypoxia chamber with 1% oxygen. ShControl and ShHIF1α lentiviruses were generated by transfection of HEK293T cells with transfer plasmids and third generation packaging and envelop plasmids as descried earlier [52]. In brief, HEK293T cells were grown in 100 mm cell culture dish at 40–60% confluency. 10 μg of transfer plasmids in combination with packaging and envelop plasmids were transfected by calcium phosphate method. After initial discard of culture medium containing transfection mix, the supernatants were collected between 24–96 hours at 12 hours intervals. The supernatants were filtered through a 0.45 μm syringe filter and lentiviruses were concentrated by ultracentrifugation at 23,500 rpm for 2 hours [45,52]. The pelleted lentiviruses were resuspended in 1ml of complete medium and frozen until used for transduction. Transduction was performed by mixing cells with resuspended lentiviral stock in the presence of 8μg/ml polybrene. 48 hours post transduction, cells were selected in the presence of 2μg/ml puromycin. The pCEFL-vGPCR [53] construct was a kind gift from Enrique A. Mesri (University of Miami Miller School of Medicine, Miami, FL). The vGPCR-Knock Out (vGPCR-KO, Frame shift Mutant), vGPCR-Knock Out Reversed Bacmid clones, Sh-vGPCR lentiviral and control plasmids [28] were provided by John Nicholas (John Hopkins Bloomberg School of Public Health, Baltimore, MD). Large scale maxiprep for vGPCR-Knock Out (Frame shift Mutant) and vGPCR-Knock Out Reversed Bacmids were prepared using Luria broath culture and Qiagen large construct kit (Qiagen Inc., Hilden, Germany). Electroporation of HEK-293T cells were done in 4mm cuvette on Biorad Gene Pulser Xcell electroporation system. The positive clones of HEK293T-BAC-KSHV-vGPCR-KO and HEK293T-BAC-KSHV-vGPCR-KO reversed were selected in 100 μg/ml hygromycin.
KSHV virion stocks were prepared from the KSHV positive BC3 cells as previously described [45]. In brief, KSHV reactivation was mediated by adding TPA (to a final concentration of 20ng/ml) and Butyric acid (final concentration 3mM) and the cells were placed in an incubator at 37°C/5% CO2 for 5 days. The cells and supernatant were pelleted down by centrifugation for 30 minutes at 3,000 rpm. The supernatant was filtered through a 0.45 micron syringe filter. The cell pellets were resuspended in 10 ml of 1X PBS followed by 3X freeze/thaw cycle. The lysed cells were again collected by centrifugation for 30 minutes at 3000 rpm and the supernatants were filtered through 0.45 micron syringe filter and pooled. The filtrates were subjected to ultracentrifugation at 23,500 rpm for 2 hours to collect the KSHV virions. Infection of PBMCs was carried out in the presence of 8 μg/ml polybrene as described earlier [45].
RNA isolation was performed according to standard method of phenol chloroform extraction using TRizol reagent (Ambion, Grand Island, NY). 2μg of total RNA was used to synthesize cDNA by random priming method using Superscript cDNA synthesis kit (Applied Biosystems Inc., Foster City, CA). 1 μl of 10 times diluted cDNA was used for real-time PCR using Power SYBR green PCR reagent (Applied Biosystems Inc., Carlsbad, CA) using a Step One Plus or Quant Studio system (Applied Biosystems Inc., Carlsbad, CA). All real-time PCR assays were performed in duplicates, with one experimental repeats for each gene. Real-time PCR for a select set of genes were performed at least in triplicate. The differences in fold change were calculated by delta delta CT method using default parameter settings. DNA from cells were isolated using Blood & cell culture DNA isolation mini kit (Qiagen Inc., Hilden, Germany). Gel eluted PCR product was used for DNA sequencing (DNA sequencing facility, Department of Genetics, University of Pennsylvania) using both forward and reverse primers. Short tandem repeat (STR) profiling for BJAB and BJAB-KSHV cells was performed using GenePrint 10 kit (Promega Inc, Madison, WI) at the genomics core facility, Department of Genetics, University of Pennsylvania.
Glucose concentration available in cell culture medium was estimated using the hexokinase measurement kit (Sigma Inc., St. Louis, MO). The amount of glucose uptake by cells was measured by subtracting the amount of available glucose from the total amount of glucose available in fresh medium. The value of glucose uptake was finally normalized per million of cells. The amount of lactate present in the medium was estimated using the lactate estimation kit (BioVision Inc., Milpitas, CA). In brief, a standard curve for the known amount of lactate was created. A pilot experiment was performed using 1 μl and 10 μl of 1:10 times diluted cell culture medium from control or treated cells to estimate the range of lactate released.
Intracellular reactive oxygen species was determined by fluorescence of the cell permeable dye DCFH-DA. DCFH-DA stock solution at a concentration of 5 mM was prepared in DMSO. In brief, cells were stained with 5μM DCFH-DA for 30 minutes in complete media at 37°C in cell culture incubator followed by collection of cells at 1500 rpm for 5 minutes. The cells were counted and equal numbers were used to determine fluorescence on a micro plate reader (Molecular Devices, Sunnyvale, CA).
Cell lysates were separated on SDS-polyacrylamide gels followed by wet transfer to nitrocellulose membrane. 5% skimmed milk was used for blocking at room temperature for 1 hour with gentle shaking. Primary antibody against HIF1α (Santa Cruz Biotechnology, Dallas, TX and Novus Biosciences, Littleton, CO), GAPDH (Novus Biosciences, Littleton, CO), DNMT3A and DNMT3B (Abcam, Cambridge, MA) were incubated overnight at 4°C with gentle shaking followed by washing with TBST 3-times (5 minutes intervals). Probing with IR conjugated secondary antibodies was performed at room temperature for 1 hour followed by washing with TBST. Membranes were scanned using an Odyssey scanner (LI-COR Inc., Lincoln, NE) for detection of bands. For confocal microscopy, 25,000 cells were semi-dried on 8-well glass slides followed by fixation in 4% paraformaldehyde. Combined permealization and blocking was performed in 1XPBS containing 0.3% Triton X-100 and 5% goat serum followed by washing with 1X PBS. Anti-LANA antibody (purified ascites) was diluted in PBS containing 1% BSA and 0.3% triton X-100 and was incubated overnight at 4°C. Slides were washed with PBS followed by incubation with Alexa448 conjugated anti-mouse secondary antibody. DAPI staining was performed for 15 minutes at room temperature followed by washing and mounting. Images were captured by confocal microscope (Olympus, Lambertville, NJ).
Total RNA was isolated by standard phase extraction using TRizol reagent (Ambion Inc., Grand Island, NY). Isolated RNA was analyzed for quantity and quality using a Bio-photometer (Eppendorf Inc., Hamburg, Germany). Samples for RNA sequencing were prepared using Illumina RNA sequencing sample prep kit. The indexed ready to run samples were run on Illumina platform at the University of Washington (Core services). The reads for sequencing data were aligned with KSHV and the human genome. All the RNA sequencing experiments were performed in duplicates. The fold change expression and statistical relevance of the differential gene expression was calculated by CLC bio software (Qiagen Inc., Hilden, Germany). The differential gene expression was represented by volcano plot using R-software. Intensity plots for the fold-change expression of a selected set of genes were created using Partek software (Partek Inc., St. Louis, MO).
|
10.1371/journal.pgen.1000423 | Germline Mutation in NLRP2 (NALP2) in a Familial Imprinting Disorder (Beckwith-Wiedemann Syndrome) | Beckwith-Wiedemann syndrome (BWS) is a fetal overgrowth and human imprinting disorder resulting from the deregulation of a number of genes, including IGF2 and CDKN1C, in the imprinted gene cluster on chromosome 11p15.5. Most cases are sporadic and result from epimutations at either of the two 11p15.5 imprinting centres (IC1 and IC2). However, rare familial cases may be associated with germline 11p15.5 deletions causing abnormal imprinting in cis. We report a family with BWS and an IC2 epimutation in which affected siblings had inherited different parental 11p15.5 alleles excluding an in cis mechanism. Using a positional-candidate gene approach, we found that the mother was homozygous for a frameshift mutation in exon 6 of NLRP2. While germline mutations in NLRP7 have previously been associated with familial hydatidiform mole, this is the first description of NLRP2 mutation in human disease and the first report of a trans mechanism for disordered imprinting in BWS. These observations are consistent with the hypothesis that NLRP2 has a previously unrecognised role in establishing or maintaining genomic imprinting in humans.
| A small set of genes (imprinted genes) are expressed in a “parent-of-origin” manner, a phenomenon known as genomic imprinting. Research in human disorders associated with aberrant genomic imprinting provided insights into the molecular mechanisms of genomic imprinting and the role of imprinted genes in normal growth and development. Beckwith-Wiedemann syndrome (BWS) is a congenital overgrowth syndrome associated with developmental abnormalities and a predisposition to embryonic tumours. BWS results from alterations in expression or function of imprinted genes in the imprinted gene cluster at chromosome 11p15. Although BWS may be caused by a variety of molecular mechanisms, to date, all the genetic and epigenetic defects associated with BWS have been limited to 11p15.5. We report a family with two children affected with BWS and an epigenetic defect at 11p15.5 in which the primary genetic defect mapped outside the imprinted gene cluster. Using autozygosity mapping, we found an extended homozygous region on chromosome 19q13.4 (containing NLRP2 and NLRP7 genes) in the mother. Homozygous inactivating mutations in NLRP7 in women have been associated previously with abnormal imprinting and recurrent hydatidiform moles. We identified a homozygous frameshift mutation in NLRP2 in the mother of the two children with BWS implicating NLRP2 in the establishment and/or maintenance of genomic imprinting/methylation.
| Genomic imprinting is an epigenetic modification that causes genes to be expressed according to their parent of origin. Although less than 100 imprinted genes have been identified in human and mice, many imprinted genes appear to have a critical role in prenatal growth and development [1]. Molecular genetic analysis of rare human imprinting disorders has played a critical role in elucidating the mechanisms of genomic imprinting. In particular, studies of the imprinting disorder Beckwith-Wiedemann syndrome (BWS MIM 130650) have provided important insights into the structure and function of imprinting centres [2]. BWS is a congenital overgrowth syndrome, characterised by prenatal and postnatal overgrowth, macroglossia and anterior abdominal wall defects. Additionally, variable features include organomegaly, neonatal hypoglycaemia, hemihypertrophy, urogenital abnormalities and in about 5% of children, embryonal tumours (most frequently Wilms' tumour). The genetics of BWS are complex, but involve mutation or altered expression of several closely linked genes associated with cell cycle and growth control in the imprinted 11p15.5 chromosomal region. Imprinted genes most frequently implicated in the aetiology of BWS include the paternally expressed IGF2, KCNQ1OT1 (LIT1) genes and the maternally expressed H19 and CDKN1C (P57KIP2) genes. KCNQ1OT1 and H19 transcripts are not translated but the IGF2 gene product is an important prenatal growth factor and the CDKN1C protein is a candidate tumour suppressor that negatively regulates the cell cycle [3]. The majority of BWS cases are sporadic and result from epimutations of the distal (IC1) or proximal (IC2) 11p15.5 imprinting centres (see [4] and references within). IC1 is a differentially methylated region (DMR) about 5kb upstream of H19 that has an “insulator function” regulated by the zinc finger transcription factor, CTCF. The insulator is methylation sensitive, such that when CTCF binds to the unmethylated maternal allele, the IGF2 promoters do not have access to (are insulated from) enhancers downstream of H19. Methylation on the paternal allele prevents CTCF from binding, thus permitting interaction between the IGF2 promoters and the enhancers [5]. About 5–10% of sporadic BWS cases have hypermethylation of the H19 DMR and in these cases IGF2 shows loss of imprinting (LOI) and biallelic expression [6]. The second imprinting centre, IC2 is a DMR located in intron 10 of the KCNQ1 gene and is known as KvDMR1. The unmethylated paternal allele permits transcription of the antisense transcript KCNQ1OT1 (also known as LIT1) and silencing of genes including KCNQ1 and CDKN1C. Maternal methylation at the KvDMR1 is thought to prevent transcription of the KCNQ1OT1 gene and enable expression of CDKN1C. Loss of methylation (LOM) at the KvDMR1 is seen in up to 50% of sporadic BWS and is associated with biallelic expression (loss of imprinting) of KCNQ1OT1 and silencing of maternal CDKN1C expression [7]–[9]. Apparent hypomethylation of IC2 may, in rare cases, result from a germline IC2 deletion [10]. However, most BWS patients with IC2 methylation defects appear to have an epimutation of unknown cause (although there is an increased risk of BWS with IC2 epimutation in children conceived by assisted reproductive technologies) [11]–[13]. In order to gain insights into the factors responsible for IC2 imprinting defects, we studied a family with BWS that displayed evidence of IC2 epimutations through a trans mechanism.
This study was conducted according to the principles expressed in the Declaration of Helsinki. The study was approved by the South Birmingham Research Ethics Committee (equivalent to the Institutional Review Board of Birmingham Women's Hospital) reference number CA/5175. All patients provided written informed consent for the collection of samples and subsequent analysis.
A consanguineous family of Pakistani origin with two affected children with BWS due to loss of methylation at KvDMR1 were investigated in the first instance. Following the identification of NLRP2 mutations in this family, a further 11 BWS families, each with a single case of BWS (mean age 10.8 years) with KvDMR1 loss of methylation were analysed for NLRP2 mutations. This cohort included 10 patients who also had loss of methylation at other imprinted loci. Ethnically matched laboratory control samples were analysed to evaluate the significance of novel sequence variants.
Genomic DNA was extracted from peripheral lymphocytes by standard techniques.
In preliminary examinations chromosomal abnormalities were excluded and the methylation status of IC1 and IC2 of imprinting region 11p15.5 were determined.
Methylation analysis of KvDMR1 was performed as described previously, with PCR amplification of bisulphite modified DNA and digestion with restriction enzyme BstU1 yielding different sized fragments which is separated using ABI377 or 3730 [4]. In addition, methylation status at 3 additional DMRs was evaluated at the Transient Neonatal Diabetes Mellitus (TND) locus at 6q24, 7q32 (PEG1) and the Angelman/Prader-Willi locus at 15q13 (SNRPN) as described previously [14]. Primers and methods for the analysis of the methylation status of PEG1 DMR by methylation specific PCR (MS-PCR) were obtained from previously published report by Mackay et al [15].
For linkage studies a genome wide linkage scan was undertaken using the Affymetrix 250k SNP microarray. Mutation analysis of JMJD2D, ZFP57, NLRP7 and NLRP2 was carried out by direct sequencing. The genomic DNA sequence of these genes was taken from Ensembl (http://www.ensembl.org/index.html) and primer pairs for the translated exons were designed using primer3 software (http://fokker.wi.mit.edu/primer3/input.htm). Amplification was performed according to standard protocols with Bio Mix Red provided by Bioline. PCR products were directly sequenced by the Big Dye Terminator Cycle Sequencing System with the use of an ABI PRISM 3730 DNA Analyzer (Applied Biosystem). DNA sequences were analyzed using Chromas software.
A family with complex consanguinity (Figure 1) was ascertained after the diagnosis of two children with BWS. Both pregnancies were complicated by polyhydramnios and raised hHCG levels. Both affected children were born by Caesarean section. At birth Child 1 (V-1 in Figure 1) was macrosomic (4.55 kg at 38 weeks gestation) and was noted to have macroglossia, an omphalocele, ear creases, right inguinal hernia, undescended testis and neonatal hypoglycaemia in the first two days after birth. Similarly Child 2 (V-2) was macrosomic (3.633 kg at 35 weeks gestation) with macroglossia, ear lobe creases and neonatal hypoglycaemia that was difficult to control. Subsequently a third child was born without features of BWS (between the second and third children a probable hydatidiform mole was diagnosed). Molecular studies demonstrated loss of maternal allele KvDMR1 (IC2) methylation in both affected children, but the unaffected sibling had normal methylation. H19 methylation status was normal in both affected children and MLPA analysis demonstrated no evidence of an IC1 or IC2 deletion. Genotyping revealed no evidence of paternal uniparental disomy and linkage analysis with microsatellite markers flanking IC2 (TH and D11S4088) demonstrated that the two children had inherited opposite maternal and paternal 11p15.5 alleles. These findings were consistent with IC2 epimutation resulting from a trans imprinting defect.
In view of the history of consanguinity, an autosomal recessive disorder was suspected (either affecting both children or affecting the mother). Genetic linkage studies were undertaken by genotyping the two children and both parents on an Affymetrix 250k SNP array platform. Five regions of homozygosity (>2Mbases) were shared by the two children but these did not contain a gene known to be implicated in the establishment or maintenance of genomic imprinting. However inspection of the maternal genotypes revealed an ∼8 Mbase homozygous region containing NLRP2 and NLRP7 at 19q13.4. NLRP7 is a homologue of the mouse NLRP2 gene (NLRP7 is not present in the mouse) and the human NLRP2 gene. Sequencing of NLRP2 in the mother identified a homozygous frameshift mutation ((c.1479delAG, NM_017852; Figure 2) that was predicted (in the absence of nonsense-mediated RNA decay) to result in a truncated protein (p.Arg493SerfsX32) lacking 539 amino acids from the C-terminal that includes the LRR domain. The mutation was not detected in 542 ethnically matched control chromosomes but the father was heterozygous for the mutation, child 1 was homozygous for the mutation and the other two children were heterozygous. Mutation analysis of 11 additional families with BWS did not reveal any evidence of pathogenic NLRP2 mutations.
To determine if the trans imprinting defect extended beyond KvDMR1, we analysed methylation levels at the TND (6q24), SNRPN (15q13) and PEG1 (7q32) DMRs. Both affected siblings (and all controls) had normal methylation levels at the TND and SNRPN DMRs but Child 2 demonstrated partial loss of methylation at the PEG1 DMR (Figure 3).
We identified a homozygous frameshift mutation in the mother of two children with BWS caused by epimutations at IC2 (KvDMR1). Most cases of BWS due to loss of methylation of KvDMR1 are sporadic, but a handful of familial cases have been described with maternally inherited germline IC2 deletions [10]. However, in our family there was no evidence of a germline deletion by MLPA analysis [16] and the two affected children were shown to have inherited opposite maternal and paternal 11p15.5 alleles. This suggested that in this family, KvDMR1 LOM resulted from a trans and not a in cis effect. Germline NLRP2 mutations have not been reported previously, but mutations in ZFP57 and NLRP7 can cause imprinting disorders in which epimutations at imprinted loci result from a trans effect. Thus individuals homozygous for ZFP57 mutations presented with transient neonatal diabetes mellitus (TNDM). The major cause of TNDM is aberrant expression of imprinted genes at chromosome 6q24 and about 20% of cases have LOM at the TND differentially methylated region (DMR). Patients with homozygous ZFP57 mutations have LOM at the TND DMR, but also at other imprinted loci including KvDMR1 [15]. However, there was no evidence of germline ZFP57 mutations in our family. Germline NLRP7 mutations are associated with familial recurrent biparental complete hydatidiform mole (FHM) in which there is epigenetic abnormalities at DMRs in multiple imprinting regions [17]–[19]. Although FHM associated with NLRP7 mutations is inherited in an autosomal recessive manner, in contrast to ZFP57 mutation, homozygotes have normal genomic methylation but in female homozygotes there is a failure to establish methylation imprints in their germ cells leading to hydatidiform moles and reproductive wastage (male homozygotes do not have imprinting defects in their sperm). The methylation defects in FHM are specific for imprinted loci, and DNA methylation at non-imprinted genes and genes subject to X-inactivation is unaffected [18]. The human NLRP2 and NLRP7 genes are highly homologous and the two proteins consist of 1,062 and 1,009 amino acids respectively and have identical structure and about 64% amino acid identity. Only one of the two affected children with BWS was homozygous for a NLRP2 mutation and, by analogy with FHM caused by NLRP7 mutations, the BWS phenotype most likely results from homozygosity in the mother, such that familial BWS associated with NLRP2 mutations is inherited in a similar manner to NLRP7 associated FHM and not in a conventional autosomal recessive manner.
Both FHM and ZFP57–TNDM are associated with imprinting aberrations at multiple loci, and we identified partial loss of methylation at the PEG1 DMR in one of the affected children. Nevertheless it seems that NLRP2 mutations have a less severe effect on imprinting than NLRP7 or ZFP57 inactivation. A subset of children with BWS and an IC2 epimutation display hypomethylation at multiple imprinting centres (DMRs) (in our series these children are more likely to have been conceived by assisted reproductive technologies) [14],[20]. To our knowledge, none of these cases have been familial and we did not identify NLRP2 mutation in the sporadic cases we studied. It appears that the establishment (or maintenance) of methylation at KvDMR1 is particularly sensitive to genetic and/or environmental insults. We note that one of the affected children demonstrated a partial loss of methylation at PEG1 DMR1 both by bisulphite sequencing and MS-PCR suggesting that NLRP2 mutations may be associated with an incomplete failure of imprinting establishment and/or a partial failure of maintenance methylation at this DMR. Interestingly, investigation of a mouse knockout of ZFP57 has suggested a role in both the establishment of germline methylation imprints and in the postfertilisation maintenance of methylation imprints [21].
NLRP2 and NLRP7 encode members of the NLRP (Nucleotide-binding oligomerization domain, Leucine rich Repeat and Pyrin domain) family of CATERPILLER proteins. NLRP family of cytoplasmic proteins comprises 14 members of similar structure that are principally encoded by two gene clusters on chromosome 11p15 (NLRP6, 10 and 14) and 19q13.4 (NLRP2, 4, 5, 7, 8, 9, 11, 12 and 13). Most of the family members are well conserved from C. elegans, D. melanogaster, rat, and mouse to human but there is no rodent homologue for NLRP7 and the gene is found in only a few genomes (human, primate and cow). Some NLRP proteins are components of the inflammasome that is implicated in the sensing of, and inflammatory reaction to, extracellular pathogens and intracellular noxious compounds [22]. Germline mutations in NLRP3 and NLRP12 are associated with familial cold autoinflammatory syndrome [23],[24]. NLRP2 was suggested to function as a modulator of macrophage NFKB activation and procaspase 1 [25], however we found that the two family member homozygous for a NLRP2 truncating mutation did not show any evidence of an immune or autoinflammatory disorder. Nevertheless most NLRP family proteins are widely expressed and not restricted to the immune system. In addition, many are expressed in human oocytes and embryos at an early stage of development. Thus Zhang et al. have reported that NLRP4, 5, 8, 9, 11, 12, 13, and 14 were highly expressed in oocytes and then gradually decreased in embryos with a very low level in day 5 embryos, whilst NLRP2 and NLRP7 progressively decreased from oocytes to day 3 embryos then showed a sharp increase on day 5 [26]. These observations are consistent with NLRP2 and NLRP7 having a similar role in early development/imprinting establishment. Although it has been suggested that FHM might result from an immune-related defect in oogenesis or early embryo development (with methylation changes being a secondary phenomenon) the specific association of the methylation defects with imprinted DMRs suggests a more direct role in the establishment or maintenance of imprinting marks. Such a view is supported by the identification of germline NLRP2 mutations in BWS and should prompt further investigation of the role of NLRP2 and NLRP7 in genomic imprinting. The apparent involvement of NLRP proteins in genome methylation and the sensing and inflammatory response to extracellular pathogens and intracellular noxious compounds is intriguing given the suggestion that cytosine methylation may have evolved as a host response to transposons [27]. It is interesting that the third child in the family we report was unaffected. Most mothers with NLRP7 mutations have recurrent molar pregnancies only, but at least three families have been reported in which affected women had liveborn offspring (see [28] and references within). In one family, three affected members had, in addition to the molar phenotype, several miscarriages and three term pregnancies [29]. In one of the term pregnancies the baby was born with severe intrauterine growth retardation, but grew into a healthy adult with normal methylation levels [30]. In another term pregnancy the baby was born with unilateral cleft lip and palate and later manifested idiopathic delayed mental and motor development [31]. Murdoch et al. detected a homozygous splice site mutation in NLRP7 in all three affected women of this family [32]. Thus homozygous NLRP7 mutations may be associated with clinical heterogeneity/incomplete penetrance, possibly resulting from genetic modifier or environmental effects. In the light of these observations and the apparently milder phenotypic effects of maternal NLRP2 inactivation than NLRP7 inactivation (Beckwith-Wiedemann syndrome and molar pregnancy respectively) it might be predicted that clinical heterogeneity/incomplete penetrance would be a feature of maternal NLRP2 inactivation. Although maternal NLRP2 mutations appear to be a rare cause of familial BWS, the identification of these cases is important, as the inheritance pattern differs from the autosomal dominant inheritance (with parent of origin effects) associated with other inherited forms of BWS. The inheritance of NLRP2-associated BWS has similarities to other autosomal recessive disorders in which homozygous mothers are well, but there is a high risk to their offspring (e.g. FHM and treated maternal phenylketonuria).
|
10.1371/journal.pcbi.1000157 | Nonlinear Muscles, Passive Viscoelasticity and Body Taper Conspire To Create Neuromechanical Phase Lags in Anguilliform Swimmers | Locomotion provides superb examples of cooperation among neuromuscular systems, environmental reaction forces, and sensory feedback. As part of a program to understand the neuromechanics of locomotion, here we construct a model of anguilliform (eel-like) swimming in slender fishes. Building on a continuum mechanical representation of the body as an viscoelastic rod, actuated by a traveling wave of preferred curvature and subject to hydrodynamic reaction forces, we incorporate a new version of a calcium release and muscle force model, fitted to data from the lamprey Ichthyomyzon unicuspis, that interactively generates the curvature wave. We use the model to investigate the source of the difference in speeds observed between electromyographic waves of muscle activation and mechanical waves of body curvature, concluding that it is due to a combination of passive viscoelastic and geometric properties of the body and active muscle properties. Moreover, we find that nonlinear force dependence on muscle length and shortening velocity may reduce the work done by the swimming muscles in steady swimming.
| In this article we develop a computationally tractable model for swimming in animals such as eels, lampreys, and aquatic snakes. The model combines motoneuronal activation, muscle dynamics, passive elasticity and damping in the spinal cord and body tissues, and simplified hydrodynamic reaction forces, thus allowing us to probe how neuromechanical interactions give rise to body shapes and, ultimately, motion through the water. We use it to investigate the sources of an interesting experimental observation in freely swimming fish: that waves of curvature propagating along the body lag behind and travel more slowly than the muscular activation waves measured by electromyography. By selectively “lesioning” components of the model, we deduce that the speed difference, at least in this type of fish, is largely due to passive viscoelasticity and body geometry. We also find that nonlinear muscle properties are responsible for a significant reduction in energy expenditure and that they can also contribute to the wave speed difference. This work is a key step in a general program to build integrated “whole animal” models of locomotion and other behaviors that will also allow us to incorporate proprioceptive and exteroceptive neural feedback. Such integrated models can contribute both to our understanding of how living systems work and to the further development of robot systems.
| Most fish swim by rhythmically passing neural waves of muscle activation from head to tail, alternating left and right. This yields travelling waves of local muscle shortening, which in turn produce travelling waves of body curvature. These mechanical waves interact with the water, developing reactive thrust that pushes the animal forward. Breder [1] divided this type of swimming into two classes, depending on the proportion of the body undergoing undulations. In the anguilliform mode, as exhibited by, e.g. lampreys and eels, most or all of the body is flexible and participates in the propulsive movement. In carangiform swimming, as exhibited by, e.g. mackerel, the amplitude of lateral motion is concentrated near the tail. See [2] for an overview of animal locomotion, and [3]–[5] for vertebrate swimming in particular.
At any point on the body, rhythmic cycles of muscle activation alternate with silence, causing cycles of muscle shortening and lengthening (see Figure 1A). However, in all species which have been studied [8] except the leopard shark [9], delays between the onsets of activation and of shortening increase along the body from head to tail (see Figure 1C), i.e., the wave of shortening travels more slowly than the wave of activation. In consequence, near the tail the greater portion of the activation phase occurs during muscle lengthening, giving rise to negative work during part of the cycle. There are a number of possible functions assigned to this change in timing (e.g., providing stiffness as the tail moves laterally through the water, thereby contributing to power transmission, or tuning the resonant body frequency to match tailbeats [10]), but the mechanism or mechanisms responsible for it are not known [11]. In this paper, we throw light on this phenomenon.
Previous computational models of anguilliform swimming have incorporated the known timing of muscle activation within a mechanical representation of the body and water [12],[13], resulting in a travelling mechanical wave. In [13] no phase delay was seen between the waves of activation and curvature, and in [12], none was reported. However, both models assumed specific scalings of muscle density with body location, and that muscle force was simply proportional to activation. In reality, the force developed by activated muscle takes time to develop. Furthermore, because of the changing relative timing of activation and curvature, the patterns of muscle length and velocity vary significantly along the body length. This results in changing patterns in the developed muscle force, and such variation is further complicated by the body taper.
In the present study we investigate this phenomenon by incorporating a revised version of a kinetic muscle force model, originally due to Williams et al. [14], in the continuum mechanical model for anguilliform swimming of [13]. The resulting integrated neuromechanical system models the swimmer as an elastic rod with time-dependent preferred curvature arising from interactions of muscles with the body configuration. The model's modular structure—coupled sets of differential equations—allows us to selectively “lesion” it to probe the sources of its collective behavior. We find that the wave speed difference results primarily from the body's tapered geometry and passive viscoelastic damping, and that it does not require prioprioceptive sensory feedback. Depending on force density, the nonlinear dependence of force on muscle length and shortening velocity can also contribute to the wave speed difference, although it is not necessary for it. In a preliminary study, however, we find that length and velocity dependence can reduce the mechanical work output during swimming. When further coupled with a central pattern generator and motoneurons, this integrated muscle-body-enviroment model will also allow us to examine proprioceptive feedback, cf. [15].
This paper is organized as follows. In the methods section we review the equations of motion of the actuated rod and the fluid loading model. We show that the discretized rod equations are equivalent to equations describing a chain of interconnected links. This allows us to relate torques at the joints, and the forces responsible for them, to the preferred curvature and elastic properties of the rod. The model for muscle forces is developed in the penultimate subsection and in the final subsection we combine the muscle and body models to produce an integrated computational model. Simulations of the model are presented in Results and a discussion ensues in the concluding section, in which some larger implications of the work are noted.
We model the swimmer's body as an isotropic, inextensible, unshearable, viscoelastic rod that obeys a linear constitutive relation and is subject to hydrodynamic body forces. We assume that passive material properties such as density and bending stiffness remain constant in time, but allow them to vary along the rod. We endow the rod with a time-dependent preferred curvature in the form of a traveling wave, representing muscular activations. We adopt the conventions of [16],[17], and use an elliptical cross section to compute hydrodynamic reaction forces, although we restrict to planar motions, since lampreys and eels in “normal” steady swimming flex their bodies primarily in the horizontal plane [18],[19]. The calcium kinetics and muscle force model, which produces the preferred curvature, is described in the penultimate subsection and the integrated model is summarized in the final subsection of this section. The material of the first three subsections below is drawn from [13], to which the reader should refer for further detail, and where the numerical method and validation tests are also described.
The independent variable s∈[0,l] denotes arc-length along the rod, and a configuration of the rod is given at each time t by the space curve s ↦ r(s,t) = (x(s,t),y(s,t)) describing its centerline in the inertial (x,y)-plane. Derivatives with respect to s and t will be denoted by subscripts. The inextensibility condition |∂r/∂s| = 1, can be written in terms of the angle φ between the tangent to the curve t = ∂r/∂s and the inertial x-axis:(1)see Figure 2. The normal to r is then given by n = (−sin φ, cos φ). Each element of the rod is subject to contact forces f = (f,g), a contact moment M, and body forces W = (Wx,Wy) per unit length, vector components again being referred to the inertial frame. The contact forces and moment are those exerted on the region (s,s+ds) by [0,s), which maintain the inextensibility constraint, and the body forces arise from interactions with the fluid environment.
Balance of linear and angular momenta yields the equations of motion (cf. [17],[20]):(2)(3)(4)where ρ is the volumetric material density and A and I the cross-sectional area and moment of inertia of the rod. For an elliptical cross-section with semi-axes a and b, as in Figure 2, A = πab and the moment of inertia for motions in the (x,y)-plane is . We assume that ρ is constant, but allow A = A(s), I = I(s) to vary (both remaining strictly positive); specifically, we will study a tapered elliptical cross section based on lamprey body geometry.
In [13] the activation of the rod was determined by an externally-specified function κ(s,t), representing its intrinsic or preferred curvature. The muscle model developed later in this section effectively replaces κ with a function that depends on neural activation and the local curvature and its rate of change, but we retain the usual linear constitutive relation [20] so that moments are proportional to departures from preferred curvature:(5)Here E>0 and δ≥0 are the Young's modulus and viscoelastic damping coefficient and the flexural rigidity EI, with SI units N m2, determines the overall stiffness. The equations of motion (Equations 2–4), the constraints (Equation 1), and the constitutive relation (Equation 5), along with specified body forces and suitable boundary and initial conditions, form a closed system of evolution equations. Natural boundary conditions for free swimming are that contact forces and moments vanish at the head and tail: M = f = g = 0 at s = 0,l.
In swimming the local body forces are due to hydrodynamic reactions that depend on the global velocity field of the fluid relative to the body. To avoid the complexity and computational expense of solving coupled rod and Navier-Stokes equations, we adopt the model of G. I. Taylor [21] in which W(s,t) depends only on the local relative velocity. This approximation accurately predicts forces on a straight rod in steady flow, but fails to capture unsteady effects including vortex shedding, which are undoubtedly important in swimming propulsion [22],[23]. We believe that it suffices as a first approximation for the present purpose, since we are mainly concerned with the interaction of muscle forces and configuration dynamics. Unlike the Kirchhoff and Lighthill theories [24],[25], we neglect added mass effects. See [13] for further discussion.
Taylor models the force on a rod of radius a due to perpendicular flow of fluid of density ρf and dynamic viscosity μ with speed v as(6)where the drag coefficient CN varies between 0.9 and 1.1 for Reynolds numbers 20<R<105, and CT is closely approximated by in the range 10<R<105, cf. Figure 1 of [21]. Drag forces for smooth oblique cylinders can be decomposed into normal and tangential components in terms of the normal and tangential velocities v⊥ and v∥ at (s,t) as:(7) and the body forces are given by(8)where n and t denote the normal and tangential unit vectors to the rod's centerline at s.
In calculating W, we consider only the height 2a of the rod, assuming that fluid reaction forces are equal to those on a cylinder of radius a, although the constant CN does change slightly for elliptical rods. Further, we set CN = 1, since Reynolds numbers for lampreys and eels lie well within the range 20<Re<105; for example, in their work on the eel Anguilla rostrata, Tytell and Lauder cite Re = 60,000 based on body length l = 20 cm for a specimen swimming at 1.4l/s. [22], and speeds reported in [23] range from 0.5 to 2 body lengths per second. In terms of Taylor's body-diameter-based Reynolds number, this corresponds to R≈2000–8000.
We discretize the rod equations with spatial step size h = l/N in the arclength variable s, letting xi(t) = x(ih,t), i = 0, …, N, and similarly for the other field variables yi,φi and parameters Ai,Ii: see Figure 3. The inextensibility constraints in Equation 1 are approximated by(9) and Equations 2–4 are approximated by the ordinary differential equations (ODEs):(10)(11)(12)where mi = ρAih and Ji = ρIih. The constitutive relation in Equation 5 becomes:(13)The force and moment free boundary conditions M = f = g = 0 at s = 0,l become:(14)
The finite-difference discretization of Equations 10–13 is closely related to representions of the body as a planar chain of rigid links subject to forces and moments. In modeling lamprey Bowtell and Williams [26],[27] take a chain of N massless rigid rods each of length h, with mass mi at each pivot and at both free ends. The pivots are actuated by passive springs, dashpots, and active force generators. Ekeberg [12],[28] adopts a similar configuration but in place of time-dependent force generators, the spring constants vary with time, and instead of point masses at the pivots, the center of mass of each link is placed at its midpoint. Here we adopt the mass distribution of [12], and include active muscle elements, to be described in succeeding subsections, in the force-generating components. The configuration of the ith link is described by its midpoint (xi,yi) and the angle φi between its centerline and the inertial basis vector êx (Figure 3). Equaions 9 then express the constraint that links remain connected at the joints. Letting (fi,gi) and Mi denote the components of contact force and the torque at the joint connecting link i to link i+1 and (hWxi,hWyi) be the body force acting on the midpoint of link i (Figure 4a), balances of linear and angular momenta yield Equations 10–12 above with mass mi = ρAih and moment of inertia of the ith link. The discrepancy between the discretized rod equations and the equations for the chain of N pivoted rods thus consists only in the terms in the moments of inertia, and the two models coincide in the limit h → 0. We employ the exact formula above for the moments of inertia Ji in all the calculations below, although the approximation Ji = ρhIi yields results (not shown) that are nearly identical, even for quite large values of h≈1.
As shown in section 4.3.4 of [13], for the large segment numbers typical of eels and lampreys, the behaviors of the discrete and continuum models are very close. Additionally, the discretization reveals how activation determines preferred curvature κ(s,t) and affects bending stiffness EI of the continuum model. As in [26], the joint connecting each pair of links of length h is actuated by a pair of spring-dashpot-actuators in parallel, with spring constant ν and damping coefficient γ, anchored to arms of length w that project normally from the links' midpoints (Figure 4b). These arms represent myosepta, the connective tissue layers to which the muscle fibres connect. The linear springs and dashpots represent passive tissue viscoelasticity, and the actuators generate prescribed contractile muscle forces fLi and fRi on the right and left sides of the body respectively. Suppressing the dependence on i and denoting the relative extensions and of the spring-dashpot-actuators as ΔR and ΔL (Figure 4c), the total forces on the right and left sides may be written(15)Since the relative extensions are dimensionless, stiffness ν and damping γ have the units N and N s. respectively. The springs are in tension (and hence generating contractile forces) when ΔR, ΔL>0. The forces are applied at a distance w from the centerline of the rod, so elementary trignometry gives:(16)where ψi = φi+1−φi is the angle between neighboring links and . Finally, computing the moment arms LR,LL to the joint along normals from the lines AB and CD on which the forces act (Figure 4c):(17) we find that, for small angles ψi, the resulting torque at joint i is given by(18)
Comparing the linearized moment in Equation 18 in the limit h → 0 with the discretized constitutive relation in Equation 13 we see that the link and discretized rod models coincide if the stiffness EIi, intrinsic curvature κi and viscoelastic damping δ are interpreted as follows:(19)
We propose that the stiffness ν and damping γ are proportional to cross-sectional area A(s). Thus we set(20)so that the stiffness and damping have units N/m2 and N s./m2 respectively. To approximate a uniform distribution of the muscle, we set w = b/2, where b is the half-width of the body. Equations 19 now become(21)In particular, using I = πab3/4 we can write Young's modulus in terms of the spring stiffness as . One of the questions we address is the influence of force density as a function of arclength. We take up this question after a discussion of force generation in muscle fibers.
Recordings such as those of [29] show that waves of motoneuronal activity consisting of bursts of closely-spaced action potentials (APs), separated by near-silent interburst periods, travel the length of the lamprey spinal cord (see Figure 1A and 1B). The waves are generated spontaneously by a distributed central pattern generator (CPG) within the spinal cord [30], which has been modelled as a chain of coupled oscillators [31]–[33]. The waves are in antiphase contralaterally and maintain approximately constant duty cycles (burst/cycle period ratios) and segment-to-segment ipsilateral phase lags, regardless of overall frequency. This activity pattern is transmitted via nerves that enter the myotomes through the ventral roots [34], producing muscle activation with similar phasing, evident in electromyograms (EMGs) [7]. Each myotome corresponds to a segment of the spinal cord.
Bundles of myofibrils make up the muscle fibres within the myotomes. The AP bursts cause calcium release from the sarcoplasmatic reticulum (SR) that surrounds the myofibrils and is encircled by T-tubuli at repeated intervals. The resulting muscle contraction occurs in three phases. (i) A motoneuronal AP arrives at the neuromuscular junction, producing an AP at the motor end plate which spreads along the surface and T-tubular membranes of the muscle fiber. (ii) This depolarization opens gates in the SR and releases Ca2+ ions into the muscle protein filaments. (iii) Ca2+ causes conformational changes in the thick filaments which form cross-bridges to the thin filaments; a subsequent conformational change then develops a force tending to slide the thin filaments over the thick ones [35], shortening the muscle (unless overcome by opposing force via the muscle attachments). This is followed by resequestering of Ca2+ by the SR, resulting in relaxation of the muscle. The force developed during muscle activation is dependent upon both the length of the muscle and the velocity of its shortening [36]. Traditionally, shortening is taken as positive, but here we use the opposite convention, referring to the time derivative of muscle length as velocity, which is negative for shortening.
To describe the forces fR(t) and fL(t) in Equations 15, 18, and 19, we adapt the model developed by Williams et al., who carried out experiments on portions of single myotomes of lamprey muscle [14]. Intermittent tetanic stimulation was applied during isometric and constant-velocity movements, and analysis and modelling of the resulting force trajectories were used to predict the trajectories recorded during applied sinusoidal movement. Experimental data are reproduced in Figure 5 below (for details of experimental protocol, see [14]). We follow a modified form of the simple kinetic model used in that study, including calcium ions, SR sites and contractile filaments (CF). The rates at which calcium ions are bound and released approximately follows the principle of mass action (see Figure 6). For example, the rate of binding of calcium ions to the CF is proportional to the product of concentrations of free calcium ions and unbound filaments, with rate constant k3. The resulting equations for the kinetics of the calcium, sarcoplasmic reticulum sites and bound filaments are as follows:(22)(23)(24)(25)(26)where brackets denote concentrations of the relevant quantity. When the muscle is activated, k1>0 and k2 = 0; in the absence of activation k1 = 0 and k2>0.
We assume that the total number of calcium ions, SR binding sites and filament binding sites per liter remain constant so that [cs]+[c]+[cf] = CT, [cs]+[s] = ST, and [cf]+[f] = FT. This allows us to reduce the five Equations 22–26 to a system of two in [c] and [cf]. We further scale by the number of filament sites FT, writing Caf = [cf/FT], Ca = [c]/FT and introducing the new constants C = CT/FT and S = ST/FT. Since the number of bound filament sites cannot exceed FT, Caf≤1, Ca≤C, and Caf = 1 when all of the filaments are bound. Although appropriate values for C and S are not known, general knowledge of skeletal muscle indicates that C is large enough for the filament binding sites to be saturated during tetanic stimulation and that S is large enough to reduce free calcium to a negligible amount during rest. We obtain similar data fits over a range of values for these constants, so we arbitrarily set C = 2 and S = 6. Thus twice as much calcium is available than is necessary to bind all of the filaments and thrice as many binding sites are available in the SR than are required to bind all the calcium.
Following Hill [37], each myotome is modeled as a contractile element (CE) in series with an elastic element (SE). (The Hill model includes a second elastic element in parallel [38], but for our purposes this can be included in the linear spring of Figure 4b.) Because they are in series, the CE and SE experience equal forces at steady-state. We begin by describing them separately, as a force P exerted by the SE, and a force Pc developed by the active element CE.
The SE is modelled as a linear spring and hence P is proportional to the length ls of this element minus its resting length ls0: P = μs(ls−ls0). This force is never negative. The total length L of the segment is the sum of ls and the length lc of the contractile element. The length and velocity vc = l ˙c of the contractile element are therefore given in terms of the length and velocity V = L ˙ of the segment and the force P as follows:(27)(28)
We assume that the the force Pc exerted by the contractile element can be described by independent multiplicative factors of its length lc and velocity vc,(29)where the constant P0 is the force exerted in isometric tetanic contraction (Caf = 1) at the optimum length lc0. The functions λ(lc) and α(vc) are estimated from force measurements (described below), from which we obtain a piecewise linear function for α and a quadratic for λ:(30)(31)We additionally restrict these functions such that 0≤α(vc)≤αmax and 0≤λ(lc)≤1. The fact that αp>αm>0 (see Table 1) reflects the ability of muscle fibers to exert progressively greater forces during lengthening than in shortening.
If we set Pc = P, the calculation suffers from instability, and in reality the stretch of the SE due to activation of the CE is not instantaneous. We therefore model the transfer of force from the CE to the SE by simple linear kinetics:(32)
Combining Equations 22–32 and using the three conserved quantities CT, ST, and FT, we obtain three ODEs for the concentrations of free calcium, bound calcium and the force exerted by the preparation:(33)(34)(35)
The parameters of the model are determined from analysis of the data of [14], as follows. μs and ls0 are determined from quick-release experiments [37]. The maximum values of force P0 in the three isometric experiments (Figure 5) are used in Equations 27, 29, and 31 to determine the values of λ2 and lc0. The results of constant-velocity ramp experiments are then used with Equations 27–31 and the parameters λ2 and lc0 to determine αm and αp. The limiting value of αmax was not determined in [14], so αmax is taken from results in dogfish [39]. In practice, results vary little over a range of values for αmax.
We set the time constant k5 = 100 s−1, so that Pc closely tracks P. The remaining time constants k1, k2, k3, and k4 are found by fitting force trajectories from the experimental data, using the least-squares curve-fitting facilities in the software XPPAUT devised by G. Bard Ermentrout and available at http://www.pitt.edu/phase/.
The parameters k1–k4 are fit in two different ways. The isometric fit follows the approach in [14] by using only data from the isometric experiments at the three lengths L = 2.7 and 2.7±0.125. The main aim of [14] was to show that a model based on isometric and constant-velocity experiments could be used to approximately predict forces that occur during swimming, even though it excludes known properties such as the observation that the length-tension and force-velocity relationships change during muscle activation and relaxation [36]. Such secondary features cause discrepancies between the predictions and the data seen in the sinusoidal traces of Figure 5, but the model nonetheless produces forces during sinusoidal movement that capture the overall behavior well.
The present study demands our best estimate of force development during swimming, and for this reason we have made a second, dynamic fit of the time constants k1–k4 based not on isometric data but on muscle force data during sinusoidal movement at 1 Hz. To best match swimming behavior, we chose the experiment with a delay of 0.1 from onset of stimulation to onset of shortening (cf. Figure 1), and as the upper panels of Figure 7 show, the resulting force trace is much closer to the data than the fit to isometric data. The discrepancy between the isometric data and the prediction using these parameters is primarily in the repolarisation phase (Figure 7, lower panel), reflecting the model's inadequacy during this phase of the force trajectory. Values for both fits, along with the other muscle parameters, are given in Table 1. The most striking difference is in the rate constant k2 (uptake of free Ca2+ by the SR), which doubles. Using this, the dynamic fit captures the rapid force decay seen in the sinusoidal data at low phase delays.
Sinusoidal forcing data were only available at 1 Hz [14] and in most of the simulations described below we retain this frequency, but we also briefly investigate swimming behavior at 2 Hz. The muscle parameters are listed in Table 1. It is worth noting that neither set of time constants is unique: in both cases it was possible to find more than one set of time constants that gave a good fit, by starting from different initial guesses. The primary goal of this study is not to discover accurate parameters, but to find a good prediction of muscle behaviour for use in our neuromechanical model.
Muscle dynamics is incorporated into the discretized rod model as follows. The forces PRi and PLi generated by the right and left myotomes associated with the ith link are modeled by two sets of the three Equations 33–35, with maximal force P0 scaled by cross-sectional body area at that location. Thus, if the entire body length is actuated, 6(N−1) first order ODEs describe the muscle forces in the N-link chain, and with the 3N second order ODEs in Equations 10–12 they jointly determine the body dynamics. Unlike in the simplified model of [13], the time course of force development now depends on the proportion of activated filaments (Caf) and on the lengths and velocities of the muscle fibers, via appropriately scaled versions of Equations 27–32. At joint i the lengths and velocities are(36)(see Figure 4 and the discussion in the preceding subsections). Equations 36 provide the explicit coupling between the muscle and body equations. As in [13] the preferred curvature at joint i is given by , and the force at each segment is given as a scaled multiple of the force PR,L of the fibers on either side of the joint. Since the number of fibers typically depends on cross-sectional area, our first approach was to take fR,L∝abPR,L, giving a preferred curvature κi∝(PR−PL)/b, but simulations with such a relation exhibited much greater motions toward the tail than those seen in the swimming animal. After extensive simulations with various scalings (not shown), we found that scaling the preferred curvature as κi∝b2(PR−PL) and the stiffness as EI∝ab2 provides the best qualitative match to behavior. This suggests that the Young's modulus, and hence , increases along the length of the body, while not only the magnitude, but also the density of muscle forces decreases. The former is consistent with the fact that the notochord takes up a proportionally greater portion of the cross-section of the animal toward the tail. This scaling thus corresponds to , and fR,L∝ab3PR,L, cf. the middle equation of Equation 21.
In the experiments described above the stimulus applied was tetanic, which does not occur normally. We assume that during swimming the muscle is stimulated in such a manner that it can be scaled linearly with respect to the tetanic stimulus. We thus scale the forces with a constant ζ that is chosen ad hoc, so that(37)Equation 37 completes the loop, so that upon imposing a traveling wave of activation which releases calcium by setting k1 and k2 of Equation 33 on and off in a piecewise constant square wave (approximating the EMG recordings [29]), we obtain a closed system of ODEs.
We now explore the behavior of the discretized model actuated by forces generated by segmental muscles. It is important to note that the entire body length of a fish is usually not equipped with swimming muscles. In lamprey the head and part of the gill region lack such muscles, and we shall henceforth assume they occupy 1/10 of the total body length, and that their passive material properties are the same as those of the rest of the body. We shall refer to the remaining 9/10 of the body that is capable of activation [40] as the activation region.
Only a fraction of the myotomes are activated on either side of the body at each instant, in a region that travels from head to tail during normal swimming. The mean temporal duration of activation at a given location is ≈0.36 of the mean cycle duration [7] (see Figure 1). This defines the square wave referred to above, and implies that the activated portion on either side also has length 0.36 times the activation wavelength. We will generally assume that the activation wavelength is one body length [6] (i.e., greater than the length of the activation region). Unless otherwise stated, in the simulations reported below we apply a stimulation rate of 1 Hz, and the activation wave thus travels down the body at speed of 1 body length/s.
The value of the Young's modulus E (or , cf. Equation 19) for the lamprey is not known with any precision. However, the studies in [10] suggest a value of E≈0.1 MPa for the eel, and the lamprey's passive stiffness is thought to be much smaller. Indeed, the stiffness of an anesthetized lamprey is so low that is difficult to measure, but preliminary studies suggest that values in the range 10−3−10−2 MPa are not unreasonable [41]. Other parameters for which we have no firm lamprey data are the overall scale of the muscle force ζ and the damping . The values specified here are selected based on extensive exploratory simulations and the studies of passive elastic and geometric properties in [13].
We take a tapered rod of length l = 21 cm with constant height 2a = 2 cm, to account for the dorsal and anal fins, and width 2b(s) given by b(s) = 1−(4/5)(s/l) cm. Unless otherwise stated, the following body parameter values are used throughout this section: Young's Modulus E = 10−3 MPa, damping , and force density ζ = 0.05 N/m3. Fluid density and viscosity are ρf = 1 g/cm3 and μ = 10−3 Pa·s – the values for fresh water – and the body is discretized into N = 21 links. We refer to these as the standard or control parameters.
We use a numerical method adapted from that of [13], the appendix of which contains a detailed description. The main difference is incorporation of two sets of the muscle force Equations 33–35 at each joint. These are in turn coupled with the rod equations through the preferred curvature κ, explicitly via the length and velocity of the muscle as described by Equations 36 above. The method employs discrete versions of the integrated constraint equations in Equation 1 that express link positions in terms of that of the head (x1,y1) and the link angles φi, thus guaranteeing that the inextensibility constraint is precisely satisfied for the discrete system in Equations 9–13 and eliminating the need to solve ODEs for (xi,yi) i = 2, …, N. Since the head region lacks activation, the number of ODEs required to describe muscle forces reduces to 6(N−1)×(9/10) = 108 in the present case.
Simulations readily yield results that are qualitatively similar to real anguilliform swimmers. For example, Figure 8 compares tracings from a film of a lamprey in a swimmill that approximate its body centerline at various times with centerline snapshots from a model simulation. The characteristic swimming behavior is clearly captured, in particular the larger amplitude at the tail end. Figure 9 shows snapshots of the body over one activation cycle. When the force density magnitude fR,L is the same on both sides, the center of mass travels in a nearly straight line, with small lateral oscillations that arise due to slight asymmetries in body shape (see section 4.3.2 of [13]). The mechanical wave travels down the body at a speed of 0.78 body lengths/s, producing a forward swimming speed that rises asymptotically to a value of 0.40 body lengths per second, giving a speed ratio or slip of 0.51. Slip values are not available for swimming lamprey, but the expected value for eels swimming at the same speed is 0.66 [23]. It is likely that eels are more efficient swimmers than lampreys, since they do not exhibit the side to side movements of the head seen in lamprey (Figure 8). Turns can be evoked by reducing the magnitude of force density on one side, so that the average of the rod's intrinsic curvature is nonzero: see Figure 10.
We now attempt to determine the mechanism(s) causing the difference in wave speeds of activation (EMG) and curvature. In the simulations of [13], the preferred curvature κ = κ(s−ct) was externally prescribed, specifically, as that of a traveling sine wave. The curvature φs that emerged depended on the passive elastic properties of the rod and the hydrodynamic body forces, but in [13] κ itself was independent of the body dynamics and of φs. In the present model the ODEs in Equations 33–35 couple the preferred curvature to the state of the rod, via the length and contraction speed of muscle fibers (cf. Equation 36) that appear in the functions α(vc) and λ(lc) of Equations 29–31. Hence κ now depends on φs, and we are able to investigate what role this dependence plays in wave propagation.
Figure 11 shows the relative timing of activation, muscle force development, and muscle shortening in a typical simulation. Activation waves travel the length of the active region with a frequency of 1 Hz, as in Figure 9. The left panel shows time courses of muscle length and force in two segments on the same side of the body; the right panel shows the relative timing of activation and curvature in the same format as Figure 1. We calculate the average wave speed of the maximal concave and convex curvatures by linear regression, first approximating the angle φ(s) along the rod by a cubic spline interpolant of the joint angles φi. This yields a continuous function of arclength s, from which we estimate the maximal and minimal curvatures. In all cases the mean speeds of convex and concave curvatures agree to 3 decimal places, so we report a single ratio of curvature speed to activation wave speed. As in the lamprey (Figure 1), the mechanical wave is slower than the activation wave, the wave speed ratio being 0.78, within the range of values 0.72±0.07 (SD) observed in lamprey [6].
The wave speed difference could be due to several separate effects, or to some combination of them. Ostensibly, any or all of the following could play roles:
We now examine these items individually and in combination. First we consider the effect of fluid loading. By setting W≡0, we remove fluid forces, a situation approximated in the laboratory by stimulating a lamprey to “swim” on a slippery bench [26]. Figure 12 (left panel) shows the results of one such simulation. A difference in wave speeds persists, although in this case the speed of the mechanical wave tends to decrease slightly midbody and then increase toward the tail. Eliminating hydrodynamic reaction forces has the effect of further reducing what is already a very small body stiffness. Under the same muscle activations the rod flops around violently.
We varied several parameters, including stiffness, viscoelastic damping, the length of the activated region and wavelength of the activation, and body geometry. As noted above, our value of Young's modulus, E = 10−3 MPa, is extremely small, but simulations with higher values did not yield realistic results. For example, with E≈0.1 MPa and an increase in muscle force density by a factor of 3, the ratio of curvature to activation speeds is 0.9, mean swimming speed increases to 0.5 body lengths per second, but the phase delay between activation and shortening is approximately zero throughout the rod.
We found that two further properties are necessary to create the observed difference in activation and response wave speeds: taper in the body and the presence of viscoelastic damping. Figure 12 (middle panel) shows results of simulations performed on an untapered rod (b(s)≡1), for which the wave speeds become almost identical. The strongest effect of taper is probably via the reduced muscle cross section, and hence smaller force generation, toward the tail (recall that the “hydrodynamic cross section” used in Equation 7 remains fixed at a = 1 cm, and that a wave speed difference persists in the absence of hydrodynamic forces.) The right panel of Figure 12 shows results of simulations performed without viscoelastic damping (). In this case the speed ratio also increases significantly, to 0.96.
Next we consider the effects of eliminating the dependence of muscle force on length and/or velocity, by setting the functions λ(lc) and/or α(vc) of Equations 27 and 28 identically equal to constants. For the former we choose λ(lc)≡0.86, because this is the value of λ(lc) at the middle length (2.7 mm) used in the isometric experiments of Figure 5, and it corresponds to the average length during typical swimming motions [14]. For the latter we take α(vc)≡1, corresponding to zero velocity. Removing both effects and maintaining all other parameter values, including force density ζ = 0.05 N/m3, we find that the wave speeds are approximately equal, but that mechanical wave amplitudes become unrealistically large (Figure 13, left panels). Upon reducing ζ to 0.025 N/m3 to achieve reasonable amplitudes, we obtain the result shown in the right panels of Figure 13: i.e., a speed ratio nearly equal to the case in which length and velocity dependence are present, but body motions are now more pronounced near the head, unlike the shapes of Figure 8. The swimming speed also drops slightly from 0.40 to 0.39 body lengths per second, and, as reported in the following subsection, the swimming efficiency is sharply reduced when length and velocity dependence are removed.
The multiplicative dependence of muscle force on the factors λ(lc) and α(vc) also allows us to separate these effects. In the simulation illustrated in the left panel of Figure 14, we set λ(lc)≡0.86 but retain the function α(vc), thus eliminating length dependence alone. The resulting speed ratio of 0.79 is almost unchanged from the control value for the full model (cf. Figure 11). The right panel of Figure 14 shows the result when only velocity dependence is abolished, by setting α(vc)≡1 and retaining λ(lc). The speed ratio 0f 0.77 is again nearly equal to the control value, although phase lags are reduced over the first half of the body length. Thus, removing either length or velocity dependence alone does not significantly affect the difference in wave speeds. In both these cases, and all those to follow, we retained the standard force density ζ = 0.05 N/m3.
The difference in wave speeds changes the relative timing between muscle activation and shortening as waves travel down the cord, as shown in Figure 1C. The changes in this relationship under all the conditions that we have investigated are illustrated in Figure 15, in which the delay from the beginning of muscle activation to the time of maximal convex curvature (approximately the beginning of shortening) is plotted against body position. The broken line at the top reproduces values from Figure 1C, experiments of [6] showing that the delay increases from 0.10 of a cycle at 24% of the body length to 0.23 at 76% body length. Data from the full control simulation of Figure 11 are shown by the thick blue line. Although the resulting phase lags are smaller than those observed in the animal, the phase gradient is qualitatively correct. Data from the simulations of Figure 12 are also shown, illustrating that with these changes in mechanical properties, the phase lag values are very different from normal. Abolition of length and velocity dependence, as in Figure 13, has little effect, when accompanied by halving the force density. Removing only the velocity dependence, as in Figure 14 (right panel), however, abolishes the phase lag in the most rostral segment.
The preceding simulations were all done for swimming at 1 Hz, the frequency for which muscle force data is available. Lampreys can of course swim over a range of speeds, by varying both activation levels and frequencies. Ichthyomyzon unicuspis has been recorded as swimming at frequencies up to ≈7 Hz., although this probobaly does not represent steady swimming. To verify that our model can accomodate frequency variations, we performed simulations at 2 Hz, keeping all other parameters at their standard values. Figure 16 shows that body shapes and amplitudes remain similar to those at 1 Hz, although the wave speed difference is somewhat magnified, the ratio decreasing to 0.71.
As noted above, removing the length and velocity dependence in muscle forces, while simultaneously halving the force density ζ, leads to a nearly identical ratio of curvature to activation wave speeds with only a slight reduction in swimming speed. Since nonlinear muscle properties are not required to produce the observed speed difference, we were prompted to ask what other differences they make. Here we investigate their effect on swimming efficiency, by comparing the work done by the muscles over a full activation cycle with length and velocity dependence present and absent. We calculate the work done by the muscles on either side of joint i by computing the integralswhere fRi,Li and VRi,Li are the right- and left-hand muscle forces and velocities defined in the last two subsections of Methods (the negative sign is due to our convention that VRi,Li are lengthening velocities).
The left panel of Figure 17 shows the work done at each joint, illustrating that, in spite of the reduced force density used for the case without length and velocity dependence, 67% more work is done than when length and velocity dependence are included, although there is a slight reduction in swimming speed. The difference is largest near the head; the work done near the tail being slightly larger for the latter case. The center and right panels show time courses of work done over one cycle at specific locations in these two regions (joints 3 and 18), with activation beginning at the time on the left axis in both cases. In addition to substantial differences in magnitudes due to reduced muscle cross section near the tail, these panels reveal that negative work is done at the tail in the beginning of the activation phase, while muscles are still lengthening. As we have noted, this may play a role in stiffening the tail as it moves laterally through the water.
Overall, these results suggest that the length and speed dependencies of the muscle fibers may provide a mechanical advantage to the animal in swimming.
This paper is primarily concerned with the role of muscle activation in the production of anguilliform swimming motions: a process that involves multipath coupling among active filaments, passive body tissues, hydrodynamic reaction forces, and proprioceptive and exteroceptive sensory feedback. To better parse this complex coupled system, here we address the influence of “feedforward” neuromechanical coupling alone by means of a mathematical model.
Our model substantially extends previous ones [12],[13],[42] by its inclusion of nonlinear muscle dynamics, which is characterised by known physiological properties with parameters fitted to experimental data. Coupled with appropriate passive viscoelasticity and geometry of the body, this gives rise to a difference in the wave speeds of neural activation and mechanical response, as seen in swimming animals, and the model enables us to investigate the sources of this difference. We find that three factors are primarily responsible for it and for the associated lags between activation and curvature onsets, namely: viscoelastic damping, taper, and the nonlinear dependence of muscle force on length and shortening velocity. The first two factors, which are properties of passive tissues and body geometry, are necessary for the appearance of the wave speed difference. The third factor, nonlinear muscle dynamics, contributes to the values of the changing phase lags, and may also contribute to the efficiency of swimming.
Figure 15 shows that the phase relationship between muscle activation and shortening produced by the model is similar to that seen in the lamprey. Significantly better data fits can be obtained by varying parameters outside the normal ranges, but rather than explore this systematically, we have instead used parameter values that best describe the lamprey.
The present study illustrates the power of integrative mathematical models in revealing biological function, by allowing “experiments” which cannot be done on animals. It partially answers questions posed by Altringham and Ellerby, who conjectured that the progressive phase lag is associated with “change in muscle function along the body [11].” Our study shows that, at least for anguilliform swimmers, muscle and mechanical properties need not vary along the body for wave speed differences to emerge. It also shows that, during steady swimming, proprioceptive feedback is not necessary to produce this basic phenomenon. This supports the suggestion of Brown and Loeb that, in stereotypical movements, neural feedback (reflexes) can be partially or wholly replaced by mechanical feedback (called “preflexes” by Brown and Loeb (section 3 of [43]), who define a preflex as “the zero-delay, intrinsic response of a neuromusculo-skeletal system to a perturbation.”), and therefore might not be required for stability [43]–[45]. Further model-based and experimental support for this hypothesis has recently emerged in legged locomotion studies [15]. However, mechanosensitive “edge cells” exist within the lamprey's spinal cord, which can influence the timing of muscle force generation and phase relationships via feedback to the CPG and motoneurons [46]. This mechanism may account for the deficit in phase lags produced by the model (Figure 15), and it is are presumably important during changing conditions and maneuvers.
The muscle model we described in Methods cannot perfectly fit both the isometric and the sinusiodal forcing data. We chose to fit it to sinusoidal data with an activation-to-curvature phase difference of 0.1, close to values seen in the data and the control simulations. This is not ideal, and may influence the results described in the results section. We plan to extend the model to include secondary muscle properties responsible for the discrepancies in its predictions. Moreover, we have used a linear model for flexural stiffness (M = EI(φs−κ), Equation 5), although the lamprey's body stiffness is nonlinear. More accurate estimates of body stiffness may also influence the results.
In our discretization the arms to which muscles are attached project perpendicularly from the center of each link toward the periphery (see Figure 4). In the lamprey, however, the myosepta to which the swimming muscles attach project obliquely backwards from the notochord toward the body wall so that the muscle layers interleave in a somewhat complicated fashion (albeit considerably less complicated than in bony fish; see [11]). We have not examined the consequences of this attachment geometry, but it can be expected to affect torques at the joints, and we intend to include it in a future study. It is of interest to note, however, that Katz et al. [47] have shown that in spite of more more complicated interleaving of muscle layers in teleost fish, the swimming muscles undergo length changes similar to those expected for a homogeneous, continuous beam, and that curvature of the midline gives a reliable measure of muscle length at any point along the body.
A further shortcoming of the present study, also noted in the methods section, is our use of an oversimplified model for fluid reaction forces. While Taylor's approximation in Equation 7 suffices for straight rods in uniform steady flow, it does not capture unsteady effects such as vortex shedding that are characteristic of swimming. These effects are likely important not only in creating propulsive thrust [22],[23], but the resulting reaction forces on the animal may also influence the speed at which the mechanical wave of curvature travels along its body. This would in turn affect the mechanical waves shown in Figure 11, perhaps changing the relative speed of activation and response. A more realistic model of vortex generation will also be needed to determine if negative work and tail stiffening are important in thrust generation, and to enable more definitive studies of swimming efficiency.
We also propose to use the present model, with the further addition of distributed CPG and motoneuron models [33],[48], to study proprioceptive feedback mechanisms in lamprey. In particular, it will allow us to investigate the influence of the aforementioned edge cells on the timing of muscle force generation. In recent experiments the isolated notochord/spinal cord preparation is rhythmically bent from side to side and the resulting edge cell feedback to motoneurons and CPG interneurons studied [49] (cf. [46]). This work complements our model in that it removes muscle activation, body elasticity and hydrodynamic forces, to reveal how an isolated sensory pathway can influence CPG phase and frequency relationships.
|
10.1371/journal.pcbi.1002305 | Balanced Synaptic Input Shapes the Correlation between Neural Spike Trains | Stimulus properties, attention, and behavioral context influence correlations between the spike times produced by a pair of neurons. However, the biophysical mechanisms that modulate these correlations are poorly understood. With a combined theoretical and experimental approach, we show that the rate of balanced excitatory and inhibitory synaptic input modulates the magnitude and timescale of pairwise spike train correlation. High rate synaptic inputs promote spike time synchrony rather than long timescale spike rate correlations, while low rate synaptic inputs produce opposite results. This correlation shaping is due to a combination of enhanced high frequency input transfer and reduced firing rate gain in the high input rate state compared to the low state. Our study extends neural modulation from single neuron responses to population activity, a necessary step in understanding how the dynamics and processing of neural activity change across distinct brain states.
| Neurons in sensory, motor, and cognitive regions of the nervous system integrate synaptic input and output trains of action potentials (spikes). A critical feature of neural computation is the ability for neurons to modulate their spike train response to a given input, allowing task context or past history to affect the flow of information in the brain. The mechanisms that modulate the input-output transfer of single neurons have received significant attention. However, neural computation involves the coordinated activity of populations of neurons, and the mechanisms that modulate the correlation between spike trains from pairs of neurons are relatively unexplored. We show that the level of excitatory and inhibitory input that a neuron receives modulates not only the sensitivity of a single neuron's response to input, but also the magnitude and timescale of correlated spiking activity of pairs of neurons receiving a common synaptic drive. Thus, while modulatory synaptic activity has been traditionally studied from a single neuron perspective, it can also shape the coordinated activity of a population of neurons.
| Correlations between the spike trains of neuron pairs are observed throughout the central nervous system [1]. The correlation between a pair of neurons' spike trains can change depending on the state of their neural circuit. For instance, correlated neural activity is altered by stimulus properties [2], [3], anesthetics [4], [5], stimulus adaptation [6], focus of spatial attention [7], [8], and the behavioral context of a task [9]. The level of spike train correlation between neuron pairs has implications for the accuracy of population codes [10], the formation of neural assemblies [11], and the propagation of neural activity [12], [13]. Nonetheless, only recently has attention been given to the mechanisms by which correlated activity is modulated [14], [15], [16], .
Cortical neurons receive a mixture of excitatory and inhibitory synaptic inputs, resulting in spiking activity that is driven by input uctuations rather than the input mean [21], [22]. This state is often described as balanced, to denote that the mean excitatory and inhibitory inputs that neurons receive are approximately equal [23], [24]. Balanced activity is inuenced by stimulus properties and history [25], [21], as well as internal brain state [26]. These changes can modulate the integration properties of single neurons, strongly inuencing neuronal activity [22]. For example, increases in the firing rate of balanced pre-synaptic activity afferent to a neuron can reduce single neuron firing rate gain [27], [28], [29], [30], [31], [32]. Further, an increase in the temporal correlation between the arrival times of excitatory pre-synaptic inputs increases the firing rate of a post-synaptic target neuron [33], [34], [35], while correlations between excitatory and inhibitory inputs can reduce output activity [34], [36]. The impact of such shifts in the temporal structure of synaptic input is amplified when the post-synaptic cell has a small integration timescale, as expected for neurons in the high input rate, balanced state [22]. These examples deal with synaptic activity convergent to a single target cell. However, what is less studied is the role that the balanced state plays in modulating the responses of a pair of neurons subject to a common synaptic input. In this study, we consider this latter scenario and show that shifts in balanced pre-synaptic population activity modulate the magnitude and timescale of the correlations of spike trains from pairs of post-synaptic neurons.
We first explore a model system and show that output spike train correlations from a pair of neurons are modulated by varying the rate of uctuating, balanced excitatory and inhibitory inputs. Specifically, we demonstrate that an increase synaptic input rate leads to an increase of short-timescale output correlation (i.e. precise spike synchrony) while correlation at long timescales (i.e firing rate co-variation) remains unaffected, or even decreases. Due to the differential affects of our mechanism on short and long timescale spiking activity we label the combined modulation correlation shaping. Correlation shaping has been observed in various sensory systems [2], [3], [37], [38], yet the core mechanisms underlying the modulation remain unknown. We present linear response analysis showing that the enhancement of output synchrony through an increase of input rate results from a shift in single neuron integration properties that favors the transfer of high frequency inputs. Dynamic clamp recordings from cortical neurons verify our theoretical predictions. Finally, in a feedforward network model, we show how correlation shaping supports a selective propagation of network responses, so that activity can be gated by correlations in complex neuronal networks. In total, our work extends mechanisms of single neuron firing rate control include the control of pairwise correlations, thereby providing a bridge between single neuron and network state modulation.
We modeled neurons as leaky integrate-and-fire units receiving conductance input [39]. Each neuron had an intrinsic timescale ms and leak reversal potential mV. Excitatory and inhibitory synaptic input caused conductance changes and with reversal potentials mV and mV so that the membrane potential dynamics followed:When reached a threshold voltage mV, the neuron spiked and the voltage was reset to mV.
We modeled the excitatory and inhibitory synaptic conductances as Poisson processes with rates and consisting of series of -functions with heights and . This framework was used for all of the simulations presented and provides a minimal model that captures our main results (for simulations of other models, see Supplementary Figures). These inputs consisted of independent processes private to each neuron as well as a shared component presynaptic to all neurons, yielding where superscripts and denote independent and shared components, respectively. For large rates, this input was approximated as a diffusion process [39], [40], [41], [42](Figure S1):where was a Gaussian white noise process with unit intensity. This allowed us to write our voltage equation in the form(1)where , , and . Note that as the rates of excitation and inhibition and increase in a balanced manner, decreases, increases, and does not change substantially because of the excitation and inhibition balance.
For our simulations and calculations, we set . This approximation ignored the multiplicative nature of the noise, which in our simulations did not substantially change the results (Figure S1), since the change in and were sufficient to modulate neuronal responses. To simulate pairs of neurons receiving correlated input, we set the fluctuating input to each neuron to be(2)where was shared across both neurons while was independent for each neuron. We note that, although the correlation in output spike trains depended on the degree of pre-synaptic overlap, Eq. (2) shows that , and hence the firing rate of neurons in our model, was independent of . The rate of excitatory input in the low state was 1.50 kHz and 6.16 kHz in the high state, with the inhibitory rate chosen to elicit a firing rate of 15 Hz in both cases. Simulations were performed using an Euler-Maruyama numerical integration scheme with a simulation timestep of 0.005 ms.
We next developed a theoretical framework to study the behavior of the above system and compared our theory against simulations of the stochastic system. For completeness, we write the governing equations used to calculate the single neuron power spectrum and transfer function ; these techniques are fully presented in [42] and we refer the reader there for further details. Letting , the voltage distribution associated with the stochastic differential equation (1) obeys the Fokker-Planck equation:where is the probability flux [43]. The boundary conditions for the probability distribution and flux at threshold are and , where is the firing rate. Furthermore, the flux obeys for and is 0 otherwise.
For time independent and the steady state distribution obeys:Using the normalization condition , we can solve for the steady state firing rate .
In order to study the system's response to a correlated, fluctuating input, it is necessary to study the system's response to time-dependent inputs. This is done most effectively by writing a time-dependent Fokker-Planck equation in the Fourier domain:where () denotes the Fourier transform of and is computed with initial condition . Solving this equation yields the Fourier transform of the first passage time density [42]. The power spectrum , where is calculated from the well known renewal relation [44].
Finally, we compute the transfer function . Suppose that we add a time-varying periodic current to the right hand side of Eq (1). If we let be sufficiently small, we can compute the spike train response to these time-dependent modulations. Decomposing the probability density, flux, and firing rate into steady state and modulated components:and then solving the Fokker-Planck equation for the time-dependent terms, we obtain a new set of equations:with boundary conditionsThese equations were solved numerically [42] obtaining a solution for the transfer function .
Surgery: Somatosensory (S1) cortical slices were prepared from CBJ/Bl6 mice age P19-26. All surgical procedures followed the guidelines approved by the Carnegie Mellon Animal Welfare Committee. The mice were anesthetized with isoflourane and decapitated. The brain was exposed, removed from the skull and immersed, in ice cold oxygenated () ACSF (in mM: 125 NaCl, 2.5 KCl, 25 , 1.25 , 1.0 , 25 Dextrose, 2 ) (all chemicals from Sigma, USA). Coronal slices (300 m) of barrel cortex made using a vibratome (Leica, Place). The slices were maintained in ACSF at for 30 min then rested at room temperature () for 1 hr prior to recording ().
Electrophysiology: L2/3 pyramidal neurons were visualized using infrared-differential interference contrast microscopy (Olympus, Center Valley, PA). Whole cell, dynamic clamp recordings were performed using a MultiClamp 700B amplifier (Molecular Devices, Union City, CA). Data were low pass filtered (4 kHz) and digitized at 50 kHz using an ITC-18 (Instrutech, Mineola, NY) controlled by custom dynamic clamp software (R. Gerkin; http://rick.gerk.in/software/recording-artist/) written in IgorPro (Wavemetrics, Lake Oswego, OR). Pipettes were pulled from borosilicate glass (2.0 mm, outer diameter) on a Flaming/Brown micropipette puller (Sutter Instruments, Novato, CA) to a resistance of 6–10 M. The intracellular solution consisted of (in mM) 130 K-gluconate, 5 KCl, 2 , 4 ATP-Mg, 0.3 GTP, 10 HEPES, and 10 phosphocreatine.
Stimulation: Pyramidal cells (n = 8) were directly stimulated by a series (50–100 trials) of simulated noisy synaptic currents in dynamic clamp. Each trial was 4 s in duration with a 5 s inter-trial interval; the period of rest was used to ensure that stability of the recordings. For each trial, excitatory (: 0 mV) or inhibitory (: −60 mV) synaptic conductance inputs were simulated as Poisson distributed spike times convolved with alpha function . ( nS, nS, ms, 8 ms). The Poisson rates for excitatory and inhibitory inputs were equal to one another (), and were set to 3 kHz in the low state and 7.5 kHz in the high state. These rates were higher than in the simulations to ensure high spike time variability, since the input variability is attenuated by the finite temporal extent of the synaptic timescales. For each state, half of these inputs were common to all neurons stimulated and half were newly generated on each trial for every neuron. This produced an input correlation, , of 0.5 between any given pair of neurons. This setup permitted pairwise comparisons. Since the synaptic drive was subthreshold, a bias current (0.3–0.7 nA) was added such that the balanced conductance fluctuations produced a mean cortical firing rate of (4–6 Hz) in both the low and high states.
We studied a layered network in which a population of 100 leaky integrate-and-fire neurons (Layer 2) received balanced input from a pre-synaptic layer (Layer 1) with and provided excitatory input to two distinct downstream targets. Neurons in Layer 1 were assumed to be Poisson as in previous sections, and the total input to a Layer 2 neuron was therefore approximated by a diffusion process. In particular, the voltage dynamics of each Layer 2 neuron followed Eqs. 1 and 2.
The downstream target was also modeled as leaky integrate-and-fire neuron. Because we wished to fix the timescale of the downstream target, we assumed delta-function, current-based synapses so that the voltage of the downstream neuron followed:where indexes the neurons in Layer 2 and indexes the spikes in each Layer 2 neurons' spike train. We compared ms and ms. For ms, we set mV and for ms, mV so that the neurons fired at comparable rates given identical input. Other parameters, including leak, threshold, and reset voltages were identical to the model previously studied.
In general, it is difficult to determine the specific changes in a neural system's dynamics that cause changes in spike train correlations. We studied a framework in which common inputs drive the correlations between the spike trains of a pair of neurons [45], [46], [47]. If the degree of input correlation, , is small, a linear approximation relating to the output spike correlation, , is written as:Here the quantity , termed the correlation susceptibility, determines the extent to which two neurons' spike trains will be correlated given a fixed level of correlation between the inputs they receive [17].
Throughout this study, we focused on a pair of neurons that shift their output correlation () due to a change in their pre-synaptic drive (Figure 1A). Under our linear model, two simple explanations for the shift in output correlation are possible. First, the shift may simply reflect a change in the correlation of the inputs that the neuron pair receives (; Figure 1B). While this answer appears straightforward, understanding shifts in input correlation requires detailed anatomical knowledge of the network architecture, in the absence of which simplifying assumptions are required [48].
A second explanation for the shift in output correlation is a shift in correlation susceptibility (), even when the input correlation remains fixed (Figure 1C). Because relates the correlations in the spiking output of neurons to their common input, we expect to be sensitive to how each neuron integrates its input. Indeed, single neuron response properties such as firing rate and neural excitability determine the extent to which neurons become synchronized by shared input [49], [17], [18], [19], [20]. There has been substantial work on how single neuron properties, such as firing rates, are modulated [27], [28], [29], [30], [31], [32], [50], [51], [52], [53], [54],suggesting that should also be open to modulation. We focused on this second mechanism and established how modulations of single neuron responses also modulated pairwise correlations in cortical populations.
We first investigated the transfer of input correlations to output spike train correlations in a simplified two-neuron network. Each neuron received conductance-based, pre-synaptic inputs from a mixed population of excitatory and inhibitory neurons (Figure 2A). To model the stochastic nature of cortical activity, the arrival times of both excitation and inhibition were modeled as Poisson processes. We set the relative strengths and rates of excitation and inhibition so that the mean input was balanced [23], [24], and the average membrane potential was below spiking threshold. Balanced pre-synaptic activity results in large membrane fluctuations that trigger spikes in a random, aperiodic pattern, consistent with in vivo recordings from cortical neurons [21], [22].
Shifts in the activity level of a recurrent cortical population are observed in many neural systems and have been shown to affect the response properties of neurons in vitro and in vitro [22], [55], [31]. To explore the modulatory effects of balanced synaptic input, we considered the neuron model in two states: a low state, in which pre-synaptic input arrived at a low rate, and a high state, in which pre-synaptic input arrived at a high rate (Figure 2A). While the level of balanced fluctuations may lie on a continuum, we compared two representative points, analogous to high and low activity states in a cortical network [56], [26]. A clear consequence of the shift from low to high states was an increase in the variability of both the input current and membrane potential response, due to greater fluctuating input (Figure 2B). This increase of input variability was reflected in an increase in spiking variability, with the coefficient of variation of the inter-spike intervals increasing from 0.73 in the low state to 0.91 in the high state. A second consequence of an increase in pre-synaptic rate was the reduction of the membrane time constant (Figure 2B). This was expected, since the membrane time constant , with the membrane capacitance and the total membrane conductance [57]. As is roughly proportional to the pre-synaptic rates, an increase in the rate of synaptic input lead to a decrease in . Taken together, the shift from the low to high state evoked a more stochastic and faster membrane potential response.
We first examined the effect of balanced synaptic input on firing rate gain, the slope of the firing rate curve when plotted as a function of excitatory input strength. When the rate of balanced excitatory and inhibitory synaptic input changed from low to high, the neuron's firing rate gain was substantially reduced (Figure 2C). This gain decrease in the high background state has been studied extensively in theoretical and in vitro work [27], [28], [29], [30], [31], [32] as well in vivo under specific stimuli conditiona [31]. In the high state,larger membrane potential fluctuations increased firing rates for weak inputs. However, there was also a decrease of the net membrane input resistance, causing an increase in the rheobase current (minimum steady current required to recruit spiking). The combination of these two effects lead to an overall reduction in firing rate gain [29]. We next explored the consequences of gain modulation via balanced activity for correlation transfer by pairs of neurons.
To study the effects of balanced excitatory and inhibitory inputs on pairwise spike train correlations, we extended our model to include a pair of post-synaptic neurons receiving overlapping pre-synaptic inputs (Figure 3A). Previous work has shown that the output firing rate affects correlation susceptibility [17]. To preclude any firing rate-induced effects, the synaptic input was adjusted so that the average output firing rate of each neuron remained at 15 Hz in low and high states (Figure 2C). Furthermore, there was a fixed overlap in the input populations, so that the input correlation also remained constant in both network states (Figure 3A). Thus, any change in the output spike train correlation induced by changing synaptic input will be due exclusively to a shift in correlation susceptibility (Figure 1C).
We found that the timescale over which the two spike trains were correlated was dependent on the level of balanced synaptic activity (Figure 3A, Right). When the synaptic rate increased from the low to high state, the magnitude of the peak of the cross-correlation function near zero lag increased, reflecting greater spike time synchrony between the neurons. However, this increase was not present for longer lags, and the spike train cross-correlation function was unchanged or reduced for sufficiently long lags ( ms).
To quantify this change in output correlation over a range of timescales, we first counted the number of spikes and that the two neurons emitted in intervals of milliseconds. We next computed the spike count correlation as a function of window size:(3)where Cov and Var denote covariance and variance, respectively. In the framework of our simple circuit (Figure 3A), correlation in output spike trains was a consequence of a shared input correlation . For small , linear response theory [17] takes the output correlation to be a linear function of the input correlation (Figures 1B,C; 3B):(4)In our model, this linear relationship held for a range of , in both low and high states and at both short and long (Figure 3B). Further, the values produced were, in magnitude, consistent with in vivo recordings from a variety of systems [58], [3], [2], [6]. When comparing for the low and high states at fixed , a differential change of correlation at different timescales was evident. Specifically, for small (Figure 3B, = 3 ms), while for large (Figure 3B, = 50 ms). This differential modulation of correlation occured over a broad range of timescales, with and intersecting only once (Figure 3C), and we label the modulation a shaping of correlation [38]. This substantial change in both the magnitude and timescale of correlation must involve a nontrivial change in how the neurons process their inputs, since the input correlation and firing rate were the same in both low and high states. We note that the qualitative results of our study are also valid for larger (Figure S2) and different synaptic strengths (Figure S3).
Since as [59], changes in at small are necessarily smaller in magnitude. However, synchrony at short timescales can have large effects on downstream targets sensitive to coincident pre-synaptic spikes [12] and indeed the peak of the cross-correlation function increased substantially in the high state (Figure 3A, Right). To properly compare correlation shaping at small and large we considered the ratio , providing a relative measure across the low and high states. The ratio was a decreasing function of , with substantial changes in correlation at both short and long timescales (Figure 3D). The negative slope of the curve indicates that increases in the rate of balanced synaptic activity favor spike synchronization rather than long timescale correlation. Finally, the spectral measure of spike train coherence between the two spike trains in both states exhibited a decrease for low frequencies but a significant increase for high frequencies in the high state (Figure 3E). Here, the increase for high frequencies, which occurs over a broad range of frequency space, is related to the increase in short timescale synchrony, consistent with the spike count correlation shaping.
Correlation shaping is an unexpected feature of balanced synaptic activity. For subthreshold membrane potential dynamics (or any other linear system) the ratio is equal to 1 for all assuming a fixed input correlation (Figure 3D, gray line). The mechanism that shapes correlation transfer so to promote spike train synchronization over long timescale correlation in the high state (Figure 3D) is the focus of the next section.
Correlation shaping is a property of the joint statistics of a pair of neurons. However, since the input correlation was the same in the low and high states of our model, then the mechanism underlying the shaping is hypothesized to be related to changes in single neuron input integration and spike emission across the two synaptic states (Figure 1C rather than 1B). In this section, we show that correlation shaping is a consequence of a shift in the single neurons' frequency response across the low to the high input state.
The spike train auto-correlation and cross-correlation functions are written as:(5)where , with labeling the spike time from neuron . Here is the mean firing rate of neuron . We are interested in the joint spike count correlation for the neuron pair, where the spike count for neuron over a window of length is (we take the neuron's stochastic dynamics to be in statistical equilibrium). The spike count variance and covariance are related to integrals of auto- and cross-correlation functions [44], yielding an alternate expresion for :(6)In the second equality we have, for simplicity, assumed that (or equivalently ). These integrals can be transformed to the frequency domain, using the Wiener-Khinchin theorem [44] to relate correlation functions to their spectral analogues , yielding(7)Here is the Fourier transform of the triangular weighting term in Eq. (6). Our strategy was to relate the cross spectrum between the spike trains, , to single neuron integration properties.
Single neuron input-output transfer is typically expressed through its spectral transfer function . The transfer function measures the ratio of the amplitudes of a neuron's firing rate response and a small amplitude sinusoidal signal of frequency (Figure 4A). For very slow inputs, the transfer function equals the firing rate gain, since this measures the sensitivity of firing responses to static () inputs. For , is the susceptibility for a neuron's trial averaged response to be locked to a time varying signal. The transfer function is experimentally measurable [60], and is related to the more commonly reported spike triggered average [61]. In general, for neurons in the fluctuation-driven regime, is a decaying function of (Figure 4B).
If each neuron receives a small shared signal , then we can write the expectation of the Fourier transform of the spike train from neuron as:(8)where the brackets denote an average over repeated frozen presentations of the shared signal with different realizations of the independent noise driving the neurons [62]. Here, is the linear response of the system to the perturbation . Finally, averaging the quantity over different realizations of the process yields the cross-spectrum between neurons 1 and [17], [62], [63], [64]:(9)For the case of white noise input, we have that . With Eqs. (7) and (9) we calculated the spike count correlation coefficient between the two neurons receiving shared white noise input as(10)Our theory then relates single neuron transfer and power spectrum to the joint pairwise response .
The theoretical predictions given in Eq. (10) gave a very good quantitative match to simulations of the leaky integrate-and-fire neuron pair (Figures 3B–E, compare solid curves to points), capturing the correlation shaping between the two states. Eq. (10) has been previously derived [17], [18], however, the model neurons considered in those studies were current driven model neurons. We considered conductance driven model neurons, meaning that the calculation of and must account for the linked shifts of the membrane time constant and membrane potential fluctuations from the low to the high state (Figure 2B). For our conductance based integrate-and-fire model neurons, the quantities and were calculated by numerically integrating the Fokker-Planck equation associated with the stochastic differential equation expressed in Eq. (1) (see [42] and Methods). The distinction between current and conductance based neural integration will be shown to be critical for correlation shaping.
Before correlation shaping is related to the shifts in between the low and high states, we first discuss the dependence of susceptibility on the window size (Figure 3B). This dependence enters equation (Eq. (10)) through the weighting term , which determines the contribution of across frequency to . For long timescales (large ), is low-pass, so that only the neurons' response to low frequencies contributes to correlation susceptibility. In contrast, for short timescales (small ), weighs the transfer function approximately equally across all frequencies. Hence, the neurons' high frequency response determines precise spike synchrony. Indeed, for we have that , while limits to a constant function on . Therefore, for large , only the zero-frequency components of contribute to the integral, while for small , all frequencies contribute.
A mechanistic understanding of correlation shaping (Figure 3D) requires knowledge of how the rate of balanced synaptic activity affects the transfer function. As discussed previously, the increase in synaptic input from the low to the high state decreased the effective membrane time constant of the neuron while it increased the input variability (Figure 2B). The decrease in corresponded to a decrease in the timescale over which a neuron integrates inputs and hence an attenuation of the neuron's transfer function. For low frequency inputs, this reduction was precisely the firing rate gain control known to occur with increased synaptic input (Figure 2C). Increased variability and shunting due to heightened conductance reduced the neuron's ability to respond to slow depolarizing inputs. However, the reduction in the transfer function from the low to high state was not uniform across all frequencies (Figure 4C, Left). This was because the smaller value of in the high state enhanced the tracking of fast inputs, mitigating the attenuation of the transfer function for high frequencies. The combination of the non-uniform attenuation of the transfer function and increase in from the low to high state determined the shaping of the correlation susceptibility (see Eq. (10)).
To illustrate the shift in single neuron response between the low and high states, we considered the quantity , the strength of the input fluctuations multiplied by the input transfer function. The ratio was an increasing function of frequency (Figure 4C, Center), indicating that high frequency transfer is favored in the high state. In general, a favoring of high frequencies corresponds to a favoring of synchrony, measured over only small (since is nearly flat across for small ). Thus, the high state is expected to favor small correlation transfer compared to the low state (Figure 4C, Right). In contrast, for large which corresponds to low frequencies, correlation transfer was disfavored in the high state (since only weights low for large ). This ratio allowed us to intuitively link correlation shaping over different timescales to the shaping of the transfer function over different frequencies.
We argue above that a change in the effective membrane time constant is central to the correlation shaping we discuss. To demonstrate this fact, we computed the transfer function and correlations for a current-based model in which remained unchanged in the low and high state, although increased by the same amount. If firing rates were again fixed at 15 Hz, the transfer function was again reduced in the high state, but the ratio remained close to unity (Figure 4D, Left and Center). As a result, no substantial correlation shaping was observed (Figure 4D, Right). The above comparison shows that this shaping requires the modulation of cellular properties that is allowed by a conductance-based model.
Finally, we note that, although our analysis has focused on the numerator of Eq. (7), the denominator also affects the correlation for large time windows (Figure S4). For these values of , the denominator was increased in the high state, reflecting the higher variability of firing due to stronger input fluctuations. This further attenuated the value of for large in the high state.
To avoid changes in correlation owing to firing rate [17], we chose the balance between excitation and inhibition in previous sections so that firing rate was fixed across both the low and high states (Figure 2C). However, it is unlikely that firing rates will remain fixed as a network shifts from a low conductance to a high conductance state. Thus, it is important to understand how correlation shaping via balanced excitatory and inhibitory inputs interacts with the correlation changes expected due to firing rate changes. In this section we show how the modulations of correlation due to balanced excitatory and inhibitory inputs and those due to firing rate changes from imbalanced inputs are distinct.
The firing rates of our output neurons were determined by the input rate of both the excitatory () and inhibitory () inputs. In fact, for any desired output rate, there was a curve in () space that achieved that rate (Figure 5A). For moderate input rates, a balanced shift in input (approximately linear in and ) preserved output firing rate. A change in output firing rate (switching from one curve to another in Figure 5A), can occur from a shift in , a shift in , or some combination of the two. When we fixed to its value in the low state and increase so that the output rate increased, increased over all timescales (Figure 5B, top), as expected [17]. A similar effect occured if we repeat this in the high state (Figure 5B, bottom). Thus, the modulation of by a rate change due to an imbalanced shift of simply scales for all (collapsed blue and orange curves in Figure 5C). Nevertheless, after correcting for the rate scaling of , the shaping of correlation between the low and high states remained clear (Figure 5C), demonstrating that correlation shaping due to a change from low to high states is distinct from correlation shifts due to arbitrary output firing rate changes.
To illustrate this, we considered a shift from 8 Hz in the low state to 35 Hz in the high state. In the shift from the low to high state, the effective membrane timescale shifted from 10.8 to 2.9 ms and the amplitude of the input fluctuations from 0.16 to 0.37 nA. These shifts changed significantly (as discussed in the previous section), and changed the timescales over which the neuron pair was correlated. This was contrasted by a shift from 8 Hz to 35 Hz in the low state: a change in firing rate without a change between low and high states. Here, shifted from 10.8 to 10.2 ms and the input fluctuations from .16 to .18 nA, having little influence on other than a uniform scaling due to the output rate change. In total, by changing both and , it was possible to not only change the output firing rate so as to amplify or attenuate , but also to shape the timescales over which a neuron pair was correlated.
Our two-neuron framework for studying correlation transfer (Figure 1A) permited an experimental verification of correlation shaping with balanced, fluctuating conductance inputs. We performed in vitro patch clamp recordings from cortical pyramidal neurons receiving simulated excitatory and inhibitory inputs. Unlike past experimental studies of correlation transfer [16], [17], our model involved conductance-based, rather than current-based synapses. Therefore, we simulated synaptic input using dynamic clamp [65] (see Methods), which affected the membrane integration timescale as well as membrane potential variability. We chose maximal excitatory and inhibitory conductances of 1 nS and synaptic timescales of 6 and 8 ms, respectively, producing a synaptic input that was more biophysically realistic than the diffusion process used in previous sections (Figure 6A). The shift from low to high state caused a near two-fold reduction in firing rate gain (Figure 6B), in qualitative agreement with our model simulations (Figure 2C) and past dynamic clamp studies [29]. Further, as was done in the model, we set the synaptic balance in the low and high states to produce approximately the same firing rate ( Hz in the low state and Hz in the high state).
The correlated input for a given neuron pair was a mixture of shared and independent excitatory and inhibitory inputs, mimicking the input provided to the model (Figure 3A). The partial overlap in the synaptic input produced correlated membrane potential and spike dynamics for every neuron pair in both the low and high states. Our recorded spike trains showed a dependence of spike count correlation on that was qualitatively similar to that of the model, apparent in the ratio of in the high and low states (Figure 6C). The ratio was a decreasing function of , indicating a bias toward synchrony in the high state compared to the low state. This shape was consistent with our model results (Figure 3D), although the ratio did not fall substantially below unity in the limit of large . This suggested that the decrease in gain and the increase in variability from the low to high state were of similar magnitudes, since in the limit of large correlation susceptibility is proportional to (Methods). A conductance-based simulation using the same synaptic parameters used for dynamic clamp stimulation produced results in agreement with the experiment (Figure S5). The favoring of synchrony ( = 2 ms) over long timescale correlation ( = 200 ms) in the high state was statistically significant in a pairwise analysis across the dataset (Figure 6C, inset; , paired t-test). The experiments demonstrated that an increase in the rate of balanced conductance input shapes pairwise correlation so as to favor synchronization over long timescale correlation, thereby verifying the main theoretical predictions of our study.
Our theoretical treatment has ignored the timescale of synaptic input, and has associated all filtering to the membrane and spike properties of the model (Figure 4). Correlation transfer with realistic synaptic timescales did quantitatively differ from the case with instantaneous synaptic input (Figure S5B). Nevertheless, our theoretical work captured the main effects of correlation shaping when synaptic timescales were realistic (Figures 6 and S5). This is because only the effective membrane time constant was sensitive to a shift in input firing rate, which our theory accounts for, while synaptic filtering did not change between low and high states. We remark that, for synapses with very long timescales, correlation shaping should only be present for large , since correlations at small will be negligible.
The spike train correlations between neuron pairs substantially influence the propagation of neural activity in feedforward architectures [13]. For example, while our study has so far focused on the transfer of correlation for neuron pairs receiving common input, the firing rate of a single downstream neuron also depends on the correlation between neurons in its pre-synaptic pool [12]. If the integration timescale of the downstream target is small, only precise spike synchrony will effectively drive the neuron. In contrast, neurons that slowly integrate inputs will be sensitive to long timescale correlations. In our study, we demonstrated that an increase in the rate of synaptic input increases spike count correlation at small while simultaneously decreasing the correlation at large (Figure 3D). We therefore expected that this correlation shaping would influence the extent to which activity can be propagated to a downstream layer. Further, that the magnitude of this effect would depend on the integration timescales of the downstream targets.
As an illustration of this effect in a simplified system, we studied the firing rate of a downstream neuron receiving input from an upstream population of correlated neurons (Figure 7A; Methods). The level of synaptic drive from layer 1 shaped the correlation of pairs of layer 2 neurons (Figure 7A, insets). The network was constructed so that the activity of any given pair of neurons in Layer 2 was equivalent to that of the neuron pairs studied in previous sections. As the correlation of layer 2 spike outputs was shaped, so too was the magnitude and timescale of the synaptic drive to the downstream target neuron (Figure 7B). For comparison, we show that downstream target's synaptic input when the layer 2 neurons were uncorrelated (Figure 7B, bottom), showing significantly reduced variability [12]. In the uncorrelated case, the firing rate of the downstream target was much less than 1 Hz, indicating that correlated input was necessary for its recruitment.
We study how correlation shaping of the layer 2 projections affected the recruitment of the downstream target neuron. In particular, we focused on how the changing timescale of correlation recruited downstream targets differentially, depending on their own integration properties. We varied the rate of balanced synaptic input from layer 1 to layer 2 in a smooth manner (following the and path for 15 Hz output in Figure 5A), gradually shaping the correlation function between any given layer 2 neuron pair. The shaping included the low and high states described earlier as near endpoints on a continuum (Figure 7C). When the downstream target had a smaller time constant (3 ms), its firing rate was increased when the pre-synaptic population was in the high state (Figure 7C, dashed line). This contrasted with the decreased firing rate in the high state when the downstream target had a longer time constant (20 ms) (Figure 7C, solid line). This differential effect was due to matching between the correlation timescale of layer 2 and the integration timescale of the downstream target. In the high state, synchrony drove the neuron with the short integration timescale, while, in the low state, long timescale correlations drove the slower neuron. Note that the firing rate of layer 2 neurons was unchanged in all cases studied. This simple example demonstrates that the structure of correlations between pre-synaptic neuron pairs can differentially drive downstream targets depending on their integration properties.
We have demonstrated that the rate of balanced synaptic input changes the correlation timescale of spike trains of a pair of neurons receiving partially correlated input. High rate synaptic input promoted precise spike time synchrony, while low rate synaptic input enhanced long timescale correlation. This correlation shaping was independent of changes in input correlation or the output firing rate of the neuron pair. Rather, it required a thresholding nonlinearity between input and spike train response as well as a state-dependent integration timescale. Both of these are properties of many neurons in the central nervous system, and hence we expect that similar correlation shaping may occur in a variety of brain regions.
Correlated neural activity continues to receive increasing attention [1], prompting investigations of the mechanisms that determine the transfer of correlation. Correlations are typically measured only at one timescale, but as we have shown, the magnitude of correlation depends on the timescale being considered, as does the likely significance of this correlation for activation of downstream neurons. Past studies have highlighted the dependence of spike train correlations on the magnitude of input correlation [15], [16], the form of spike excitability [19], [66], or the firing rate of the neuron pair [17], [18]. However, how the timescale of correlations are modulated through plausible mechanisms had not been addressed. Changes in membrane conductance have been widely studied and strongly influence the dynamics of single neuron activity [22]. In our study, we found that timescale-specific changes in neural correlations are a necessary consequence of conductance based modulation schemes. Previous work that has examined how correlated activity is transferred has used linear response methods to examine the response of neurons to current fluctuations, thereby leaving membrane integration invariant [17], [18]. As a result, cellular properties such as timescale were not modulated (see Figure 4D). We showed that when synaptic conductance is considered, it is possible to shape both the magnitude and timescale of output spike train correlations. This is a novel result that is nevertheless consistent with, and complementary to, the observation that firing rate also modulates correlations (see Figure 5).
The widespread use of multi-unit recording techniques to study population activity has produced an increasingly clear picture of how neuronal spike trains are correlated in a variety of neural states. Recently, there has been particular interest in noise correlations, which are specific to within trial comparisons and cannot be directly attributed to a common signal [10].
Several groups have reported noise correlation measurements, ranging from small positive values [58], [2], [3], [6], [7], [8] to values that are, on average, zero, with positive and negative values equally represented [4], [67], [48]. Furthermore, in cases where significant noise correlation is measured, it can be modulated on distinct timescales. In the visual system, for example, noise correlation measured on timescales less than 100 ms is largest for cells with similar preferred stimulus orientations being driven at that orientation, observed in both spike responses [2] and synaptic input [37]. Further, while increasing stimulus contrast enhances short timescale correlation, it reduces long timescale ( ms) correlation [2]. In primate area V4, stimulus attention reduces noise correlation when measured on timescales that are larger than 100 ms, yet has little influence on short timescale correlation [7], [8]. In contrast, other groups have shown that stimulus attention enhances spike synchrony measured at the gamma frequency timescale (20–40 ms) [68]. In the electrosensory system, long timescale noise correlation is reduced by recruitment of a non-classical receptive field, while synchrony is increased under the same conditions [3]. Thus, spike train noise correlations provide an excellent framework to study how the magnitude and timescale of correlations are shaped by neural state changes.
While a shaping of output correlation observed in these systems may be inherited from a state-dependence of input correlation (Figure 1B), single neuron response properties are often also modulated by network state. This suggests that a shift in correlation susceptibility may underlie a shift in pairwise correlation (as in Figure 1C). Indeed, firing rate gain is modulated by attention [69], stimulus contrast [21], and the recruitment of a non-classical receptive field [70]. In many cases, intracellular recordings have established that gain control is mediated by an increase in the rate of excitatory and inhibitory synaptic inputs [21], [31], in a fashion similar to the case presented in our study. Dual intracellular experiments that measure both input and output correlation across distinct neural states [45], [46], [37] are required to parcel the contribution of correlation inheritance and correlation transfer to the full shift in noise correlations.
A central result of our paper is that changes in synaptic input rate shape the correlation between the output spike trains from a pair of neurons. This is a consequence of how synaptic input modulates the timescale of membrane integration and response sensitivity of the two neurons. Our theoretical analysis formalizes this concept by explicitly relating the spike train correlation coefficient to the single neuron transfer function. Though we focused on modulation by balanced synaptic inputs, the relationship between transfer function and correlation is general, requiring only that the input correlation be sufficiently small. Thus, we predict that any synaptic or cellular mechanism that modulates single neuron transfer will necessarily affect spike train correlations.
Modulation of single neuron transfer with the level of synaptic input rate is well studied [27], [28], [29], [30], [31], [32]. However, how other cellular processes affect neuronal transfer is equally well studied. For example, increases in the spike after-hyperpolarization [50] or decreases in the spike after-depolarization [51] reduce the gain of the firing rate response to static driving inputs. Sustained firing often recruits slowly activating adaptation currents that also reduce gain [52], [53]. We predict that these modulations will reduce long timescale spike rate correlations. In contrast, the presence of low threshold potassium currents in the auditory brainstem [71] promotes high frequency single neuron transfer and thus may also promote pairwise synchronization. In total, our result gives a general theory that links the modulation of single neuron and network responses, thereby expanding the applicability of studies of single neuron modulation.
How the brain selectively propagates signals is a basic question in systems neuroscience. One control mechanism is through an ‘unbalancing’ of feedforward excitation to inhibition, with disinhibited populations propagating activity and excessive inhibition silencing propagation [72]. Modulation of correlation is an alternative mechanism to control signal propagation. The correlation between spike trains from neurons in a population enhances the ability of that population's activity to drive downstream targets [12], [13]. We have shown that modulating the timescale of correlation in the upstream population to match the integration timescale of the downstream population improves signal propagation (Figure 7). Matching the integration dynamics of distinct neuronal populations to one another is a common theme in the binding of distributed activity [73]. In previous studies, the phase relationship between distinct neuronal populations both oscillating at some frequency gated the interaction between distinct brain regions. Our study did not assume rhythmic population dynamics, but rather only matched integration timescales.
The nonlinearity of spike generation allows for the transfer of shared input to multiple neurons to be controlled in complex ways. We have shown that well-studied mechanisms of single neuron response modulation, such as firing rate gain control, have direct relations to changes in correlation for neuron pairs. Thus, state dependent shifts in single neuron transfer also influence how populations of neurons coordinate their activity. Our results are a step in understanding how the collective behavior of neuronal networks can be controlled in different brain states.
|
10.1371/journal.ppat.1005205 | Transcription Factor SomA Is Required for Adhesion, Development and Virulence of the Human Pathogen Aspergillus fumigatus | The transcription factor Flo8/Som1 controls filamentous growth in Saccharomyces cerevisiae and virulence in the plant pathogen Magnaporthe oryzae. Flo8/Som1 includes a characteristic N-terminal LUG/LUH-Flo8-single-stranded DNA binding (LUFS) domain and is activated by the cAMP dependent protein kinase A signaling pathway. Heterologous SomA from Aspergillus fumigatus rescued in yeast flo8 mutant strains several phenotypes including adhesion or flocculation in haploids and pseudohyphal growth in diploids, respectively. A. fumigatus SomA acts similarly to yeast Flo8 on the promoter of FLO11 fused with reporter gene (LacZ) in S. cerevisiae. FLO11 expression in yeast requires an activator complex including Flo8 and Mfg1. Furthermore, SomA physically interacts with PtaB, which is related to yeast Mfg1. Loss of the somA gene in A. fumigatus resulted in a slow growth phenotype and a block in asexual development. Only aerial hyphae without further differentiation could be formed. The deletion phenotype was verified by a conditional expression of somA using the inducible Tet-on system. A adherence assay with the conditional somA expression strain indicated that SomA is required for biofilm formation. A ptaB deletion strain showed a similar phenotype supporting that the SomA/PtaB complex controls A. fumigatus biofilm formation. Transcriptional analysis showed that SomA regulates expression of genes for several transcription factors which control conidiation or adhesion of A. fumigatus. Infection assays with fertilized chicken eggs as well as with mice revealed that SomA is required for pathogenicity. These data corroborate a complex control function of SomA acting as a central factor of the transcriptional network, which connects adhesion, spore formation and virulence in the opportunistic human pathogen A. fumigatus.
| Invasive fungal infections affecting immunocompromised patients are emerging worldwide. Among various human fungal pathogens, Aspergillus fumigatus is one of the most common molds causing severe invasive aspergillosis in immunocompromised patients. The conidia, which can evade from innate immunity and adhere to epithelial cells of alveoli in human lungs will start to germinate and cause the disease. Currently, the understanding of the molecular mechanisms of adherence of fungal cells to hosts is scarce. The transcription factor Flo8 controls adhesion to biotic or abiotic surfaces and morphological development in baker’s yeast. Flo8 homologues in the dimorphic human pathogenic yeast Candida albicans or the filamentous plant pathogen Magnaporthe oryzae are required for development and virulence. We found in this study that the Flo8 homologue SomA of A. fumigatus is required for adhesion and conidiation. Two independent invasive aspergillosis assays using chicken eggs or mouse demonstrated that deletion of the corresponding gene resulted in attenuated virulence. SomA represents an important fungal transcription factor at the interface between adherence, asexual spore formation and pathogenicity in an important opportunistic human pathogen.
| Adherence to host cells represents a key step for pathogenesis of bacterial or fungal microorganisms. Prerequisites at the molecular level include cell wall adhesive or hydrophobic proteins, carbohydrate components of the cell wall and the extracellular matrix. Gene families responsible for adherence comprise the FLO adhesins (flocculins) of Saccharomyces cerevisiae or the ALS agglutinins (agglutinins like sequence) of Candida albicans [1, 2]. Conidial adherence of the opportunistic human pathogen Aspergillus fumigatus requires the hydrophobin RodA, the laminin-binding protein AspF2 or glycans as important constituents of the cell wall [3–5].
Adherence is triggered by different environmental stimuli which are sensed by receptors and which induce various signaling pathways [2, 6]. A prominent example is the cyclic adenosine monophosphate (cAMP) dependent signaling pathway, which is highly conserved from bacteria to mammals. In eukaryotic cells, the cAMP dependent protein kinase A (PKA) signaling pathway is activated by the G protein-coupled receptors [7]. The corresponding Gα subunit activates adenylate cyclases, which convert ATP to cAMP. This secondary messenger binds to the regulatory subunits of PKA. The catalytic subunits of the enzyme are released and activate downstream transcription factors by phosphorylation [7]. The cAMP/PKA pathway plays a crucial role in development and pathogenesis in animal or plant pathogenic fungi such as C. albicans, Cryptococcus neoformans, Magnaporthe oryzae and Ustilago maydis [8–12]. This link between development, virulence and the cAMP/PKA pathway is conserved in the filamentous fungus and opportunistic pathogen A. fumigatus. Components of the A. fumigatus cAMP/PKA pathway include the GpaA and GpaB Gα subunits of the heterotrimertic G protein, the AcyA adenylate cyclase, the PkaR regulatory and the PkaC1/PkaC2 catalytic subunits of PKA. Deletion of gpaB or acyA results in reduced conidiation and a reduced growth rate in the ΔacyA strain [13]. The regulatory PkaR and the catalytic PkaC1 proteins are required to promote germination, growth and accurate conidiation [14–16]. Null mutants of the previous described genes showed attenuated virulence and indicate a role of the cAMP/PKA pathway in the pathogenicity process. This process needs to be further elucidated because of the importance of the A. fumigatus opportunistic pathogen, which can cause invasive aspergillosis in immunocompromised individuals with mortality rates of more than 60% [17–19].
The cAMP/PKA pathway activates several downstream factors like the S. cerevisiae transcription factor Flo8, which controls adhesive and filamentous growth. It induces the expression of the FLO11 gene for an adhesin, which is required for flocculation or the establishment of biofilms in haploid and pseudohyphae formation in diploid yeast strains [2]. The Flo8 counterpart of the dimorphic yeast C. albicans regulates hyphal development and virulence factors [20]. In both yeasts, Flo8 interacts with additional co-activators as for example Mfg1, which is also required for invasive and hyphal growth [21, 22]. The current knowledge about the transcription factors and their corresponding genes which represent the homologues of the FLO8 gene are limited. Among filamentous fungi, only the gene for MoSom1 corresponding to Flo8 in the plant pathogenic filamentous fungus M. oryzae has been examined. The Flo8 counterpart of the dimorphic yeast C. albicans is the only analyzed protein in a human pathogen. Representatives of constitutively filamentous fungi of human pathogens have not yet been studied. MoSom1 carries like the corresponding yeast protein the N-terminal LUFS (LUG/LUH-Flo8-single-stranded DNA binding) domain and can complement adhesive growth in a Δflo8 yeast mutant strain. MoSom1 controls the gene for the hydrophobin MoMpg1, which is required for fungal attachment to plant leaves during infection. Deletion of the Mosom1 gene results in loss of asexual or sexual development and impairs pathogenicity [23].
The development of Aspergilli has been primarily analyzed in the model fungus Aspergillus nidulans [24, 25]. The C2H2 zinc finger transcription factor BrlA represents a central regulator of asexual development in A. nidulans as well as in A. fumigatus and controls the formation of vesicles, which are required for conidiation at the top of aerial hyphae. BrlA induces the expression of the downstream abaA and wetA regulatory genes, which induce differentiation of phialides as spore forming cells and the subsequent maturation of conidia, which represent the asexual spores [26]. The flbB, flbC and flbD regulatory genes are genetically located upstream the expression of brlA [25]. MedA and the APSES (Asm1, Phd1, Sok2, Efg1, and StuA) protein StuA regulate transcription of the brlA gene in A. nidulans where they are required for metulae cell formation from vesicles followed by phialide cell formation [27]. A. fumigatus forms only phialides as asexual spore forming cells but does not produce an additional layer of metulae cells. Though, lack of either MedA or StuA also impairs conidiation in A. fumigatus where their exact molecular function is yet unknown [28, 29]. Additionally, MedA and StuA control adhesion and virulence in A. fumigatus by regulating the gene uge3 encoding uridine diphosphate (UDP)-glucose-epimerase, which is essential for adherence through mediating the synthesis of galactosaminogalactan [30].
The main objective of this study was to examine the function of A. fumigatus SomA and putative interaction partners. SomA corresponds to the Flo8/Som1 regulator described in other fungi. Our data show that SomA in collaboration with its co-regulator PtaB plays a key role in a transcriptional network controlling conidiation and adhesion and that SomA is required for virulence of filamentous pathogens A. fumigatus.
Flo8 is a regulator of S. cerevisiae dimorphism and its counterpart Som1 in the filamentous fungus M. oryzae is required for plant pathogenicity in rice [20, 23, 31]. These proteins share with the A. fumigatus protein SomA (AFUA_7G02260) the LUFS domain, which contains a LisH (Lis homology) motif for protein dimerization and tetramerization at the N-terminus (Fig 1A). SomA shows identities of 15.7% and 20.5% to the Flo8 proteins of S. cerevisiae and C. albicans, respectively, and 39% to Som1 of M. oryzae. In addition, there is a conserved nuclear localization signal (NLS) PSPSKRPRLE in filamentous fungi. These data suggest that the proteins derived from all homologous genes have a nuclear function. Exons of somA were identified by comparing the DNA sequence of the genomic locus with cDNAs, which were amplified from total mRNA. Sequencing of the resulting plasmid revealed that somA carries five exons of a size of 486 bp, 152 bp, 1279 bp, 267 bp and 171 bp (pME4192) resulting in a deduced protein of 784 amino acids with a molecular weight of 84.59 kDa. An additional splice variant (pME4193) was found (Fig 1A) Both splice variants of somA could be identified in the ΔakuA strain after 20 h vegetative growth in liquid minimal medium (MM).
Cross-species complementation of the somA gene was performed in yeast flo8 mutant strains to verify whether both variants share similar functions with Flo8. Expression of either somA or its splice variant (pME4194 and pME4195) under the MET25 yeast promoter could rescue invasive growth (cell-surface adhesion) in the flo8 (truncated FLO8) haploid mutant (BY4742) on solid agar (Fig 1B). Flocculation (cell-cell adhesion) in liquid medium was complemented similar to Flo8 (pME4197) (Fig 1C). In addition, expression of somA in Δflo8 diploid strain (RH2660) restored pseudohyphal growth (Fig 1D). These data support that SomA and Flo8 can fulfill similar cellular functions in yeast.
Flo8 is a transcription factor, which binds and promotes transcription of the FLO11 gene encoding the flocculin Flo11 [32, 33], which is a key determinant for adhesion in yeast [2]. We performed β-galactosidase assays with the 3 kb FLO11 promoter fused to the bacterial LacZ reporter gene to examine whether SomA complements the adhesive phenotypes in flo8 or Δflo8 yeasts (Fig 1B and 1D) by activating FLO11 gene expression. As shown in Fig 2A, both SomA and its splice variant showed significantly increased FLO11 promoter driven LacZ activity in comparison to the mutant strain transformed with the empty plasmid.
We took a more detailed look at the FLO11 promoter to determine whether SomA and Flo8 bind to similar regions of the promoter. A set of 14 reporter constructs containing 400 bp FLO11 promoter fragments that overlap by 200 bp [34] (Fig 2B) was analyzed in the flo8/Δflo1 yeast strain (Y16870). As shown in Fig 2C, two promoter regions were affected by both Flo8 and SomA. Comparison of Fig 2B and 2C indicated that these two regions are located at 1.8 kb and 1.2 kb upstream of the start codon of FLO11. SomA presumably recognizes two additional regions located at 1.4 kb and 1 kb upstream of the FLO11 open reading frame. These data indicate that SomA and Flo8 share molecular functions in recognizing and controlling similar regions of the FLO11 promoter and hence complemented adhesion and filamentous growth in flo8 and Δflo8 yeast strains.
Yeast Flo8 is part of a protein complex required for regulating cellular development, and Mfg1 represents another subunit of this complex [21, 22]. We analyzed whether the similarity of SomA to Flo8 and the similar function of both proteins are reflected by similar interaction partners of SomA in A. fumigatus. A GFP tagged somA gene was constructed (AfGB75) to identify interaction partners of SomA. A GFP-Trap was performed and the recruited proteins were analyzed by LC/MS. Proteins identified in the GFP control strain were considered as unspecific background identifications (LC/MS raw data in S1 Table). Apart from SomA itself, the PtaB protein (AFUA_2G12910), a homologue of yeast Mfg1, was identified by LC/MS. This protein was absent in GFP control strain (Fig 3A). The detailed LC/MS data are shown in S2 Table.
We further performed a co-immunoprecipitation to verify whether PtaB and SomA are interaction partners (Fig 3B and 3C). A strain expressing SomA-GFP and PtaB-RFP fusion proteins was constructed (AfGB117). Application of the α-GFP antibody recognized GFP in the trap enrichment of the GFP control strain (Fig 3B). Several signals in the GFP-Trap of the strain expressing both fusions (somA-gfp/ptaB-rfp) presumably represent SomA-GFP and its degradation products. Only a single signal in the RFP-Trap was detected by the α-GFP antibody. The single signal (arrow, Fig 3B) at approximately 170 kDa was verified by LC/MS as SomA-GFP. This suggests that the SomA-GFP protein has been recruited through the RFP-Trap by PtaB-RFP. The reciprocal experiment using an α-RFP antibody and the same trap enrichments resulted in the recognition of RFP in the RFP control strain and several signals in the RFP-Trap presumably representing PtaB-RFP and its derivatives. Two signals at 120 kDa in the GFP trap enrichment (arrow, Fig 3C) were identified by the α-RFP antibody and were determined as the PtaB-RFP protein with a calculated size of 106 kDa by LC/MS. In addition, several SomA-GFP peptides were identified by LC/MS which presumably correspond to the SomA-GFP degradation products which are visible in Fig 3B (GFP-Trap lane). Taken together, these data suggest that SomA and PtaB physically interact in A. fumigatus similar to their counterparts Flo8 and Mfg1 in the two yeasts S. cerevisiae and C. albicans [22].
The somA gene was deleted in a ΔakuA background strain (AfS35) to analyze the function of this gene in correlation with growth, adhesion and development. Cultivation on solid MM plates revealed slow growth (2.8 mm/day) of the ΔsomA strain (AfGB77) in comparison to ΔakuA strain (6.1 mm/day). This ΔsomA growth phenotype was verified by complementation with the respective wild type gene. The complemented strain (AfGB105) showed improved growth rate (5.1 mm/day), indistinguishable from ΔakuA strain (Fig 4A). We also analyzed PtaB as the physical interaction partner by genetic analysis. The deletion of ptaB (AfGB115) resulted also in a reduced growth rate phenotype (4.6 mm/day) which was less pronounced in comparison to the ΔsomA phenotype. In addition, a delayed conidiation was observed in the ptaB null mutant (arrow, S1 Fig). The growth defect and the delayed asexual development of the ΔptaB strain were restored by complementation with a ptaB-rfp fusion (S1 Fig).
Asexual spores are normally produced at conidiophores consisting of aerial hyphae with a vesicle on top where the conidia are pinched off [27]. The ΔsomA strain formed exclusively aerial hyphae and was incapable of forming conidiophores. To have a detailed look on conidiation, the strains were inoculated on MM-agar coated microscope slides or MM agar on microscope slides and incubated for 28 h at 37°C. As shown in Fig 4B, the ΔsomA mutant showed no mature conidiophores. In contrast, the ΔakuA strain and the somA complemented strain revealed conidiophores (white arrow) and vesicle formation (black arrow) on top of the aerial hyphae. Furthermore, macroscopic inspection indicated that the ΔakuA strain produces aerial hyphae and conidiophores similar to the somA complemented strain. In contrast, the ΔsomA mutant showed only aerial hyphae (Fig 4B). The defect in vesicle formation of the ΔsomA mutant was similar to the defect in a ΔbrlA strain except of the growth retardation (Fig 4A) [35]. We analyzed the SomA dependent step in asexual development in more detail. A Tet-somA strain (AfGB74) was constructed by replacing the promoter region with the inducible Tet-On system [36] which could conditionally express the somA gene upon addition of doxycycline to the medium.
The growth and conidiation phenotype of the ΔakuA strain were not affected by the presence of doxycycline (Fig 4A). The Tet-somA strain grew as slowly as the ΔsomA mutant and had severely impaired sporulation when doxycycline was absent (Off state). In contrast, these impaired phenotypes were complemented when the promoter was induced by doxycycline (Fig 4A). Further observation showed that the Tet-somA strain revealed conidiophores (white arrow) and vesicle formation (black arrow) on top of the aerial hyphae as the ΔakuA strain only under inducing conditions (Fig 4B). Taken together, these results support a function of SomA in asexual development and fungal growth.
Flo8 is required for adherence of S. cerevisiae by regulating FLO gene expression [2]. Therefore, the impact of the loss of the somA gene on the adherence to plastic or fibronectin were examined. Due to the fact that the ΔsomA mutant has a defect in asexual development, we used the Tet-somA strain to perform the adherence assay and the ΔakuA strain was used as control. As a pilot test the adherence of germlings was tested. Germlings of the ΔakuA strain and Tet-somA mutant (On state) displayed 25% adherence to polystyrene plates and fibronectin-coated plates. In contrast, the Tet-somA germlings (Off state) showed only 5% adherence to both surfaces (Fig 5A).
The polysaccharide galactosaminogalactan (GAG) from the fungal cell wall is composed of α1,4-linked galactose and N-acetylgalactosamine and plays a role in fungal adherence [30]. We tested whether the loss of germling adhesion in the Tet-somA mutant (Off state) is due to reduced GAG production. We cultivated the strain under inducing and non-inducing conditions and after precipitation and hydrolysis of GAG, the amounts of galactose and galactosamine were measured by GC-MS [30]. The amount of galactose was reduced to 31% and the amount of galactosamine was reduced to 6% in the Tet-somA strain (Off) compared to the Tet-somA strain (On), respectively (Fig 5B).
The yeast Flo8-Mfg1 complex is required for biofilm formation [22]. We analyzed whether SomA and PtaB play a similar role in the A. fumigatus life style. As shown in Fig 5C, the hyphae of the Tet-somA strain (On) formed biofilm when the promoter was induced by doxycycline (+). The ΔakuA strain with (+) or without the drug showed similar biofilm formation (Fig 5C). In contrast, the complete mycelium was washed off when the Tet-somA strain was at Off state. A similar phenotype was detected for PtaB. The ΔptaB mutant strain resulted in a defect of biofilm formation and this phenotype could be rescued by re-introducing the ptaB-rfp fused gene in the deletion strain (Fig 5C).
Taken together, these data show a common function of SomA and PtaB in biofilm formation. Furthermore, SomA is required for germling adherence to plastic surfaces or fibronectin and GAG production.
The cellular function of SomA as a transcription factor involved in asexual development and adherence was examined by quantitative transcript analysis of putative target genes. We could show that the Tet-somA strain has a similar asexual development as the ΔsomA mutant when doxycycline is absent (Fig 4). In addition, we showed that SomA plays an important role in adherence and GAG production using the Tet-somA strain (Fig 5). Hence, we used the Tet-somA strain to test the role of SomA in gene regulation. The ΔakuA mutant and the Tet-somA strain were incubated in liquid minimal medium (MM) for 18 h. Afterwards, the mycelium was shifted to liquid MM for 8 h and solid MM plate for 24 h with or without doxycycline. The drug had no effect on gene expressions in the ΔakuA mutant (S2 Fig). Transcript analysis revealed that the Tet-somA strain (Off) abolished brlA expression in contrast to the On state of the Tet-somA strain (Fig 6). In A. fumigatus, FlbB is necessary for flbD expression and FlbD might be essential for expression of brlA [25]. The expression of flbBCD genes in the Tet-somA strain was decreased in the Off state in comparison to the On state (Fig 6). The velvet domain protein family and LaeA also control fungal development and secondary metabolism in filamentous fungi including conidiation [24, 37, 38]. The expression of members of the velvet domain family was not significantly affected except for transcription of the velC gene, which was impaired by the Tet-somA Off state (Fig 6). These results suggest a broader role of SomA in fungal development.
Gravelat et al (2013) showed that medA and stuA genes are required for adhesion and regulate some putative adhesins [28, 30] and the transcript levels of medA and stuA were significantly reduced in the Tet-somA Off state (Fig 6). Possible adherence genes located downstream of the medA and stuA genes were further evaluated. Three genes (AFUA_3G13110, AFUA_3G00880 and uge3) encoding possible adherence-associated proteins with high scores in bioinformatic prediction [39] were analyzed. The transcript levels of all three genes are reduced in the absence of somA (Off state) (Fig 6). A similar transcript analysis was also observed in the ΔsomA mutant in comparison to the ΔakuA background strain and the somA complemented strain (S3 Fig).
Deletion of the SomA interaction partner PtaB also resulted in a delayed conidiation (S1 Fig) and, a defect of biofilm formation (Fig 5C). This suggests that PtaB might also contribute to the SomA control of gene transcriptions. The transcript levels showed that the ΔptaB mutant strain had a significant effect on the expression of the development and adherence related genes which are also controlled by SomA (S4 Fig).
SomA, FlbB, MedA and StuA represent fungal transcription factors controlling a complex developmental regulatory transcriptional network. To identify the interaction between SomA and these three transcription factors, an epistatic analysis was performed. The overexpression of somA did not change the phenotype of either ΔakuA background, ΔflbB, ΔmedA or ΔstuA mutant strains and had also no significant effect on colony growth (Fig 7A). Double deletion strains revealed a different picture. An additional somA deletion in the ΔflbB background resulted no more in a ΔflbB but in a ΔsomA colony phenotype including the reduced growth rate of the colony. The same phenotype indistinguishable from the ΔsomA mutant was observed in the double mutant strain with ΔmedA (Fig 7B). The ΔstuA ΔsomA double mutant (AfGB114) showed a more complex phenotype which does not completely correspond to the ΔsomA deletion. More aerial hyphae on the surface and a lighter color on the back are visible compared to ΔsomA single or the other double deletion strains.
Taken together, our combined genetic and transcriptional analysis supports that SomA regulates asexual development regulatory genes flbB and flbD and, through this pathway, affects the brlA master gene of conidiation. There is presumably an additional combinatory effect between SomA and the regulator StuA. The network of SomA, PtaB, StuA and MedA finally results in a SomA-mediated control of various adhesins encoding genes in A. fumigatus.
SomA is presumably required for adhesion by affecting medA expression and a ΔmedA deletion results in reduced virulence in a mice model [28]. Hence, we addressed whether SomA plays a role in virulence in animals. We established the Tet-On system in an egg infection model as a pilot study to carry out the virulence experiments. This model mimics the pulmonary invasive aspergillosis model in mice by infecting the chorioallantoic membrane in eggs [40].
In an egg infection model, the ΔsomA mutant was not included due to the severely impaired conidiation. The eggs infected with the inactive Tet-somA strain without doxycycline (Off) had no significant difference in mortality of infected eggs compared to the PBS control (p = 0.58; log-rank test). The Tet-somA (Off) showed attenuated virulence compared to the ΔakuA strain, the somA complemented and the Tet-somA (On) (p<0.05). In contrast, the On state of the Tet-somA strain which was induced by doxycycline showed similar virulence to the ΔakuA strain or the somA complemented strain (Fig 8A).
We verified the egg infection experiments by virulence assays of the mutant compared to the ΔakuA background strain in a mouse infection model for pulmonary aspergillosis [41]. As in the egg model, the ΔsomA mutant was not included due to the severely impaired conidiation. Addition of doxycycline in mouse model was not performed due to the fact that the somA complemented strain and the Tet-somA (On) strain show the virulence in the egg infection (Fig 8A). In this model, the Tet-somA mutant (Off) displayed attenuation in virulence, which was statistically significant compared to the ΔakuA background and the somA complemented strain (p<0.05) (Fig 8B). In neutropenic mice, a cellular immune response is severely restricted and development of invasive aspergillosis is characterized by massive invasive growth of the fungus. Accordingly, the presence of invasive mycelia was confirmed by histopathology in mice infected with the ΔakuA background and the complemented conidia. Even the two mice that died after infection with Tet-somA (Off) showed fungal growth within the lung tissue. However, no mycelium could be found in the majority of mice that survived the infection with Tet-somA conidia (Off) (Fig 8C). We had shown that the ΔptaB mutant had a defect of biofilm formation (Fig 5C) and PtaB interacts with SomA (Fig 3). This suggests that PtaB might also contribute to the SomA control of pathogenesis. However, a mice infection model showed that the ΔptaB mutant strain had normal virulence and invasive mycelia as the ΔakuA background strain (Fig 8B and 8C).
The resistance of oxidative stress and cell wall integrity are important virulent factors in A. fumigatus infection [42, 43]. Due to the fact that loss of somA was avirulent whereas the ΔptaB strain had normal virulence (Fig 8B), we performed a stress test to study the function of these two genes in stress response. As shown in S5 Fig, the ptaB null strain was resistant to H2O2 (3 mM) in comparison to the ΔakuA background strain. In contrast to this result the loss of somA does not increase resistance to H2O2.
Taken together, these data show that the Tet-On system is functional in the egg infection assay of A. fumigatus. The egg as well as the mouse model as established infection assays support that SomA is contributing to virulence of the opportunistic fungal pathogen A. fumigatus.
The current understanding of Flo8/Som1 homologues and their role in adhesion and virulence is primarily based on yeasts with their dimorphic life style switching between a single cell yeast growth form and a pseudohyphal or hyphal growth mode [20, 31]. In addition, Som1 had been analyzed in plant pathogenic and saprophytic filamentous fungi [23, 44]. Here, we show that similar to the Flo8-Mfg1 complex in yeast the corresponding pair SomA-PtaB is required for biofilm formation in the opportunistic human pathogenic filamentous fungus A. fumigatus. Application of the Tet-On system revealed that the Flo8/Som1 counterpart SomA of this fungus functions in development and virulence in embryonated hen egg as well as a mouse infection model.
The mechanism of adherence has most extensively been studied in the yeast S. cerevisiae. Since adhesion is highly correlated with pathogenicity in fungi, S. cerevisiae has been used as a model for detecting adhesins encoding genes and identifying control genes of adhesion from C. albicans [45] and the filamentous fungus Verticillium longisporum [46]. Flo8 is one of the most prominent yeast regulators of adhesion and it had been demonstrated that Flo8 functions downstream of the cAMP/PKA pathway [47]. The binding of Flo8 to target promoters is regulated in budding yeast by Tpk2, which is one of the catalytic subunits of PKA, and loss of either Flo8 or Tpk2 blocks pseudohyphal growth [48]. We showed that heterologous SomA protein complements the defects of flo8 yeast in haploid adhesive (cell-cell and cell-surface) and diploid pseudohyphal filamentous growth, which require the expression of FLO11 or FLO1 [2, 32]. Furthermore, the expression of FLO11 could be activated by SomA in flo8 yeast. These results suggest that SomA might be activated by Tpk2 in S. cerevisiae and it can activate downstream genes as FLO11.
Higher expression levels of FLO11::LacZ reporter were measured by heterologous SomA compared to the Flo8. Further analysis of FLO11 promoter regions indicated that SomA binds to two similar promoter regions as Flo8 and two additional regions. All these four regions contain the Flo8 consensus binding sequence TTTGC [49]. The higher promoter binding activity of SomA might be due to a higher overall binding affinity to the FLO11 promoter compared to Flo8. Flocculation requires both adhesins Flo1 as well as Flo11 [33, 50]. The observed decreased flocculation by SomA suggests a poor binding activity to the promoter of FLO1. The fact that heterologous SomA activates the expression of FLO11 by binding to similar regions of the promoter as Flo8 supports an evolutionary conserved strategy of gene activation in yeast and filamentous fungi.
Flo8/Som1 is a transcription factor that regulates downstream targets together with other interaction partners [22, 23]. Here, we showed that SomA interacts with PtaB, the A. fumigatus homologue of yeast Mfg1. PtaB and Mfg1 proteins are members of the LIM-domain binding protein family, which play a role in development in eukaryotic cells [51, 52]. In S. cerevisiae, the Mfg1 protein forms even a ternary complex with Mss11 and Flo8 leading to efficient FLO11 expression and hence mediating invasive growth and pseudohyphal formation [22]. Deletion of either partner gene leads to the loss of these growth forms. In C. albicans, this heterotrimeric complex is also required for hyphal growth [21], but overexpression of FLO8 complemented the defect of Δmss11 in hyphal growth [53]. This indicated that Flo8 and Mss11 might share functions in development. The closest relative of an Mss11 encoding gene in A. fumigatus is somA with no other putative paralogue present in the genome. One possible explanation is that the yeast Flo8-Mss11-Mfg1 complex might correspond to a SomA-SomA-PtaB complex in A. fumigatus. The MSS11 gene in yeasts might be either a product of gene duplication in yeasts or might have been lost in filamentous fungi as the Aspergilli.
Increased protein sizes of SomA-GFP and PtaB-RFP were detected with the respective antibodies in Western experiments (Fig 3B and 3C). These results imply that SomA-GFP and PtaB-RFP is modified post-translationally. Bioinformatic prediction tools [54] suggest 10 putative ubiquitination sites (medium confidence) in SomA and three additional sites in the deduced PtaB primary amino acid sequence. Consistently, SomA-GFP is simultaneously recognized with α-Ubi and α-GFP antibodies (S6 Fig).
Asexual development in Aspergilli is a morphological change, which is reminiscent to the dimorphic life style of yeasts. Aerial hyphae are formed which can differentiate into conidiophores, containing many single cell conidia with a single nucleus per cell. These asexual spores are released into the air for dispersal of the fungus [25, 27]. The combined data of the deletion analysis and the regulated promoter suggest that SomA is a regulator of asexual spore formation and is controlling developmental steps after the formation of aerial hyphae during the asexual cycle (Fig 4B). The similar developmental phenotypes in vesicle formation between the ΔsomA (AfGB77) and the ΔbrlA (A1176) strain and the fact that SomA regulates the expression of flbB, flbD and brlA genes suggest that SomA and BrlA might be part of the same regulatory pathway. This is consistent with earlier findings where lack of conidiation was also observed in M. oryzae and A. nidulans when Som1 is inactivated [23, 44].
The connection between Flo8/Som1 and the PKA pathway seems to be conserved between yeasts and filamentous fungi. The ΔsomA strain in A. fumigatus was reduced in its growth rate in comparison to the wild type and resembles the ΔacyA mutant phenotype, which is deficient in the adenylate cyclase producing cAMP where growth was reduced and nearly no conidiation was observed [13]. MoSom1 interacts with the CpkA catalytic subunit of protein kinase A [23]. A. fumigatus pkaC1 and pkaC2 encode two cAMP dependent PKA catalytic subunits. PkaC1 belongs to the class I PKAs similar to Tpk proteins of S. cerevisiae whereas PkaC2 is dispensable for conidiation. In contrast, PkaC1 is responsible for conidiation and vegetative growth [15]. This suggests that the protein kinase PkaC1 may regulate activation of SomA and subsequently control asexual development.
SomA controls conidiation and adhesion primarily by affecting the expression of the three regulatory genes flbB, stuA and medA (Fig 9). FlbB is a bZIP transcription factor which controls together with the cMyb transcription factor FlbD the expression of the major regulatory gene brlA. The resulting protein BrlA is a C2H2 zinc finger transcription factor, which plays a key role in asexual development in the pathogen A. fumigatus and the model fungus A. nidulans [35, 55]. Deletion of either flbB or flbD results in fluffy phenotypes resembling the ΔbrlA mutant strain in A. nidulans. The FlbB impact on conidiation is similar in A. fumigatus, but the FlbD impact is less pronounced. The flbB deletion abolishes flbD expression and delays brlA expression [56]. In A. fumigatus, expression of the flbD gene requires in addition to FlbB also FlbE as further developmental regulator. Consequently, conidiation is delayed and reduced in a ΔflbB mutant [25, 56–58].
SomA controls stuA and medA expressions in the correlation between conidiation and adherence in A. fumigatus. StuA and MedA contribute to accurate spatial and temporal expression of brlA in A. nidulans [59]. Consistently, disruption of stuA and medA results in abnormal conidiophores and reduced conidiation in A. fumigatus [27–29]. The severe impairment of conidiation in the somA deletion mutant can be attributed to both StuA and MedA, which are also required for activating expression of genes for adherence. Uge3 is an UDP-glucose epimerase that interconverts UDP-glucose and UDP-galactose and mediates formation of galactosaminogalactan (GAG) [30]. This compound is part of the extracellular matrix and is required for biofilm formation as well as adherence and therefore plays a prominent role in pathogenesis of A. fumigatus [60]. The StuA binding sites (A/T)CGCG(T/A)N(A/C) had been defined [61] and are also present in the promoter regions, of the uge3 (position -1651 and -1108 bp) or AFUA_3G00880 (position -3627 bp) genes for adhesion and as well in the brlA promoter region (position -507, -753 and -3276 bp) for asexual development (S7 Fig). This indicates that StuA has a dual role in directly activating the transcription of genes for adhesion and conidiation by binding to the corresponding promoters (Fig 9). SomA is required for the expression of stuA and medA and thereby plays an important role for conidiation as well as adhesion.
In this study, an attenuated virulence of the Tet-somA mutant (Off state) resulting in significantly reduced cellular SomA protein levels was observed in infection models with embryonated egg or with mice (Fig 8). An interesting question is which of the phenotypes which are observed in the absence of sufficient amounts of SomA protein is causing the reduction in virulence. SomA is required for the normal fungal growth rate, for asexual spore formation and for adherence. The ΔstuA mutant which is located downstream of the somA gene (Fig 9) showed abnormal impaired conidiation with a nonsignificant reduced virulence [29]. Loss of the gpaB gene, which is the Gα subunit of a heterotrimeric G proteins upstream of somA resulted in no growth retardation but attenuated virulence in a mouse model [16]. Loss of somA gene resulted in abolished biofilm formation and adherence to various surfaces as well as decreased uge3 expression and GAG production (Figs 5 and 6). Gravelat et al. showed that adhesion is an important factor for full virulence in mice model. The deletion of uge3 resulted in a normal growth and morphology together with an impaired adherence and reduced virulence in mice model [30]. In contrast, the ΔptaB mutant showed normal virulence in mice model incorporate with reduced expression of uge3 gene and resistance to oxidative stress (Fig 5C and S4 and S5 Figs). These data support a complex interplay of SomA with several genetic networks which ultimately has an important impact on fungal virulence in host cells (Fig 9).
The Flo8/Som1 protein family functions downstream of the cAMP/PKA signaling pathway and this pathway regulates asexual development in both A. nidulans and A. fumigatus [25]. We showed that SomA is upstream of flbB, medA and stuA genes (Fig 7). Overexpression of the somA gene in the ΔstuA mutant had no significant phenotype compared to the ΔstuA mutant. In contrast, the ΔstuA ΔsomA double mutant strain resulted in a complex phenotype with more aerial hyphae and different lighter color in the back as a hint of changes in secondary metabolite production. This distinct phenotype of the ΔstuA ΔsomA double deletion strain could reflect the activation of additional signal pathways. Previous studies indicated that StuA is required for conidiation and the regulation of secondary metabolite clusters [29, 35]. These studies suggested that the difference of color on the back might be due to the absence of StuA protein in the double deletion mutant. The stuA gene was expressed, but was reduced, in the ΔsomA mutant (Fig 6 and S3 Fig). Macheleidt et al. showed that expression of stuA is regulated by the cAMP/PKA pathway [62]. SomA might affect the expression of the stuA gene but the interplay between both gene products is presumably more complex and might rather resemble a genetic network than a single signal transduction pathway (Fig 9).
There is an important interplay between conidiation and cell-cell adhesion. The formation of aerial hyphae results in vesicles, which further differentiate into spore forming cells by a polar budding process reminiscent to the yeast single cell growth mode [27, 63]. In A. nidulans, this results primarily in the formation of metulae cells, which are absent in A. fumigatus where directly the spore forming phialides are formed. Phialides are the cells, which produce the small hydrophobic non-motile conidia in a process similar to pseudohyphae formation where at the beginning the spores are attached to each other. Pseudohyphae formation in yeast requires adhesins such as Flo11, which mediate cell-cell adhesion [2]. StuA controls metulae and phialide differentiation in A. nidulans and consistently, Phd1 as homologue of StuA, governs pseudohyphal growth in S. cerevisiae [63]. Asexual spore formation also requires AbaA, which is located downstream of BrlA in the developmental cascade. Consistently, the AbaA corresponding yeast protein Tec1 is required for pseudohyphae formation. Tec1 can be replaced by A. nidulans AbaA to repair the defect of a Δtec1 S. cerevisiae mutant strain [64].
Several proteins providing adherence in filamentous fungi have been identified. SomA affects several adhesion related genes (Fig 6), which results in the SomA mediated plastic adherence observed in this study. Hydrophobins also play a specific role in adhesion and are amphiphilic proteins involved in aerial hyphae formation and conidiation [65]. Hydrophobins as Mhp1 of the plant pathogen Magnaporthe grisea or Mpg1 of M. oryzae are responsible for appressorium development and subsequent entry into the plant host [66, 67]. RodA is a spore hydrophobin of A. fumigatus which prevents immune recognition and provides adherence of conidia to collagen or albumin [3]. The expression of the rodA gene depends on regulators as BrlA and AbaA [25] and the rodA expression is affected by SomA (Fig 6). Recently, galactosaminogalactan was found to be an adhesive compound produced by the A. fumigatus Uge3 epimerase, which is required for virulence [30]. We showed that SomA and PtaB regulate uge3 expression as well as stuA and medA (Fig 6 and S4 Fig), which have been shown to control transcription of uge3. In addition, loss of somA showed reduced galactose and galactosamine production which are required for galactosaminogalactan formation (Fig 5B). Interestingly, the homologues of these genes are also present in the non-pathogenic A. nidulans. StuA, MedA and Som1 are required for normal conidiation in A. nidulans [27, 44] and StuA binding sequence is present in the promoter of the uge3 homologue. This indicated that SomA might regulate other unknown genes required for adhesion or pathogenicity.
Most of the filamentous pathogenic fungi show host specificity and can only infect and cause disease in either plants or animals. Only a limited number of fungal pathogens such as Aspergillus flavus or Fusarium oxysporum can cause infections in both kingdoms. Plants and animals possess distinct protective systems to prevent fungal infections. This includes the plant cuticle as barrier or increased temperatures of 37°C and higher in humans [68]. Genes which are required for cuticle degradation or appressorium formation are specific for plant pathogens [69]. The set of genes from filamentous fungal pathogens, which is important for plant as well as human pathogenic fungi is limited and includes siderophores for the uptake of iron ions and adhesins as hydrophobins [70, 71]. The transcription factor Flo8/Som1 is an interesting control gene which is required for adherence, development and pathogenicity of pathogens which include filamentous fungal plant pathogens as M. oryzae as well as dimorphic and constitutively filamentous fungal human pathogens as C. albicans and A. fumigatus, respectively.
The fungal strains used in this study are listed in Table 1. The AfS35 (ΔakuA) was used as the background strain [72]. A. fumigatus was grown at 37°C in minimal medium (MM) as previously described [73, 74]. 1% D-glucose as carbon source and 10 mM ammonium tartrate as nitrogen source were supplemented. 2% agar and 1 mg/L pyrithiamine were used for solid medium and selection, respectively. For pyrithiamine marker recycling, 0.5% xylose was supplemented. Escherichia coli strain DH5α was used for construction of plasmid and was propagated in LB medium (0.5% yeast extract, 1% bacto-tryptone and 1% NaCl) at 37°C. S. cerevisiae strains BY4742, Y16870, RH2656 and RH2660 (Table 1) were used for cross-species complementation. The BY4742 (flo8) is derived from S288c carrying truncated FLO8 gene [31] and RH2660 (Δflo8) is derived from Σ1278b with deletion of FLO8 gene [75, 76]. S. cerevisiae was cultivated at 30°C in either non-selective YEPD medium (1% yeast extract, 2% peptone and 2% glucose) or in SC-3 medium (0.15% yeast nitrogen base without amino acid and ammonium sulfate, 0.5% (NH4)2SO4, 2% glucose and 0.2% amino acid mixture lacking uracil, L-methionine and L-leucine). Appropriate amino acids were supplemented as required for adhesive assay and flocculation assay. For solid medium, 2% agar was added.
To determine the growth rate of the strains, 500 conidia or a portion of mycelia of the strains were inoculated in the middle of MM plates for 5 days. The colony diameters were measured every day. Doxycycline (5 mg/L) was supplemented as required.
Recombinant DNA technologies were performed according to standard methods [77]. DNA fragments for plasmid construction, hybridization probes or sequencing were amplified by Phusion polymerase (Thermo Fisher Scientific GmbH). Primers used for plasmid construction are listed in S3 Table. Isolation of genomic DNA from A. fumigatus was performed as previously described [78]. Isolation of plasmid DNA and RNA were performed using either QIAprep Miniprep Kit (Qiagen) or RNeasy Plant Mini Kit (Qiagen) referring to user’s manual, respectively. The cDNA of A. fumigatus was generated from total RNA using the QuantiTect Reverse Transcription Kit (Qiagen) following the user’s manual.
The 5’ and 3’ UTR regions were amplified with the corresponding primers HO499/500 or HO501/502. These two PCR products were fused by amplifying with the primer pair HO499/502 to yield a fragment which contains a restriction site for SfiI in the middle and a restriction site for HindIII at both ends. Then it was cloned into pJET1.2 Blunt cloning vector (Fermentas GmbH). The self excising marker system, which harbors a xylose driven β-recombinase, a pyrithiamine resistance cassette and two flanking binding sites (six), was isolated from pSK485 [79] with SfiI restriction enzyme. This recyclable marker fragment was cloned into the SfiI restriction site in the previous plasmid containing fused 5’ and 3’ UTR regions to generate pME4188.
Transformations were performed as previously described to construct A. fumigatus mutants [80]. The ΔakuA strain (AfS35) was transformed with the deletion fragment isolated from pME4188 using HindIII restriction enzyme and was selected with pyrithiamine. The positive mutant (ΔsomA::ptrA) was streaked out on MM plates containing 0.5% xylose and glucose to remove the pyrithiamine resistance and resulted in ΔsomA mutant (AfGB77).
For complementation, the fragment containing 5’ UTR and somA gene was amplified with the primer pair HO603/601 and cloned into SmaI digested pUC19 (Fermentas GmbH) using the In-fusion HD Cloning Kit (Takara BioEurope/Clontech) to generate pME4189. Linear pME4189 amplified with primers HO711/611 was fused with the recyclable marker fragment and the 3’ UTR of somA which was amplified with the primer pair HO677/501 to yield pME4190. The complement fragment isolated from pME4190 by HindIII digestion was transformed into ΔsomA mutant and the pyrithiamine resistance was removed to generate the complemented mutant (AfGB105).
To overcome the defect of conidiation in the somA deletion mutant, a strain containing the conditional expression of somA gene was generated. First, the pyrithiamine resistance cassette and the Tet-On system were amplified with the primer pair HO116/675 using pCH008 [36] as template. This fragment was cloned into the linear pME4189 amplified with primers HO710/676 to yield plasmid pME4191. The Tet-On system fragment replaced 602 bp of the 5’ UTR region (position -602~-1) of somA gene. The fragment isolated from pME4191 using HindIII restriction enzyme was transformed into the ΔakuA strain to generate the Tet-somA mutant (AfGB74).
The deletion of the ptaB gene was carried out as followed. The 5’ UTR was amplified via PCR using the primer pair PtaB-1/PtaB-2. For the 3’ UTR the primers PtaB-3/PtaB-4 were used. As selection marker the flipper cassette published by [79] was used. The cassette was received by digestion of the plasmid pSK485 with SfiI. The three received fragments were integrated in the pBluescript II KS+ via an EcoRV restriction site. Therefore, the GeneArt Seamless Cloning and Assembly Kit (Invitrogen) was used according to user’s manual. The generated plasmid pME4361 was digested with PmeI and the received knock out fragment was integrated in the ΔakuA strain. Transformants were selected on pyrithiamine containing media and checked for correctness via Southern hybridization. To recycle the resistance cassette clones were streaked out on minimal medium containing 0.5% glucose and 0.5% xylose. The loss of the resistance cassette was checked by Southern hybridization to generate the ptaB knock out strain without pyrithiamine resistance (AfGB115).
To complement the ptaB deletion mutant, a fragment containing the ptaB gene with 5’ and 3’ UTR region was amplified with primers HO885/701 and was cloned into pUC19. The rfp gene and the trpC terminator were amplified in accordance with the primer pair HO872/873 or HO874/890 and fused by amplifying with the primer pair HO872/890 to yield a 1.5 kb fragment. This fragment was cloned into the plasmid, which contains 5’UTR-ptaB-3’UTR in pUC19, amplified with primers HO887/888 to yield plasmid contains ptaB-rfp fused gene. Afterwards, the recyclable marker was cloned into the SfiI site in previous plasmid containing 5’UTR-ptaB-rfp-trpCt-3’UTR to generate pME4362. For complementation, the ΔptaB mutant was transformed with the fragment isolated from pME4362 with HindIII restriction enzyme to generate the complemented strain (AfGB116).
To construct plasmid pME4359, the 5’ and 3’ UTR regions of the medA gene were amplified with the corresponding primers HO852/881 or HO854/855 and fused by HO852/855. The fused fragment was further cloned into pUC19. The recyclable marker was cloned into the plasmid, which contains used 5’ and 3’ UTR of medA, amplified with primers HO854/881 to generate pME4359. The construction of plasmids pME4360 and pME4358 was similar to pME4359. The 5’ and 3’ UTR regions of stuA were amplified with primer pairs HO848/880 and HO850/851. These two PCR products were fused with primer HO848/851, cloned into pUC19 and linked with recyclable marker to yield pME4360. For pME4358, two UTR regions of flbB were amplified with primers HO844/879 and HO846/847 and fused by HO 844/847. The fused PCR was cloned into pUC19. The recyclable marker was cloned into the plasmid, which contains fused UTRs of flbB, amplified with primers HO882/846 to generate pME4358. The ΔakuA strain was transformed with the deletion fragment isolated from pME4359, pME4360 or pME4358 using HindIII and ApaLI restriction enzymes to obtain ΔmedA; ΔakuA (AfGB107), ΔstuA; ΔakuA (AfGB108) or ΔflbB; ΔakuA (AfGB106) mutants.
For construction of the somA overexpression plasmid, the somA gene was amplified with primer pair HO531/532 and cloned into PmeI site in pSK379 to yield pME4363. Plasmid pME4363 was transformed into the ΔakuA background, ΔmedA, ΔstuA or ΔflbB mutants to generate the OEsomA (AfGB119), ΔmedA;OEsomA (AfGB110), ΔstuA;OEsomA (AfGB111) or ΔflbB;OEsomA (AfGB109) strains, respectively. The somA gene is ectopical overexpression. Double deletion strains ΔmedA; ΔsomA (AfGB113), ΔstuA; ΔsomA (AfGB114) or ΔflbB; ΔsomA (AfGB112) were constructed by transformation of the deletion fragment isolated from pME4188 using HindIII restriction enzyme into the corresponding strain ΔmedA, ΔstuA or ΔflbB mutant.
All mutants were confirmed by Southern hybridization which was performed as described previously [81]. Preparation of probes was carried out using AlkPhos Direct Labelling Reagents Kit (GE Healthcare) according to user’s manual. Detection of probes was performed with the CDP-Star Detection reagent (GE Healthcare) following user’s manual.
Blast searches and protein conserved domain identification were conducted at the National Center for Biotechnology Information (http://www.ncbi.nlm.nih.gov). Protein alignments were made by Clustal Omega at European Molecular Biology Laboratory–European Bioinformatics Institute (http://www.ebi.ac.uk). Nuclear localization signal (NLS) was predicted at cNLS mapper (http://nls-mapper.iab.keio.ac.jp/). The protein name and gene number of A. fumigatus are according to the AspGD (http://www.aspergillusgenomes.org) [82]. Protein and DNA sequence analysis was performed using Lasergene software (Dna Star Inc., Madison, WI, USA).
To complement flo8-deficient yeasts, two SomA cDNA variants were amplified with primers HO441/HO442, digested with SpeI/HindIII and cloned into pME2786 and pME2787 to generate plasmids pME4192, pME4193, pME4194 and pME4195, respectively (Table 2). For positive control, S. cerevisiae Flo8 (FLO8) was amplified with primers HO446/447 and cloned into SmaI digested pME2786 and pME2787 to generate plasmids pME4196 and pME4197, respectively. Empty plasmids pME2786 and pME2787 were used as negative control. Transformation of S. cerevisiae was performed as described previously [83]. The pME2786 based plasmids were transformed into haploid strain BY4742 and selected on SC-Leu medium; meanwhile, the pME2787 based plasmids were transformed into haploid strain BY4742 and diploid strain RH2656 and RH2660 using selection medium SC-Ura.
For haploid adhesion assays, the transformed strains were grown on corresponding selective plates for 3 days at 30°C, and then those plates were washed with water until the negative control was washed away. The plates were photographed before and after washing. The flocculation assay in haploid yeasts was performed as previously described [33]. Briefly, the transformed BY4742 strains were inoculated in either SC-Leu or SC-Ura liquid medium and incubated for one day at 30°C. Following, 1 mL of EDTA (0.5 M, pH8) was added and the samples were vortexed until flocculation were disrupted. The value of flocculation was determined by F = 1-B/A, where A is OD600 in solution without 0.1% CaCl2 and B is OD600 in the presence of 0.1% CaCl2. For pseudohyphal growth, diploid transformants were grown on SLAD (synthetic low ammonium dextrose medium) for 6 days at 30°C, and then the plates were photographed.
To verify that SomA could act on the FLO11 promoter, both pME4192 and pME4193 (Table 2) were co-transformed with plasmid pME2167 which contains 3 kb of the FLO11 promoter fused with LacZ reporter gene, in Y16870 strain (flo8/Δflo1) and transformants were selected on SC-Ura-Leu medium. Transformations of plasmid pME2167 with either FLO8 (pME4196) or empty vector (pME2786) in Y16870 strain were used as positive or negative control, respectively. To identify the specific DNA motif that was regulated by SomA and Flo8, 14 reporter constructs containing 400 bp FLO11 promoter fragments which have 200 bp overlaps and were cloned in front of CYC1::LacZ fused gene [34]. These 14 reporter plasmids (pFLO11-2/1 to pFLO11-15/14) were co-transformed with plasmids pME4192, pME4193 or pME4196 in Y16870 and selected with SC-Ura-Leu medium. The regulation of SomA and Flo8 on the FLO11 promoter was determined by β-galactosidase assays.
The assays were performed as previously described [75]. Yeast strains were grown in SC-Ura-Leu liquid medium overnight as pre-culture, 1 mL of the pre-culture was added to 10 ml of SC-Ura-Leu-Met liquid medium as main-culture for 6 h. β-galactosidase activities were calculated by following formula (OD420 × 0.35) / (0.0045 × protein concentration × extract volume × time) [84]. Protein concentrations were determined by OD595 with Bradford assay [85].
To analyze mycelial morphology, strains were grown on agar-coated slides with thin layer of MM agar for 28 h at 37°C and then were observed under the Olympus Axiolab microscope with 400-fold magnification. For aerial hyphae visualization, strains were grown on agar slides with either MM agar or MM agar containing 5 mg/L doxycycline for 28 h at 37°C, and were observed under the microscope as well.
To construct SomA-GFP tagged protein, gfp gene was amplified with primers HO210/713 from pME4435. The 3’ UTR region of somA gene was amplified with primer set HO648/677. This fragment and gfp fragment were cloned into linear pME4189 containing 5’UTR-somA gene amplified with primers HO611/712 using In-fusion Kit resulting plasmid carrying somA-gfp fused gene together with 5’ and 3’ UTR regions. Then, the recyclable marker fragment was cloned into previous plasmid amplified with HO697/501 to yield pME4198. The ΔakuA background strain was transformed with somA-gfp fused fragment which was isolated from pME4198 using HindIII restriction enzyme and selected by pyrithiamine to yield somA-gfp strain (AfGB65). The ΔakuA strain was transformed with pME4435 containing overexression gfp for GFP-Trap control. To perform the co-immunoprecipitation experiments, the fragment isolated from pME4362 with HindIII restrict enzyme was transformed into the somA-gfp mutant and resulted in ptaB-rfp/somA-gfp strain (AfGB117).
A GFP overexpression strain and two independent SomA-GFP strains were grown in 200 mL of MM for 24 h at 37°C. Total proteins were extracted by mixing grinded mycelia with B* buffer (300 mM NaCl, 100mM Tris pH 7.0, 10% glycerol, 2 mM EDTA, 0.02% NP40, 2 mM DTT, 1 mM PMSF, 2 protease inhibitor pills/1 mL (Complete, EDTA-free, Roche)). The GFP trapping was performed by following steps: The crude protein extracts were mixed with 15 μL GFP-Trap beads (ChromoTek) which has been washed with B* buffer and incubated for 2 h on a rotating machine at 4°C. After the incubation, the beads were washed twice with 1.5 mL and 1 mL of B* buffer. After the centrifugation for 1 min at 4500 rpm at 4°C, the supernatant was removed. The beads were resuspended with 40 μL 6 X loading dye (250 mM Tris pH 6.8, 15% β-mercaptoethanol, 30% glycerol, 7% SDS, 0.3% bromophenol blue) and were boiled for 6–8 min at 95°C to separate the proteins from the beads. The eluted protein were applied to 12% SDS-PAGE. Gel pieces were isolated and performed trypsin digestion. Trypsin digestion was performed as previously described [86, 87].
The somA-gfp/ptaB-rfp strain (AfGB117), GFP (AfGB76) or RFP overexpression (AfGB118) strains were grown in 1 L of MM for 24 h at 37°C for co-immunoprecipitation experiment. Protein extraction and immunoprecipitation were performed as previous described. RFP-Trap beads (ChromoTek) was used for the RFP trapping.
20 μg of eluted protein from co-immunoprecipitations were subjected to 12% SDS-PAGE after heating in SDS loading dye at 95°C for 10 min and were transferred to a nitrocellulose membrane (Whatman). The enhanced chemiluminescence method was used for detection as previous described [88]. Signals were recorded with a Fusion-SL7 detection system (Peqlab) and the Bio 1D software (Peqlab). Detection of GFP or RFP fused proteins was carried out using α-GFP (Santa Cruz) or α-RFP antibody (ChromoTek), respectively. UbiA antibody (Genescript) was used for ubiquitin detection. The horseradish peroxidase-coupled rabbit (Invitrogen) or mouse antibody (Jackson ImmunoResearch) were used as secondary antibodies.
Mass spectrometry analysis was performed as previously described [87, 89]. Orbitrap raw files were analyzed with the Proteome Discoverer 1.4 software (Thermo Scientific, San Jose, Ca, USA) using the Mascot and SequestHT search engines with an A. fumigatus protein database with the following criteria: Peptide mass tolerance 10 ppm; MS/MS ion mass tolerance 0.8 Da, and up to two missed cleavages allowed. The variable modification considered was methionine oxidation, and carbamidomethylation was considered as fixed modification. High peptide confidence and a minimum of two peptides per protein were used as result filters. Heatmaps of MaxQuant results (version 1.4.1.2 v) were made with Perseus (version 1.4.1.3 v) software [90].
Two adherence assays were performed as previous described [25]. For biofilm formation 1 mL of Sabouraud broth containing 105 conidia of strains were inoculated into 6-well Nunclon Δ surface culture plates (Nunc) at 37°C for 24 h and 5 mg/L doxycycline was supplemented in the medium of indicated strains. Afterwards, the culture medium was removed and the wells were washed three times with PBS. Biofilm was visualized by applying 2 mL of 0.1% crystal violet solution for 5 min. Excess stain was removed and the plates were washed three time with 3 mL of sterile water. The remained stain in biofilm was extracted by adding 1 mL of ethanol. The biofilm density was determined by measuring the absorbance of the destained solution at 570 nm. The biofilm formation assays were performed in triplicate and experiments were repeated 3 times.
To test the adherence of mutants to plastic and fibronectin, 104 conidia in 1 mL of Sabouraud broth were incubated at 37°C for 8 h for germination and 5 mg/L doxycycline were supplemented in the medium of the indicated strains. The 6-well culture plates were untreated or coated with 0.01 mg/mL of fibronectin overnight and were inoculated with 150 germlings of the mutants and incubated at 37°C for 30 min. Afterwards, wells were washed 3 times with 3 mL of PBS and covered with YEPD agar. Fungal colony was quantified after incubation at 37°C. The adherence assay was performed in triplicates and experiments were repeated 3 times.
Quantitative real-time PCR (qRT-PCR) was performed with MESA GREEN qPCR MasterMix Plus for SYBR (Eurogentec) using CFX Connect Real-Time PCR system (Bio-Rad). The ΔakuA strain and mutants were grown in liquid MM medium for 20 h at 37°C and RNA was extracted using RNeasy Plant Mini Kit (Qiagen). DNase digestion and subsequent cDNA synthesis was carried out using 0.8 μg of RNA with the QuantiTect Reverse Transcription Kit (Qiagen). To perform the shift experiments, the indicated strains grown on liquied MM medium for 18 h at 37°C. Afterwards, the mycelium were shifted to liquid MM medium for 8 h or solid MM plate for 24 h. Addition of Doxycycline (5 mg/L) was noted. Each sample was performed in duplicates and the experiment was repeated two times. The histone H2A (3G05360) was used as reference gene for normalization. The relative expression of the gene of interest was calculated using the ΔΔCT method as previously described [91]. All the primers used for quantitative real-time PCR were listed in S3 Table.
The assay was performed as previous described with modification [30, 92]. 5*107 spores of the strain Tet-somA were inoculated in 100 ml modified Brian medium with 50 μg/ml doxycycline and grown for 20 hours at 37°C. The mycelium was shifted to 100 ml modified Brian medium with and without doxycycline. After 24 hours of growth at 37°C, the culture supernatants were precipitated with 2.5 volumes of ethanol. The pellets were washed with 150 mM NaCl and extracted with 8 M urea. The supernatants were dialyzed against water and dried. Polysaccharides were hydrolyzed (4 M TFA, 100°C, 4 h) and the obtained monosaccharides were derivatized with MSTFA and measured in Gas chromatography-mass spectrometry (GC-MS) (Varian CP-Sil8 CB for amines (30 m, 0.25 mm, 0.25 μm)).
Fresh harvest spores were counted and adjusted to 106 spores/ml. Approximately 2000 spores were spotted on minimal medium plates and incubated for three days at 37°C. For inducing cell wall stress conditions SDS was used at a final concentration of 0.01%. For oxidative stress H2O2 was used at a final concentration of 2 mM and 3 mM, respectively
Mice were cared for in accordance with the principles outlined by the European Convention for the Protection of Vertebrate Animals Used for Experimental and Other Scientific Purposes. According to that, chicken embryo is not animal (http://conventions.coe.int/Treaty/en/Treaties/Html/123.htm). All animal experiments were in compliance with the German animal protection law and were approved by the responsible Federal State authority “Thüringer Landesamt für Lebensmittelsicherheit und Verbraucherschutz” and ethics committee “Beratende Komission nach § 15 Abs. 1 Tierschutzgesetz” with the permit Reg.-Nr. 03-001/12. Fertilized chicken eggs were obtained from a local producer and stored at 4°C for maximum of 4 days before incubation. Embryonation was performed by incubation of the fertilized eggs in an incubator with 60% humidity at 37°C in the lab. The embryonated eggs were infected with spores on day 10 of embryonation and the infection experiments was terminated on day 17 of embryonation.
Egg infection model was performed as described previously [40]. Eggs were incubated in an incubator with 60% humidity for 10 days at 37°C. Each A. fumigatus strain was grown on malt extract agar (Oxoid) with 5 mg/L doxycycline for 7 days and the conidia were harvested freshly on the day for infection. Each egg was inoculated with 1000 conidia in 100 μL PBS with 5 mg/L doxycycline (corresponding to the volume of the egg) and 20 eggs were inoculated for one strain. Doxycycline was not added to the PBS solution in the Tet-somA (Off state) and the complemented strain. The embryonic death was determined by the loss of movement. Survival data were plotted by Kaplan–Meier curves and statistically analyzed by log rank test using Graph Pad Prism 5.00 (GraphPad Software).
Virulence of A. fumigatus mutant strains was tested in an established murine model for invasive pulmonary aspergillosis [41]. To induce neutropenia in female CD-1 mice (Charles River), cyclophosphamide (140 mg/kg; Sigma-Aldrich) was injected intraperitoneally on days -4, -1, 2, 5, 8 and 11, with an additional subcutaneous dose of cortisone acetate (200 mg/kg) on day -1. Mice were anaesthetized and intranasally infected with 20 μL of a fresh suspension containing 2 x 105 conidia. A control group was mock-infected with PBS.
The health status was monitored twice a day for 14 days and moribund animals, defined by severe dyspnoea, severe lethargy, or weight loss > 20%, were sacrificed. Infections were performed with a group of 10 mice for each tested strain. A control group of 5 mice was not infected (inhaled PBS). Survival data were plotted by Kaplan–Meier curves and statistically analyzed by log rank test and Gehan-Breslow-Wilcoxon test using Graph Pad Prism 5.00 (GraphPad Software).
Lungs from sacrificed animals were fixed in formalin and paraffin-embedded for histopathological analyses according to standard protocols. 4-μm sections were stained using Periodic acid-Schiff (PAS, hyphae stained pink). Sections were analyzed with the Zeiss Axio Imager.M2 microscope and images were taken with a Zeiss Axiocam 105 color camera and analyzed by Zen 2012 software (Zeiss).
S. cerevisiae: Flo8, P40068; Mfg1, Q07684; Mss11, Q03825; Flo11, P08640.
A. fumigatus: SomA, Q4WAR8; PtaB, Q4X0N1; BrlA, Q4WRE4; RodA, P41746; FlbB, Q4X053; FlbC, Q4X0E3; FlbD, Q4WK99; StuA, Q4X228; MedA, Q4X0J5; Uge3, Q4WX18; AfuA_3g13110, Q4WYI0; AfuA_3g00880, Q4WFV6; VelC, Q4WPM3.
|
10.1371/journal.pgen.1003004 | Cytokinesis-Based Constraints on Polarized Cell Growth in Fission Yeast | The rod-shaped fission yeast Schizosaccharomyces pombe, which undergoes cycles of monopolar-to-bipolar tip growth, is an attractive organism for studying cell-cycle regulation of polarity establishment. While previous research has described factors mediating this process from interphase cell tips, we found that division site signaling also impacts the re-establishment of bipolar cell growth in the ensuing cell cycle. Complete loss or targeted disruption of the non-essential cytokinesis protein Fic1 at the division site, but not at interphase cell tips, resulted in many cells failing to grow at new ends created by cell division. This appeared due to faulty disassembly and abnormal persistence of the cell division machinery at new ends of fic1Δ cells. Moreover, additional mutants defective in the final stages of cytokinesis exhibited analogous growth polarity defects, supporting that robust completion of cell division contributes to new end-growth competency. To test this model, we genetically manipulated S. pombe cells to undergo new end take-off immediately after cell division. Intriguingly, such cells elongated constitutively at new ends unless cytokinesis was perturbed. Thus, cell division imposes constraints that partially override positive controls on growth. We posit that such constraints facilitate invasive fungal growth, as cytokinesis mutants displaying bipolar growth defects formed numerous pseudohyphae. Collectively, these data highlight a role for previous cell cycles in defining a cell's capacity to polarize at specific sites, and they additionally provide insight into how a unicellular yeast can transition into a quasi-multicellular state.
| Many processes, including cell growth, are often regulated differently in distinct cellular regions. In the rod-shaped fission yeast Schizosaccharomyces pombe, new cell ends created by cell division initiate growth long after old cell ends inherited from mother cells. Though distributions of cell tip factors contribute to this growth pattern, we have found that the process of cytokinesis, which executes physical separation of daughter cells at the end of the cell cycle, also plays an important role in defining new end-growth competency. Defects in completing cytokinesis and remodeling the division site curb new end growth even when protein complexes that drive tip elongation constitutively associate with new cell ends. Moreover, when parts of the cytokinetic machinery persist at the division plane following constriction, S. pombe cells become highly invasive. We believe that these findings provide insight into growth transitions in pathogenic fungi, as well as into the evolution of the single-celled state from multicellular hyphal forms. Additionally, we speculate that cytokinesis-based constraints on growth polarity might be conserved in mammalian cells, which have been reported to likewise polarize only distally to the cleavage furrow at the conclusion of cell division.
| Many cells polarize in response to intrinsic and extrinsic signals. As cell polarization is generally multifaceted, cells must integrate both negative and positive cues for successful cellular morphogenesis. In various organisms, the cell cycle provides a platform on which these cues are organized (for reviews, see [1], [2]), thereby ensuring distinct polarization events occur at the appropriate location, time, and context.
The fission yeast Schizosaccharomyces pombe represents a genetically tractable organism for studying cell cycle regulation of growth polarity (for reviews, see [3], [4]). Wild-type S. pombe extend solely at their two cell tips, lengthening their rod-shaped bodies while retaining fairly constant widths. After cell division, S. pombe grow only at old ends, so-called because they served as ends of the dividing mother cell. Then, at a point in G2 known as new end take off (NETO), new ends, which arise from cell division, also initiate growth [5]. NETO is not required for cell viability, and myriad mutants defective in this process have been identified [3], [4]. Yet, beyond requirements for S-phase completion and a minimal interphase cell size [5], additional cell cycle controls on NETO have not been identified.
As in other cell polarization events, cytoskeletal rearrangements accompany growth transitions in S. pombe. Prior to NETO, microtubule plus end-associated proteins Tea1 and Tea4 ride growing microtubule ends to both cell tip cortices [6]–[9], where they anchor based on their association with membrane proteins [10], [11]. Upon NETO, Tea4 recruits formin For3, which had before only been tethered to old ends, into a complex with itself and Tea1 at new ends [8]. As over-expression of a Tea1-For3 fusion can drive NETO prematurely [8], this association likely brings For3 into the proximity of formin activators at new ends, stimulating For3 catalysis of F-actin cables that will deliver growth cargo to this tip. Not surprisingly, loss of Tea1, Tea4, and/or For3 impairs fission yeast polarization and elongation [8], [9], [12], [13]. Actin patches, which guide endocytic vesicle internalization and constitute a second F-actin structure, also re-polarize to both cell tips upon NETO [14]. Disruption of proteins comprising these structures similarly jeopardizes growth polarity establishment [15]–[17]. Thus, alteration in protein composition at cell tips is coupled tightly to cytoskeletal rearrangements.
In addition to promoting cell tip growth, several tip-localized cell polarity factors, including Tea1 and Tea4, direct the cell division plane away from cell ends and towards the cell middle for cytokinesis [18], the process by which daughter cells undergo physical separation following nuclear division. However, whether the process of cytokinesis reciprocally modulates cell polarity is unclear. Some observations hint that the cell division machinery may play a role in directing cell polarity. As was previously noted, new ends formed by cell division initiate growth well after old ends. In mutants in which cells remain physically connected at division sites for multiple cell divisions, internal cells can grow, though this occurs sub-apically adjacent to septa [19], [20]. Moreover, many polarity factors localize to the cell division site [4], [21]–[23]; nonetheless, only cell tip-localized populations of these polarity proteins have been demonstrated to contribute to growth polarity in S. pombe.
As in most eukaryotes, cytokinesis occurs in S. pombe through the assembly and constriction of an actomyosin-based cytokinetic ring (CR) [24]. In addition to actin and myosin, several accessory proteins regulate the dynamics and organization of this structure. For one, Cdc15, which contains an N-terminal F-BAR domain and a C-terminal SH3 domain characteristic of the pombe Cdc15 homology protein family [25], has been posited to link CR proteins to the cortical membrane at the division site [26]. Cdc15-binding proteins at the CR include formin, myosin, and the C2 domain protein Fic1 [27], [28]. Fic1 localizes to both interphase cell tips and the cell division site [28], though its specific functions at these sites have not been described. Fic1's budding yeast ortholog, Inn1, contributes to cytokinesis by linking the CR to the ingressing membrane and by participating in septum formation [29], [30]. Septa form in both budding and fission yeasts as cell wall is deposited behind the constricting CR [31]. A conserved signaling network, known as the septation initiation network (SIN) in S. pombe, triggers septum deposition during cytokinesis [32]. Together with the CR, septa provide mechanical force for membrane closure at the cell division site [33]. Subsequent septum degradation allows for abscission [34], [35]. Clearly, various remodeling events must occur at the cell division site for cytokinesis to complete efficiently. Whether such remodeling events also influence daughter cell behavior has never been examined.
While wild-type S. pombe classically grow in a single-celled form, multiple fission yeasts, including S. pombe, possess the ability to assume an invasive, hyphal-like state [20], [36]. The ability of pathogenic fungi to undergo such a morphogenetic switch contributes significantly to fungal infections [37]. Though non-pathogenic, S. pombe, similar to the budding yeast Saccharomyces cerevisiae [38], can transition into invasive growth as a foraging response to low nutrients [36]. Invasive S. pombe form structures that technically qualify as pseudohyphae, for, unlike as in hyphal growth, cytokinetic constriction occurs [39], [40]. Pseudohyphae likely maintain their hyphal-like structure due to cellular adherence and preferential growth at old ends [39], [40]. Intriguingly, it has been postulated that single-celled fission yeast evolved from multicellular, filamentous fungi, with transcriptional networks that ensure efficient cell separation playing predominant roles in the evolution of a single-celled state [41]. Though S. pombe pseudohyphae do not commonly exhibit aborted cytokineses or multicellularity, it is an attractive hypothesis that inefficient, but not entirely defective, cytokinesis might somehow mark new ends to impair their growth and promote the dimorphic switch in S. pombe.
In this manuscript, we define a novel cell cycle control on S. pombe growth polarity, namely that the process of cytokinesis imposes limitations on new end growth competency. Here, we focus on Fic1, which we show to be involved in the re-establishment of polarized cell growth at new ends following cell division. Specifically, we demonstrate that Fic1 polarity function requires its localization to the CR but not to interphase cell tips, and that its protein-protein interactions at the CR, including that with Cdc15, promote bipolar cell growth in the ensuing cell cycle. We present evidence that loss of Fic1 impairs disassembly of the cell division apparatus, with parts of this machinery persisting at new ends following CR constriction. Additional mutants defective in late cytokinesis also exhibit impaired new end growth. Importantly, premature activation of NETO signaling does not fully rescue bipolar growth in cells with late cytokinesis defects, suggesting that cytokinesis-based constraints on S. pombe growth polarity play a central role in defining new end growth competency. We propose that such constraints can provide a mechanistic understanding of how S. pombe and possibly other fungi transition into invasive hyphal-like growth.
Recently, our laboratory identified Fic1, which was implicated in cytokinesis based on its protein and genetic interactions and its localization to the CR [28]. In addition to defects in cytokinesis, deletion of S. pombe fic1+, which is a non-essential gene, resulted in an abnormally high percentage of cells that grew only from one end (i.e., monopolar cells) (Figure 1A–1C). Tip growth was judged using calcofluor staining, as birth scars formed at previous division sites do not stain well with calcofluor and growth can be assessed using the position of these scars relative to cell tips (Figure S1A) [5]. The growth defects observed upon fic1+ disruption suggested that Fic1 not only participates in cytokinesis but also in the establishment of bipolar cell growth.
Although the upstream NETO factors Tea1 and Tea4 localized normally to both cell tips in fic1Δ cells (Figure S1B–S1C), other cell tip proteins implicated in growth polarity regulation exhibited unusual localization patterns in this mutant. Unlike wild-type cells with mostly bipolar actin patch distribution (Figure 1D–1E), a variety of mutants defective in bipolar cell growth exhibit monopolar actin patches [8], [21]–[23]. As in such mutants, the actin patch marker Crn1-GFP [42] accumulated preferentially at one cell end in a high percentage of fic1Δ cells (Figure 1D–1E). Signaling through Rho GTPases controls actin patch organization in S. pombe [13], [43], and the Rho1 activator Rgf1 [21], which was GFP-tagged and imaged with the spindle pole body marker Sid4-RFP [44], likewise predominated on one end of many fic1Δ cells (Figure 1F–1G). Not surprisingly, in both wild-type and fic1Δ cells, Rgf1-GFP and Crn1-RFP concentrated at the same ends (Figure S1D), which were confirmed by calcofluor staining to be the growing ends of fic1Δ cells (Figure S1E). Consistent with Fic1 affecting both actin and Rho networks, deletion of fic1+ was synthetically sick with deletion of genes encoding factors involved in F-actin nucleation (WASp Wsp1) and Rho GTPase regulation (RhoGEF Rgf1 and RhoGAP Rga1) (Figure S1F). Thus, we conclude that the absence of Fic1 upsets patterning of some but not all polarity factors.
To discern whether new and/or old ends were defective in resuming growth following cell division in fic1Δ cells, we performed time-lapse DIC imaging to trace birth scars in live cells. As expected, nearly all wild-type cells underwent NETO prior to subsequent septation (Figure 2A and 2C). However, following roughly two-thirds of fic1Δ cell divisions, either one or both daughter cells failed to initiate new end growth prior to the next septation (Figure 2B–2C). The most predominant growth pattern in fic1Δ cells was that in which one daughter cell underwent NETO while the other did not (Figure 2B–2C), with nearly 70% of those daughter cells that did not exhibit NETO being the younger daughter cell. Unlike tea1Δ and tea4Δ cells, in which one daughter cell commonly fails at its new end and the other daughter cell fails at its old end (Figure 2D) [8], [22], [23], fic1Δ cells were specifically defective in the re-establishment of growth at new ends following cell division (Figure 2B–2C). Intriguingly, tea1Δ fic1Δ double mutants grew mainly in a tea1Δ pattern, though nearly one-fifth of cell divisions produced a T-shaped daughter cell (Figure 2D–2E). Consistent with this, roughly 10% of tea1Δ fic1Δ cells were T-shaped at 25°C, while T-shaped tea1Δ cells were almost never observed at this temperature (Figure 2F). T-shapes always arose in cells that the tea1Δ growth pattern dictated should grow at their new ends (Figure 2D–2E) but that actually grew at neither (Figure 2E and 2G), suggesting these cells polarize at sites other than their tips because growth is inhibited at both ends. These data confirmed that the polarity defect caused by loss of Fic1 stochastically impacts new end growth in a variety of genetic backgrounds. Importantly, fic1Δ new ends that failed to extend in one cell cycle initiated growth as an old end in the next cell cycle, suggesting the defect in growth polarity caused by loss of Fic1 was not permanent. Consistent with a delay but not a block in growth, new ends that initiated growth prior to the next septation did so much later on average in fic1Δ cells than in wild-type cells (120 min versus 75 min) (Figure 2H).
To test whether fic1Δ's polarity defect was independent of S phase completion, we arrested fic1Δ cells in late G2 using cdc25-22, a temperature-sensitive allele of the phosphatase that activates cyclin-dependent kinase at the G2-M transition. As was previously observed [5], otherwise wild-type cells blocked in G2 almost always underwent NETO (Figure 2I–2J). However, roughly half of fic1Δ cells remained monopolar (Figure 2I–2J), indicating that the fic1Δ polarity defect occurs irrespective of S phase completion. To test whether fic1Δ cells were too small to initiate NETO, we measured cell lengths at division. Though slightly shorter on average than wild-type cells (13.3 µm versus 15.3 µm), all fic1Δ cells were longer at division than the minimum length required for NETO (∼9 µm) (Figure 2K) [5]. Therefore, it is unlikely that the fic1Δ growth polarity defect is caused by reduced cell length. These data underscore that loss of Fic1 disrupts the establishment and timing of NETO independently of previously described cell cycle controls.
Though many cell polarity factors localize to the cell division site in addition to interphase cell tips, only the actions of these proteins at interphase cell tips have been demonstrated to be relevant to polarity regulation. As was observed previously [28], cytoplasmic Fic1-GFP localizes to cell tips during interphase and later to the CR during cell division (Figure 3A). We also detected another pool of Fic1-GFP lining the division site as the CR constricted (Figure 3A). Given this localization pattern and the specific new end growth defect of fic1Δ cells, we asked whether Fic1 affected the timing of NETO via its functions at the cell division site.
Like S. cerevisiae Inn1 [30], Fic1 is comprised of an N-terminal C2 domain and a C-terminal stretch of PxxP motifs (Figure 3B). As was found for Inn1 [29], [30], the C terminus of Fic1 (“Fic1C”, amino acids 127–272), expressed from its endogenous locus and GFP-tagged, was sufficient for CR localization, as judged by co-localization with the CR protein Cdc15-mCherry (Figure 3C). In contrast, a GFP-tagged N-terminal C2 domain fragment (“Fic1N”, amino acids 1–126) was never observed at the CR (Figure 3C) though it was produced in vivo (Figure S2). Importantly, medial-localizing Fic1C, unlike Fic1N, supported proper growth polarity establishment (Figure 3D–3F), and, in contrast to full-length Fic1-GFP, Fic1C-GFP was not detected at tips of interphase cells (Figure 3G). We thus conclude that Fic1, unlike other characterized growth polarity factors, does not exert its polarity function at cell tips during interphase, but instead does so at the cell division site during cytokinesis.
Because Fic1's C terminus was necessary and sufficient for proper growth polarity, we examined whether protein-protein interactions at the CR mediated by Fic1's C-terminal PxxP motifs, which bind SH3 domains, govern Fic1's polarity function. Fic1 was originally identified based on its interaction with Cdc15's SH3 domain [28]. As would be expected if association of Cdc15 with Fic1's C terminus is important in establishing the timing of NETO, calcofluor-stained cdc15ΔSH3 cells, which are viable but lack Fic1-Cdc15 interaction [28], exhibited growth polarity defects (Figure 4A–4C and Figure S3A–S3B).
To address the consequence of specifically disrupting Fic1-Cdc15 interaction, we determined which of Fic1's C-terminal PxxP motifs interact(s) with Cdc15's SH3 domain. Previous yeast-two hybrid data indicated Fic1 amino acids 190–269 mediate direct association with Cdc15's SH3 domain [28]. This region contains four of the eleven PxxP motifs within Fic1's C terminus (Figure S3C). To identify which are relevant for Cdc15 interaction, yeast two-hybrid assays using single and combinations of proline to alanine mutations were performed (Figure S3D). Mutation of PxxPs 10 and 11 in combination, or P257 of PxxP 11 alone, abolished the two-hybrid interaction (Figure S3D), and the P257A mutation eliminated co-immunoprecipitation of Fic1-FLAG3 with Cdc15 in vivo (Figure 4D). Supporting the idea that the Fic1-Cdc15 interaction is most relevant during cell division, Fic1-GFP did not accumulate preferentially in Cdc15-mCherry puncta at interphase cell tips (Figure S3E) and co-immunoprecipitation of Fic1-FLAG3 with Cdc15 was considerably stronger in mitosis than in interphase (Figure 4D). This is similar to other Cdc15 protein-protein interactions, which become enriched upon Cdc15 dephosphorylation at mitosis [26]. fic1-P257A cells exhibited monopolar growth defects similar to fic1Δ and cdc15ΔSH3 cells (Figure 4A–4C), confirming that binding of Fic1's C terminus to Cdc15 is critical for Fic1's polarity function. Even so, Fic1-P257A-GFP still localized to the CR (Figure S3F), indicating that medial localization of Fic1 during cytokinesis is necessary but not sufficient for re-establishment of proper growth polarity following cell division.
To corroborate that PxxP-mediated protein-protein interactions at the cytokinetic ring play a predominant role in Fic1's polarity function, we tested whether other interactors participate in S. pombe polarity regulation. The SH3 protein Imp2 has previously been shown to function redundantly with Cdc15 and bind Fic1 during cytokinesis [28]. Consistent with additional Fic1 interactions guiding growth polarity, loss of Imp2 also severely compromised bipolar cell growth (Figure 4A–4C and Figure S3A–S3B). In S. cerevisiae, the Fic1 ortholog Inn1 binds to another SH3 protein, Cyk3 [29], [45], and complexing of these two proteins with the Cdc15 homolog Hof1 has been suggested to direct septum formation and cell separation [29]. We found that S. pombe Cyk3 co-immunoprecipitated with Fic1 in mitosis (Figure 4E), and we also detected direct interaction between S. pombe Cyk3's SH3 domain and Fic1 via yeast two-hybrid (Figure S3G). Accordingly, these interactions appear to be conserved. As was also described in a recent study [46], we found that Cyk3-GFP localized to the CR and division site during cytokinesis, and it was retained at new ends immediately following cell division (Figure 4F). Consistent with these proteins performing a common function, loss of Cyk3 resulted in growth polarity defects similar to those seen upon loss of Fic1 or its interaction with Cdc15 or Imp2 (Figure 4A–4C). Thus, Fic1 collaborates with associated proteins at the CR to execute its growth polarity function, and we postulate that its C terminus acts as an adaptor molecule for SH3 proteins to ensure integration of distinct processes during cytokinesis. Of note, Fic1-P257A-GFP still localized to the CR in imp2Δ cyk3Δ cells (Figure S3H), indicating other CR proteins besides Cdc15, Imp2, and Cyk3 bind Fic1 and likely participate in polarity-relevant events at the division site.
To discern how loss of Fic1 scaffold function during cytokinesis impacts subsequent new end growth, we next defined what aspects of cytokinesis are perturbed in fic1Δ cells. Previous data demonstrated that fic1Δ was synthetically lethal with sid2-250 [28], a temperature-sensitive allele of the SIN kinase Sid2. Consistent with Fic1 and associated factors working in parallel to the SIN, we found that fic1Δ and cyk3Δ suppressed the hyperactive SIN allele cdc16-116 (Figure S4A), and that fic1Δ and cyk3Δ were synthetically sick or lethal with a variety of SIN alleles conferring loss of SIN function (Figure S4A–S4B). These genetic data implied that Fic1 most likely functions during late stages of cytokinesis. In line with this idea, the percentage of fic1Δ cells that had undergone ingression but were still joined at their division sites was more than four times that of wild-type cells (Figure 5A–5B). When cells were arrested in G2 using the cdc25-22 allele, this difference increased, with the percentage of joined cells roughly 15 times greater in the absence of Fic1 (Figure 5A–5B). Similar to S. cerevisiae inn1Δ cells [29] and S. pombe cdc15ΔSH3 cells [28], many G2-arrested fic1Δ daughter cells that were still joined at division sites exhibited membranous bridges (Figure S4C). These findings verified that the completion of cell division is perturbed in the absence of Fic1.
Consistent with early cytokinesis events proceeding appropriately without Fic1, time-lapse imaging of myosin regulatory light chain Rlc1-GFP [47], [48] along with spindle pole body marker Sid4-GFP revealed that the CR formed and constricted normally in fic1Δ cells (Figure 5C–5D). However, at the termination of CR constriction, parts of the CR persisted at the division plane (Figure 5E–5G and Figure S4D). During cytokinesis, the septum closes behind the constricting CR, and septum closure can be visualized using the β-glucan synthase GFP-Cps1 [49], [50]. As cytokinesis progresses, two GFP-Cps1 dots marking the leading edge of the septum can be seen getting progressively closer in the division plane, and these dots eventually join into one just as the CR completes constriction (Figure 5F). We found that Rlc1-mCherry3 remained at the division site following septum closure on average longer in fic1Δ cells compared to wild-type cells (22 min versus 8 min) (Figure 5E–5F). Consistent with these remnants representing the CR as a whole and not just Rlc1, phalloidin staining revealed atypical actin-rich masses, in addition to normal actin patches, flanking septa in fic1Δ cells (Figure 5G). By expressing LifeAct-GFP, we verified that these abnormal actin masses co-localized to a high degree with Rlc1-mCherry3 in a fic1Δ genetic background (Figure S4D). Thus, we conclude that the CR does not disassemble properly at the conclusion of cell division in fic1Δ cells.
In addition to CR-associated factors, glucanase Eng1-GFP [35] persisted at ingressed division sites significantly longer in fic1Δ cells compared to wild-type cells (on average, 51 min versus 21 min) (Figure S4E–S4F). Because glucanases execute septum degradation [34], [35], these data suggest that cell wall turnover is inefficient at fic1Δ septa. We thus conclude that loss of Fic1 jeopardized the completion of cell division, stalling remodeling of new ends in the next cell cycle.
Because faulty cytokinesis led to persistence of parts of the cell division machinery at fic1Δ division planes, we speculated that these remnants might deter subsequent polarized growth at new ends. If this were the case, one would expect other mutants with late cytokinesis defects to also show erroneous new end growth. Previous data had indicated that Fic1-associated Imp2 contributes to CR disassembly, with imp2Δ cells exhibiting abnormal actin structures flanking previous division sites [51]. Though we had shown that imp2Δ cells are defective in bipolar cell growth (Figure 4A–4C), we wanted to confirm that their growth defect was specific to new ends. Using time-lapse DIC imaging, we found that roughly 75% of imp2Δ cell divisions produced at least one daughter cell that failed at new end growth (Figure 6A). Interestingly, both imp2Δ daughter cells failed at new end growth in the majority of cases (Figure 6A–6B). Therefore, proper disassembly of CR components correlates with new end competency for polarized growth.
In addition to showing CR disassembly defects, fic1Δ cells also exhibited delays in septum remodeling at the division site. We therefore tested if disruption of septum degradation could likewise impact polarized growth. Loss of Eng1 or its cooperating glucanase, Agn1 [34], resulted in high percentages of monopolar growth (Figure 6C–6D and Figure S5A). Moreover, the growth defect of eng1Δ daughter cells was specific to new ends (Figure 6A–6B), and, similar to fic1Δ cells, eng1Δ daughter cells that initiated NETO prior to the next septation did so on average later than wild-type cells (129 min versus 75 min) (Figure S5B). Anillin-like Mid2 and the septin ring, of which Spn1 and Spn4 form the core [52], target these glucanases into a ring structure around septa [53]. Loss of any of these proteins likewise impaired bipolar cell growth (Figure 6C–6D and Figure S5A). In addition, though the majority of spn1Δ daughter cells failed at new end growth (Figure 6A–6B), those that initiated NETO prior to the next septation took longer on average to do so than wild-type cells (95 min versus 75 min) (Figure S5B). We therefore conclude that defective completion of cell wall remodeling at the division site, in addition to improper disassembly of CR components, compromises NETO efficiency.
The SIN coordinates many aspects of CR and septum regulation during late cytokinesis. Not only does SIN signaling oversee maintenance of a mature, homogenous CR [54], it mediates Cps1 targeting and accumulation at the division site [49], [50]. Loss of SIN signaling during cytokinesis can thus lead to CR fragmentation [54] and abortive septation [49]. Given these phenotypes and the synthetic genetic interactions between fic1Δ and SIN mutants (Figure S4A–S4B), we examined the relevance of the SIN to new end growth control. Temperature-sensitive alleles of genes encoding the SIN kinases Cdc7 and Sid2 caused mild but statistically significant growth polarity defects at semi-restrictive temperature (Figure 6C–6D and Figure S5A). A temperature-sensitive allele of the gene encoding Cps1, which functions downstream of the SIN, caused dramatic defects in establishing bipolar cell growth even at permissive temperature (Figure 6C–6D and Figure S5A). Additionally, a high proportion of cps1-191 cells failed specifically at new end growth (Figure 6A–6B), and those that were able to trigger NETO prior to subsequent septation did so on average later than wild-type cells (107 min versus 75 min) (Figure S5B). Not surprisingly, we were able to detect incomplete ingression of cps1-191 cells shifted to the restrictive temperature during cytokinesis (Figure S5C), again suggesting that these mutants experience remodeling errors at the division site.
Currently, the mechanism of membrane remodeling and scission at the S. pombe division site is unclear. In a variety of other organisms, endosomal sorting complex required for transport (ESCRT)-III factors contribute to this process [55]. ESCRT-III components have not been implicated in S. pombe cytokinesis regulation, though ESCRT-III-associated AMSH (S. pombe Sst2) localizes to the division site [56]. We found that deletions of genes encoding ESCRT-III components Vps2 and Vps24 or ESCRT-III-associated Sst2 were synthetically sick with a variety of loss-of-function cytokinesis alleles, including imp2Δ and cps1-191 (Figure S5D). Interestingly, loss of Vps2, Vps24, or Sst2 resulted in monopolar percentages significantly greater than observed for wild-type cells (Figure 6C–6D and Figure S5A), and nearly half of vps24Δ cell divisions resulted in one or both daughter cells that failed at new end growth prior to the next septation (Figure 6A–6B). Though these phenotypes were less penetrant than in other mutants, we speculate that ESCRT-III function guides membrane remodeling at the conclusion of S. pombe cell division to impact new end polarized growth.
Of note, deletion of rlc1+ or paxillin pxl1+, which function primarily in early actomyosin function at the CR [47], [48], [57], [58], did not alter growth polarity percentages as significantly as other mutations or deletions (Figure 6C–6D and Figure S5A). Indeed, less than half of non-septated rlc1Δ and pxl1Δ cells were monopolar (Figure 6C), and the monopolar septated percentages of these genotypes were more similar to wild-type percentages than were those of the other mutants examined (Figure 6D). We therefore conclude that early steps in cytokinesis do not impact subsequent polarized cell growth as much as the terminal steps in cell division.
If faithful remodeling of the division site is important for growth competency of new ends, then one would expect that prematurely triggering NETO signaling just after cell division should not fully rescue the growth polarity defects of late cytokinesis mutants. To test this, we constructed a mutant that would undergo constitutive NETO. As over-expression of a fusion protein linking cell tip-associated Tea1 with formin For3 induces NETO in G1 [8], we integrated a Tea1-For3 fusion (Figure 7A) into the endogenous tea1+ locus and deleted the single copy of the for3+ gene. We confirmed that the Tea1-For3 fusion protein was produced in vivo (Figure 7B) and verified that this fusion was sufficient to induce NETO in a cdc10-V50 G1 arrest (Figure S6A–S6B). As previously reported [7], double deletion of tea1+ and for3+ resulted in general cell rounding (Figure 7C). However, expression of the Tea1-For3 fusion protein in the absence of Tea1 and For3 individually caused cells to regain their rod-shaped appearance (Figure 7C). Intriguingly, a high percentage of tea1-for3 cells were either septated or exhibited cytokinesis defects (Figure 7C–7D), and tea1-for3 cells were significantly longer at division than wild-type cells (on average, 18.3 µm versus 15.3 µm) (Figure 7C and 7E). Thus, though the endogenous Tea1-For3 fusion protein functioned in prematurely triggering NETO, it also affected cell division.
To analyze tea1-for3 cells in real-time, we performed time-lapse DIC imaging. As expected, most tea1-for3 cells underwent NETO before the next cell division (Figure 8A), with nearly 75% of new ends initiating growth within 50 minutes of septum splitting (Figure 8B). Nonetheless, some tea1-for3 outliers took much longer to extend at tips created by cell division (Figure 8B). After grouping the times needed for tip growth to occur at previous division sites relative to the amount of time needed for the mother cell to complete cytokinesis, we found that newly-formed tips that took longer to initiate growth had been formed by more inefficient cytokinesis (Figure 8C–8D). As distal tip growth continued in cells undergoing division (Figure 8D) and appeared unimpeded by additional factors, these findings suggested that faulty cytokinesis imposes constraints at previous division sites that counteract positive polarizing cues. We corroborated this model by expressing the Tea1-For3 fusion in fic1Δ cells. Although tea1-for3 cells were mostly bipolar, tea1-for3 fic1Δ cells showed a high percentage of monopolar growth (Figure 8E–8G). These findings confirmed that efficient completion of cytokinesis is critical for new end growth, even when signaling networks responsible for NETO are prematurely activated.
S. pombe undergoing a dimorphic switch from single-celled to invasive form grow primarily in a monopolar fashion at old ends [39], [40]. Moreover, it has been postulated that cytokinesis errors might contribute to a hyphal-like transition in S. pombe [41]. We therefore considered that cytokinesis-based constraints on S. pombe growth polarity might facilitate invasive growth transitions. Using techniques similar to those described previously [40], [59], we tested whether various cytokinesis mutants displaying defective bipolar growth could form pseudohyphae into 2% agar. Cells lacking Fic1 or its interacting partners Cyk3 or Imp2 were significantly more invasive than wild-type cells (Figure 9A–9B). Like other invasive S. pombe mutants [39], [40], these mutants formed pseudohyphae composed of single cells oriented in filament-like projections (Figure 9C and Figure S7A). In addition to these strains, we found other cytokinesis mutants exhibiting high degrees of monopolar growth (spn1Δ, cdc7-24, and vps24Δ) to also be highly invasive and to form pseudohyphal projections into 2% agar (Figure 9A–9B and Figure S7A). Of note, the vps24Δ strain showed drastically more invasive growth than the others, though the reasons for this are currently unclear. rlc1Δ and pxl1Δ, which possess cytokinesis defects that do not considerably impact polarized cell growth (Figure 6C–6D and Figure S5A), invaded less efficiently on 2% agar than cytokinesis mutants exhibiting NETO defects (Figure 9A–9B). This supports the notion that defective cytokinesis promotes the dimorphic switch most robustly when it results in faulty NETO. As has previously been observed, tea1Δ also invaded well on 2% agar (Figure S7B–S7D). Thus, though cytokinesis-based constraints on growth polarity support enhanced S. pombe invasiveness, other polarity defects, which are not entirely specific to new ends, can do so as well.
Consistent with bipolar growth defects accompanying pseudohyphal growth, tea1-for3 cells, which experience constitutive NETO induction, almost never extended pseudohyphae into 2% agar (Figure 9D–9E). Because cytokinesis-based constraints on growth polarity partially override tip-based NETO signaling, we reasoned that tea1-for3 cells should become more invasive upon loss of Fic1. Indeed, on 2% agar tea1-for3 fic1Δ cells formed pseudohyphae (Figure S7E), which were more numerous than those observed for wild-type and tea1-for3 strains (Figure 9D–9E). Thus, perturbations in cytokinesis cause growth polarity errors that facilitate pseudohyphal growth even upon constant NETO signaling.
Lastly, we asked whether loss of polarity-relevant cytokinesis factors could partially rescue invasiveness of an asp1Δ strain, which is unable to undergo the dimorphic switch due to an inability to sense external cues [40]. Previously, it was demonstrated that asp1Δ cells form a biofilm-like colony on 0.3% agar (Figure 9F) [40]. Growth on 0.3% agar is more sensitive for assaying invasiveness of strains that invade less efficiently, as wild-type colonies form protrusions on 0.3% agar but extend relatively few pseudohyphal projections into 2% agar (Figure 9A–9B and 9F) [40]. We therefore assessed the effect of cytokinesis defects on asp1Δ invasiveness by testing whether asp1Δ strains that also lacked relevant cytokinesis factors still formed biofilms on 0.3% agar. Intriguingly, double deletion strains of asp1Δ with fic1Δ, spn1Δ, or vps24Δ did not form biofilms on 0.3% agar but instead made projections into and on the surface of the agar (Figure 9F). In contrast, double deletion strains of asp1D with either rlc1Δ or pxl1Δ still formed biofilms on 0.3% agar (Figure 9F). We thus conclude that cytokinesis-based constraints on polarized cell growth in S. pombe can foster invasiveness even in the absence of typical nutritional signals.
In this study, we have shown how cytokinesis and cell polarity crosstalk to regulate fission yeast morphogenesis. Our data support a model (Figure 10) in which Fic1 acts as an adaptor at the CR, where it guides proper completion of cytokinesis and thereby affects division site remodeling. Loss of Fic1, its interactions, or parallel pathways results in delayed growth at new ends, even upon constitutive activation of NETO signaling. Impaired bipolar cell growth resulting from defective cytokinesis in turn enhances S. pombe invasiveness.
The majority of S. pombe monopolar mutants previously analyzed fail at old end growth [4]. However, the cytokinesis mutants studied here were predominantly new end growth defective. As in other organisms [60], numerous S. pombe proteins known to affect growth polarity localize to the division site; this has fostered speculation that signaling at both cell tips and the division site might impact growth zones [23]. However, whether or not the cytokinesis functions of these proteins can specifically impact cell polarity has received little attention, especially in S. pombe. Our data provide evidence directly linking division site organization to S. pombe growth polarity. Because many factors involved in completing cell division likely also impact subsequent polarized growth, we believe our data could explain the involvement of diverse proteins in this process.
In other organisms, cytokinesis proteins appoint local regions of the cell cortex for growth following cell division [61]–[63]. For example, several budding yeast proteins, which remain at the cell cortex following cell division, have been reported to convey a cortical “tag” that marks the position of the next bud site [61], [62]. Similarly, during Drosophila melanogaster neurogenesis, cytokinetic furrow components mark the site from which the first dendrite will sprout [63]. In these cases, cytokinesis factors confer a positive polarizing cue adjacent to previous division sites, which contrasts with our findings in S. pombe where the cell division machinery impedes polarization and growth at new ends created by cell division.
The fact that S. pombe grow at old but not new ends after cell division is somewhat counterintuitive [4], especially because the cell growth machinery concentrates at the division site. Upon the completion of cytokinesis, the S. pombe growth machinery mysteriously shuttles to old ends rather than remaining at new ends. Why does the growth machinery relocate from the division site to old ends? One explanation is that new end cortices must be re-structured to become competent for tip growth. Indeed, specific lipid and cell wall variants contribute to S. pombe cytokinesis [64], [65], and local rearrangements of these may be required for growth activation. Moreover, the persistence of CR factors at new ends might create physical barriers to cytoskeletal elements, such as actin cables, required for tip growth. An inhibitory role for cell division in polarization is supported by studies of mutants that undergo multiple rounds of cytokinesis without physically separating, because internal cells in these structures do not grow into septa but branch adjacently. In a single-celled context, we speculate that when late cytokinetic events are perturbed, inherent delays in cortical re-structuring are exacerbated, causing growth polarity defects at new ends. In cases in which one daughter cell fails at NETO while the other is successful, we suspect these arise due to unequal partitioning of cytokinesis remnants and/or differences in cell cycle stages or life histories of daughter cells.
Consistent with furrow remodeling affecting polarization in other organisms, initial cellular protrusions of dividing mammalian cells orient away from the midbody linking daughter cells during abscission [66]. Only after the completion of cell division does polarization also occur near the latent division site [66]. Moreover, forced entry of HeLa cells with monopolar spindles into cytokinesis results in anucleate daughter cells that, similar to their nucleated counterparts, exhibit membrane protrusions only distal to cleavage furrows [67]. Thus, similar to our model in S. pombe, some factor at the division site cortex, and not a cell's cytosolic constituents, requires remodeling for post-cytokinetic polarization. Recent evidence indicates that mechanosensory pathways can direct cell polarization away from points of tension [68]. As modeling predicts that cortical tension peaks at the division plane during cytokinesis [69], it will likewise be important to assess the relevance of mechanical cues to cytokinesis-based polarization events.
Previous work has implied that association of microtubule-associated protein Tea1 with formin For3 at new ends is sufficient for NETO [8]. In our study, we expressed an endogenous Tea1-For3 fusion that could induce premature NETO. However, when cytokinesis was perturbed in tea1-for3 mutants, NETO was delayed. We posit that local cortical abnormalities in cell wall, membrane, or associated factors can partially override typical growth cues in S. pombe, as has been observed in some plants [70]. Upon defective cytokinesis, such abnormalities at the division site may physically inhibit cell growth at new ends. These defects can lead to the formation of T-shaped cells when old end growth is also blocked, as in tea1Δ fic1Δ mutants. Our data underscore robust completion of cytokinesis as a major determinant of S. pombe NETO.
Is it beneficial for a cell to halt new end growth until well after cytokinesis completion? As mentioned previously, human cells undergoing division initially move away from each other, creating a pulling force that could contribute to abscission [66]. Highly adherent mammalian cells can actually complete cytokinesis, with some defects, in the absence of cortical myosin from the cleavage furrow [71]. Constriction-independent cytokinesis was first observed in the amoeba Dictyostelium discoideum [72], which accomplishes this task by likewise polarizing and growing distally to the division site [73]. One could imagine that in cases where S. pombe cell separation is delayed, tip growth at old ends might contribute similar forces to aid in abscission. Premature new end growth signaling might unbalance these forces, leading to exacerbated cytokinesis delays as in some tea1-for3 cells. Premature new end growth might also interfere with remodeling during cytokinesis and thereby result in cell division defects. These findings highlight interdependence between the cell polarization and division machineries in S. pombe.
Our data indicate that Fic1's C terminus, and not its C2 domain, represents its major cytokinetic functional domain, contrasting with data reported for S. cerevisiae Inn1 [29], [30]. Why is Fic1's C2 domain dispensable for Fic1's cytokinesis, and thus polarity, functions? Sequence alignment indicates that there is in fact very low sequence identity between Fic1 and Inn1 C2 domains [30]. If Fic1's C2 domain is unable to perform functions or mediate interactions that Inn1's C2 domain can, it seems reasonable that Fic1-interacting proteins may be able to compensate. Consistent with this idea, over-expression of S. cerevisiae Cyk3 suppresses cytokinesis defects of inn1Δ mutants, suggesting Cyk3 function overlaps with Inn1 [29]. These data support that Fic1's C terminus is an efficient signaling platform, which scaffolds SH3 domain proteins through its PxxP motifs to ensure coherent integration of late cytokinesis signals.
What is the specific function of the Fic1 scaffold during cytokinesis? Our data indicate that loss of Fic1 leads to faulty CR disassembly and prolonged persistence of factors at the division site. CR disassembly defects were also observed in inn1Δ S. cerevisiae mutants, leading to speculation that Inn1 might stabilize the constricting CR by physically linking it to the ingressing membrane [30]. Subsequent findings countered that Inn1's C2 domain cannot bind phospholipids, and it was postulated instead that Inn1 cooperates with Cyk3 to coordinate cell wall deposition [29]. As in fic1Δ cells, septation and CR disassembly defects commonly accompany one another [51], [74]. Because these processes are inextricably linked, it is currently difficult to tease apart which defect precedes the other in fic1Δ cells. Moreover, completion of cytokinesis also requires lipid rearrangements in both animal cells and S. pombe [64], [75], and membrane bridges were observed in fic1Δ cells. We thus envision that Fic1's C terminus links signaling pathways that guide completion of multiple tasks during late cytokinesis and thereby affect new end remodeling. Of note, we believe that defects in early cytokinesis do not significantly alter bipolar growth establishment, as later defects more directly impinge on division site remodeling and have less time to be remedied before the next cell division. Our data furthermore support the notion that CR constriction and disassembly occur independently [74], as CR constriction but not disassembly proceeded appropriately in fic1Δ cells.
As fungal hyphae consist of long chains of cells, the transition into hyphal growth requires strict inhibition of cytokinesis. In some yeasts, Cdc14 phosphatase activates the Ace2 transcriptional program [76], which triggers expression of cell separation enzymes [77]. Upon the hyphal transition in Candida albicans, this signaling cascade is disrupted [78], and other transcription factors suppress expression of Ace2 targets [79]. Therefore, cytokinetic inhibition in hyphae is believed to operate largely on a transcriptional level, and reactivation of the Ace2 transcriptional program is thought to be responsible for the evolution of single-celled yeast growth [41].
In this study, we showed that fission yeast cells that undergo defective, yet not wholly abortive, cytokinesis exhibit enhanced invasive capacity. We believe cytokinesis-based constraints on growth polarity assist the transition into pseudohyphal growth because they force S. pombe to orient outwards and grow predominantly from old ends, a pattern commonly observed in S. pombe pseudohyphal growth [39], [40]. However, as demonstrated by tea1Δ cells, other changes in polarity can also enhance S. pombe invasiveness. Though not specifically defective at new end growth, tea1Δ cells grow predominantly in the direction of the mother cell, and these alterations in polarity might likewise favor growth orientations that are more conducive than bipolar growth to the invasive process.
We believe our data suggest that manipulation of cytokinesis proteins, and not necessarily signaling cascades that feed into downstream transcriptional pathways, can directly modulate the dimorphic switch. We thus speculate that the cytokinetic machinery might represent a direct target of the pseudohyphal developmental program. Intriguingly, loss of cytokinesis proteins that affect NETO rescued invasiveness of an asp1Δ mutant, which lacks the ability to detect nutritional cues [40] deemed important for the S. pombe dimorphic switch [36]. Because various environmental cues also regulate hyphal morphogenesis in pathogenic fungi [80], it will be important to assess the relative significance of cytokinesis-based controls on polarized growth for invasiveness in these species.
The S. pombe strains used in this study (Table S1) were grown in either yeast extract (YE) or Edinburgh minimal media with relevant supplements. fic1+, fic1N, fic1C, fic1-P257A, crn1+, tea4+, rlc1+, tea1+, for3+, and tea1-for3 were tagged endogenously at the 3′ end with GFP:kanR, FLAG3:kanR, mCherry3:kanR, RFP:hygR, V5:kanR, or V5:hygR cassettes as previously described [81]. A lithium acetate method [82] was used in S. pombe tagging transformations, and integration of tags was verified using whole-cell PCR and/or fluorescence microscopy. Introduction of tagged loci into other strains was accomplished using standard S. pombe mating, sporulation, and tetrad dissection techniques. For blocking of cdc25-22 strains in G2, cells were grown at 25°C and then shifted to 36°C for 3 h. For blocking of nda3-KM311 strains in prometaphase, cells were grown at 32°C and then shifted to 18°C for 6.5 h. For blocking of cdc10-V50 strains in G1, cells were grown at 25°C and then shifted to 36°C for 4 h. For blocking of cps1-191 cells in a cytokinesis arrest, cells were grown at 25°C and then shifted to 36°C for 3 h.
Mutants and truncations of fic1 were expressed from the endogenous fic1+ locus. To make these strains, a pIRT2 vector was originally constructed in which fic1+ gDNA with 5′ and 3′ flanks was inserted between BamHI and PstI sites of pIRT2. Mutations were then introduced via site-directed mutagenesis. The fic1(aa1-126) construct was made by inserting a stop codon after residue 126. The fic1(aa127-272) construct was created by inserting XhoI sites before both the start codon and residue 127, digesting with XhoI to release the internal fragment, re-ligating the plasmid, and adding a start codon after the remaining XhoI site. fic1Δ was then covered by these pIRT2-fic1 constructs, and stable integrants resistant to 5-FOA were isolated and confirmed by whole-cell PCR and western blotting.
To make the tea1-for3 fusion, a pIRT2 vector was originally constructed in which tea1+ gDNA with 5′ and 3′ flanks was inserted between SacI and SphI sites of pIRT2. Site-directed mutagenesis was performed to replace the tea1+ stop codon with a SmaI/SalI/PstI multiple cloning site. for3+ gDNA was amplified with a small N-terminal linker sequence and inserted between SmaI and PstI in this multiple cloning site (linker residues are Pro-Gly-Ade-Gly-Ade-Gly-Ade accounting for restriction site and added residues). tea1Δ was then covered by this pIRT2-tea1-for3 construct, and stable integrants resistant to 5-FOA were isolated and confirmed by whole-cell PCR and western blotting. An integrant was subsequently mated with for3Δ, such that we could isolate tea1-for3 strains in which tea1+ and for3+ were lacking.
Expression of acyl-GFP [83] was controlled by the thiamine-repressible nmt1 promoter of pREP3 [84], [85]. Expression of LifeAct-GFP [86] was controlled by the thiamine-repressible nmt81 promoter of pREP81 [87]. Expression from these nmt promoters was kept off by addition of 5 µg/mL thiamine to the medium, and expression was induced by washing and culturing in medium lacking thiamine for at least 24 h.
Spot assays to analyze genetic interactions were performed as previously described [88], except that all were done on YE agar. Synthetic interactions were judged based on differences in growth between double mutants and relevant single mutants.
Yeast two-hybrid analysis was performed as previously described [89], except that the bait and prey plasmids were either empty or encoded Cdc15 SH3 (aa843–927) [28], Cyk3 SH3 (aa1–59), wild-type [28] or mutant Fic1(aa190–269) fragments, or full-length Fic1.
Cells were lysed by bead disruption in NP40 lysis buffer in either native or denaturing conditions as previously described [90], except with the addition of 0.5 mM diisopropyl fluorophosphate (Sigma-Aldrich). Proteins were immunoprecipitated by anti-FLAG (Sigma-Aldrich), anti-Cdc15 [28], or anti-V5 (Invitrogen) antibodies. Immunoblot analysis of cell lysates and immunoprecipitates was performed using anti-FLAG, anti-Cdc15, anti-V5, anti-GFP (Roche), or anti-Cdc2 (Sigma-Aldrich) antibodies as previously described [88].
Live-cell bright field images as well as all still images of cells expressing proteins endogenously-tagged with GFP, RFP, or mCherry were acquired on a spinning disc confocal microscope (Ultraview LCI; PerkinElmer) equipped with a 100X NA 1.40 PlanApo oil immersion objective, a 488-nm argon ion laser (GFP), and a 594-nm helium neon laser (RFP, mCherry). Images were taken via a charge-coupled device camera (Orca-ER; Hamamatsu Phototonics) and processed using Metamorph 7.1 software (MDS Analytical Technologies; Molecular Devices). Bright field images were used in determining cell lengths at division. Time-lapse GFP images of cyk3-GFP sid4-GFP cells secured on agar pads, which were sealed by Valap (a Vaseline, lanolin, and paraffin mixture), were acquired every 3 min using this system. Z-sections were acquired for all fluorescence images and combined into maximum projections. Cells were grown to log phase at 25°C before such imaging.
Images of yeast cells and pseudohyphae on YE agar plates were acquired by focusing a camera (PowerShot SD750; Canon) through a microscope (Universal; Carl Zeiss) equipped with a 20X NA 0.32 objective.
All other microscopy was performed using a personal DeltaVision microscope system (Applied Precision). This system includes an Olympus IX71 microscope, 60X NA 1.42 PlanApo and 100X NA 1.40 UPlanSApo objectives, fixed- and live-cell filter wheels, a Photometrics CoolSnap HQ2 camera, and softWoRx imaging software. The microscopy performed using this system was as follows:
To assay pseudohyphal invasion into 2% agar, 5 µl containing a total of 105 cells were spotted on 2% YE agar and incubated at 29°C for 20 days. Colonies were subsequently placed under a steady stream of water and surface growth was wiped off using a paper towel. These methods were established in previous studies [40], [59].
To assay whether specific mutants rescued invasiveness of an asp1Δ strain on 0.3% agar [40], 1 µl containing 106 cells was spotted on 0.3% YE agar as well as onto 2% agar as a control. Plates were incubated at 29°C for 12 days, at which point colony growth and/or biofilm formation were visualized.
|
10.1371/journal.pgen.1003680 | The Yeast Environmental Stress Response Regulates Mutagenesis Induced by Proteotoxic Stress | Conditions of chronic stress are associated with genetic instability in many organisms, but the roles of stress responses in mutagenesis have so far been elucidated only in bacteria. Here, we present data demonstrating that the environmental stress response (ESR) in yeast functions in mutagenesis induced by proteotoxic stress. We show that the drug canavanine causes proteotoxic stress, activates the ESR, and induces mutagenesis at several loci in an ESR-dependent manner. Canavanine-induced mutagenesis also involves translesion DNA polymerases Rev1 and Polζ and non-homologous end joining factor Ku. Furthermore, under conditions of chronic sub-lethal canavanine stress, deletions of Rev1, Polζ, and Ku-encoding genes exhibit genetic interactions with ESR mutants indicative of ESR regulating these mutagenic DNA repair processes. Analyses of mutagenesis induced by several different stresses showed that the ESR specifically modulates mutagenesis induced by proteotoxic stress. Together, these results document the first known example of an involvement of a eukaryotic stress response pathway in mutagenesis and have important implications for mechanisms of evolution, carcinogenesis, and emergence of drug-resistant pathogens and chemotherapy-resistant tumors.
| Cellular capability to mutate its DNA plays an important role in evolution and impinges on medical issues, including acquisition of mutator phenotypes by cancer cells and emergence of drug-resistant pathogens. Whether and how the environment affects rates of mutation has been studied predominantly in the context of environmental agents that damage DNA (e.g. UV and γ-rays). However, it has been observed that conditions of chronic non-DNA-damaging stress (e.g. starvation or heat shock) also increase mutagenesis. It has been shown that in bacteria, activation of the general stress response activates a pro-mutagenic pathway and thus promotes mutagenesis during periods of stress. However, in eukaryotes, so far there has been no evidence of a stress response regulating mutagenesis. In this manuscript we demonstrate that in budding yeast, a model eukaryote, the general environmental stress response (ESR) regulates mutagenesis induced by proteotoxic stress (accumulation of unfolded proteins) at several loci. We also identify two pro-mutagenic DNA metabolic pathways that contribute to this mutagenesis and present genetic data showing that the ESR regulates these pathways. Together, these data advance our understanding of how cellular sensing and responding to environmental cues affect cellular capability for mutagenesis.
| Sensing and responding to environmental cues are ubiquitous cellular functions essential for survival. Budding yeast cells respond to a variety of stresses by inducing or repressing specific sets of genes in a stereotypical fashion that, to a certain degree, does not depend on the identity of the stress. This process is termed the environmental stress response (ESR) [1], [2]. Paradoxically, the ESR provides little protection from the initiating stress – genes required to survive the stress do not significantly overlap those that change expression in response to the stress and mutations in ESR regulators do not significantly sensitize cells to stress [3], [4]. This observation raises the possibility that ESR activation may have other cellular roles. One potential role of the ESR is suggested by observations that chronic stress can induce genetic instability in different organisms [5]–[10]. The phenomenon of stress-associated genetic instability impinges on medical issues, such as the role of tumor microenvironment in genetic instability of cancer cells and emergence of drug-resistant pathogens and chemotherapy-resistant tumors. While in Escherichia coli stress response can activate mutagenic DNA repair [11], [12], no evidence exists as yet for the involvement of the ESR in mutagenesis in a eukaryote. In this manuscript we investigate the effects of stress on mutagenesis in yeast and the role of the ESR in this process.
Numerous studies have indicated that environmental stress can affect genome stability. For instance, in mammalian cells in tissue culture hypoxia and starvation can suppress error-free DNA repair pathways (e.g. mismatch repair and homologous recombination) and cause an increase in mutagenesis [13]–[19]. In yeast, various types of stress can affect chromosome segregation and promote aneuploidy [6]. Interestingly the most potent inducer of aneuploidy is proteotoxic stress, e.g. inhibition of HSP90 protein chaperone by radicicol [6]. One explanation of this phenomenon is that HSP90 can become “overtaxed”, such that its client proteins that function in chromosome segregation would interact with their targets in a misfolded, disfunctional state, with aberrant consequences for ploidy maintenance [6]. Other instances of genetic instability, in particular mutagenesis, were reported in response to chronic osmotic and DNA replication stresses [20], [21]. These types of stress are thought to be mutagenic at least in part because they can directly cause DNA damage: osmotic stress induces DNA breaks [22] and replication stress stalls DNA replication forks and creates regions of ssDNA [23]. Finally, several groups reported the phenomenon of “adaptive mutation” (alternately termed “stationary phase” or “selection-induced” mutation) in budding yeast Saccharomyces cerevisiae ([9], [10]; for a comprehensive review, see [24]). In these experiments, starvation for an amino acid induced reversions of mutations in amino acid biosynthesis genes, enabling cells to grow on the starvation medium. Besides amino acid starvation, “adaptive” mutants were also observed after exposure of yeast cells to the drug canavanine [25]. Together these studies suggest that sensing and responding to environmental stress may have important consequences for genome stability, but mechanisms underlying this assertion and the involvement of stress responses in these phenomena remain underexplored.
Cellular pathways that function in mutagenesis in eukaryotes have been extensively studied, predominantly by identifying mutants defective in spontaneous and/or DNA damage-induced mutagenesis. These analyses have identified DNA translesion synthesis (TLS) as a key mutagenic pathway in both yeast and higher eukaryotes [26]. In yeast, TLS is largely carried out by two specialized DNA polymerases Rev1 and polymerase ζ (Polζ) that, unlike replicative polymerases δ and ε, can polymerize DNA using damaged or distorted DNA templates and thus function in DNA damage bypass pathways [27]. While Rev1 and Polζ can interact and function together in vivo, they do not have identical phenotypes in all mutation assays, suggesting that they have some independent roles [26]. In contrast to spontaneous and DNA damage-induced mutagenesis, genetic requirements for stress-associated mutagenesis are less well characterized. However, Heidenreich et al. reported that starvation-associated frameshift reversion was independent of both Rev1 and Polζ [28], instead requiring proteins that function in non-homologous end joining (NHEJ), such as Ku [29]. NHEJ is a DNA double-strand break (DSB) repair pathway that directly ligates broken ends together without relying on a homologous template [30]. Whether these mutagenic repair pathways are important for other types of stress-associated mutagenesis and are influenced by cellular stress responses has not yet been examined.
The ESR in S. cerevisiae is activated in response to any one of a large number of environmental stresses [1], [2], including DNA damaging agents, such as the DNA alkylating drug methylmethane sulfonate and inhibitor of ribonucleotide reductase hydroxyurea [31]. The genes that are repressed by the ESR largely function in translation and other growth-promoting pathways. The genes that are induced by the ESR function in several molecular processes, such as protein folding and repair of oxidative damage, but the functions of many of them are not known. Stress-driven induction of most ESR-activated genes is largely regulated by partially redundant transcription factors, Msn2 and Msn4 [1], [2], [32], [33]. Of the two proteins, Msn2 plays a greater role in transcriptional activation and its behavior has been relatively well examined. In unstressed cells, Msn2 is localized almost exclusively to the cytoplasm. Upon a sudden stress, such as a drop in glucose concentration or osmotic shock, Msn2 moves into the nucleus in the majority of cells where it binds to stress response elements in its targets' promoters to activate transcription [34]–[36].
In this manuscript, we examine the effect of the ESR on mutagenesis in S. cerevisiae by analyzing spontaneous and stress-associated mutagenesis in mutants lacking MSN2 and MSN4. We report that the drug canavanine causes proteotoxic stress and activates the ESR, and that under conditions of severe canavanine stress MSN2 and MSN4 promote certain types of mutation events, most notably single nucleotide deletions in simple repeats. Furthermore, while MSN2 and MSN4 are dispensable for mutagenesis induced by osmotic and DNA replication stresses, they can promote or suppress mutagenesis induced by different types of proteotoxic stress. Furthermore, TLS polymerases and Ku also function in proteotoxic stress-induced mutagenesis and exhibit unanticipated genetic interactions with MSN2 and MSN4. Together, these results implicate the yeast ESR in regulation of mutagenic DNA repair pathways activated by proteotoxic stress.
Canavanine is a toxic analog of arginine and can be imported into yeast cells via an arginine transporter, Can1. The CAN1 gene is a commonly used mutation reporter in yeast as can1 mutations can be selected on plates containing canavanine. In a typical experiment, CAN1 cultures grown in the absence of canavanine are plated on canavanine-containing plates, so that only can1 mutants can form colonies. The mutation rates are then calculated from the can1 colony distribution data [25], [37]. If can1 mutants form spontaneously during cell division in culture, the frequency of mutants should follow a Luria-Delbrück distribution [38]. An earlier study by Lang and Murray, designed to accurately calculate CAN1 mutation rate in culture using a large-scale fluctuation assay, detected a significant deviation of the data from a Luria-Delbrück distribution, suggesting that some can1 mutants were forming after plating the cultures on canavanine plates [25]. Raising the concentration of canavanine ten-fold (to 600 µg/ml) decreased, but did not eliminate, post-plating mutation. Using this high canavanine concentration and excluding small colonies from the calculation led to better fit of the data to a Luria-Delbrück distribution, suggesting that small colonies were largely can1 mutations that occurred after plating.
We used the same large-scale fluctuation assay and also obtained evidence for post-plating can1 mutation. Similar to Lang and Murray, we found that eliminating small colonies (Figure 1A) from the dataset improved the fit of the data to a Luria-Delbrück distribution. The frequency of large colonies exhibited a better fit to a Luria-Delbrück distribution than to a Poisson distribution, while small colonies fit a Poisson distribution better than a Luria-Delbrück distribution (Figure 1B), consistent with the conclusion that small colonies were more likely to have arisen after selection had been imposed. We performed a reconstruction experiment to investigate the possibility that small colonies were simply inherently slower growing than the large colonies (Figure 1C). Six large and six small independent can1 mutants were picked, purified and seeded into a culture of a carrier strain that harbored two copies of the CAN1 gene, and whose rate of canavanine resistance was accordingly negligible relative to the experimental strains. The carrier cultures were then plated on canavanine-containing medium. We observed that all seeded cells, whether they came from original large or small can1 colonies, formed large colonies in the reconstruction experiment (Figure 1C), ruling out any general or context-specific growth defects for cells in small colonies. In sum, the statistical modeling and the reconstruction experiment strongly supported the conclusion that the majority of small colonies were due to post-plating can1 mutations.
The only pathway known to promote “adaptive” mutagenesis in previously reported reversion assays is non-homologous end joining (NHEJ) [29]. To investigate the role of NHEJ in post-plating can1 mutation, we deleted YKU80 and observed that our yku80Δ strain exhibited an approximately two-fold reduction in the frequency of post-plating can1 mutation without any change in CAN1 mutation rate during mitotic growth (Figure 1D). We also tested the involvement of a TLS DNA polymerase, pol ζ, by deleting its subunit REV3. The rev3Δ mutant exhibited reduced CAN1 mutation frequency both in culture (two-fold) and after plating (2.75-fold). Overall, these results demonstrated that NHEJ and TLS help promote CAN1 mutagenesis during acute canavanine exposure.
We speculated that if small can1 colonies arose under different conditions than large colonies (i.e. during acute exposure to canavanine on plates), they might have been generated by different mutagenesis mechanisms and thus might have different mutation spectra. Therefore, we sequenced the can1 ORF from a large number of pre-plating (large) and post-plating (small) can1 colonies. All of the sequenced alleles contained at least one mutation in the CAN1 ORF (Table S1). The predominant classes of mutations were deletions and base pair substitutions, and the overall distribution of these broadly defined mutation classes was not significantly different between large and small colonies (Figure 2A). However, finer grained analysis of the sequence data yielded several significant results. First, a significant proportion of mutations in large can1 colonies occurred at sites previously identified as hot-spots of transcription-associated mutation (TAM) in CAN1 [39]. For example, ATΔ at position 1127 was previously identified as the most frequent mutation in actively transcribed CAN1 [39] and was, in this study, the single most frequent mutation in the large can1 colonies (Figure 2B; Table S1). This result showed that can1 mutations in large colonies were generated in cells actively transcribing CAN1, consistent with our earlier conclusion that these mutations occurred in actively growing cells in culture. In contrast, can1 mutations in small colonies did not exhibit a TAM signature, indicating that these mutations occurred under conditions of low transcriptional activity and consistent with the hypothesis that they arose after plating. Second, small colonies differed significantly from large ones (Fisher exact test P<0.002) with respect to the types of deletions in CAN1. In particular, the majority of deletions in large colonies were those of 2–5 nucleotides, while over 70% of deletions in small WT colonies were those of a single nucleotide (Figure 2C; Table S1), with 13 out of 17 such −1 deletions occurring in simple repeats (i.e. mononucleotide runs). Finally, we observed a statistically significant difference in the types of base pair substitutions in large versus small colonies (P = 0.028; Figure 2D). Together, these data confirmed our earlier conclusions that can1 mutations in small colonies arose under different conditions than those in large colonies and were generated by distinct mutagenesis mechanisms.
Our results revealed that we had recovered two classes of can1 mutants – those that arose in culture in absence of exogenous stress and those that arose in cells experiencing acute canavanine toxicity on plates. While other examples of mutagenesis in yeast during stressful conditions (including post-plating mutagenesis on canavanine [25]) have been reported, the roles of stress responses in this mutagenesis have not been examined. We addressed the role of the ESR in mutagenesis on canavanine plates by, first, asking whether canavanine activated the ESR and, second, whether attenuation of the ESR affected post-plating mutagenesis. Re-localization of Msn2 from the cytoplasm to the nucleus is a key marker of ESR activation [35]. Thus, to examine the effect of canavanine on the ESR, we examined the subcellular localization of Msn2-GFP in both CAN1 and can1 cells after plating them on canavanine medium. While Msn2-GFP was cytoplasmic in the majority of unstressed cells, it relocated to the nucleus in CAN1 cells after plating on canavanine, with about 80% of the cells showing nuclear Msn2-GFP by 12 hours after plating (Figure 3A and 3B). In contrast, Msn2-GFP in can1 cells remained cytoplasmic. This difference was not due to a general defect of CAN1 cells in responding to stress, as CAN1 and can1 cells mount a similar response to glucose starvation (Figure S1). Thus, canavanine treatment activated the ESR in yeast cells.
We next asked whether a functional ESR was required for post-plating mutation on canavanine. MSN2 and MSN4 function in a partially redundant manner, so we deleted both genes and measured the effect of the msn2Δ msn4Δ mutant (hereafter referred to as msnΔ) on CAN1 mutation in culture and on canavanine plates. Strikingly, and as predicted if ESR is necessary for post-plating mutations, the frequency of post-plating can1 mutants was reduced over 3-fold in the msnΔ strain relative to the MSN strain (Figure 3C). In contrast, CAN1 mutation rate of msnΔ cells in culture was only slightly decreased relative to that of MSN cells. To test whether reduced frequency of post-plating mutants in the msnΔ strain might be simply explained by its reduced proliferation or viability on canavanine, we measured proliferation and viability of both WT and msnΔ cells after plating them on canavanine. As expected, acute exposure to the high concentration of canavanine in the plates (600 µg/ml) was lethal to cells, but the onset of lethality was gradual, with about a third of the cells still viable after 24 hours. The msnΔ mutants were only slightly more sensitive to canavanine than WT cells (Figure 3D), consistent with previous reports that the MSN genes support acquired, rather than primary, stress resistance [3]. Also, for both WT and msnΔ strains, cell number increased by approximately 50% between 10 and 24 hrs after plating on canavanine, indicating that at least half of the cells divided during that time (Figure 3D). We also ruled out the possibility that post-plating mutant formation was simply delayed in the msnΔ strain by monitoring emergence of new can1 colonies in 72 WT and 72 msnΔ cultures for 15 days after plating. We detected no evidence that post-plating mutations simply arise later in the msnΔ strain (Figure S2). Together, these results demonstrated that the difference in post-plating can1 mutation between WT and msnΔ strains could not be attributed to differences in survival, proliferation, or a delay in mutant emergence. We conclude therefore that in the msnΔ strain the reduction in post-plating can1 mutants is due to a defect in mutagenesis.
If Msn2-Msn4 were important for generating can1 mutations on canavanine plates but not in culture, we might expect deletion of MNS2 and MSN4 to affect specifically the post-plating can1 mutation spectrum. To address this possibility, we sequenced the CAN1 ORF in pre-plating (large) and post-plating (small) can1 mutants generated in the msnΔ strain (Figure 4). The resulting mutation spectra had two similarities and two important differences compared to the WT can1 spectra. First, as in the WT strain, the overall distribution of broadly defined mutation types was not significantly different between large and small msnΔ colonies (Figure 4A). Second, as in the WT strain, only large msnΔ colonies were significantly enriched for mutations at CAN1 TAM hot-spots, such as ATΔ at position 1127 (Figure 4B). The observation of TAM hot-spots in large but not small msnΔ colonies indicated that, as in the WT strain, the majority of large msnΔ colonies were due to mutations that arose in culture, while the majority of small msnΔ colonies were due to mutations that arose after plating on canavanine. This conclusion was especially significant given that with regard to deletion and base pair substitution spectra msnΔ post-plating mutants were strikingly different from WT post-plating mutants. With respect to deletion types, in contrast to the WT strain, msnΔ small colonies showed no shift toward −1 deletions relative to large colonies (Figure 4C). Also unlike the WT strain, where small colonies showed a statistically significantly different base pair substitution spectrum from large colonies, large and small msnΔ colonies showed no difference in base pair substitution spectra (Figure 4D). Intriguingly, the unstressed base substitution spectra looked somewhat different for WT and msnΔ strains (compare large WT spectra in 2D to msnΔ spectra in Figure 4D). This difference was not statistically significant (P = 0.12) but nevertheless suggested that even in the absence of exogenous stress ESR activity may also exert subtle effects on mutagenesis. In sum, we observed that msnΔ affected specifically post-plating can1 deletion and base pair substitution spectra, indicating that the ESR controls specific mutagenesis mechanisms operating in cells exposed to canavanine.
If canavanine exerted general mutagenic effects, we should be able to detect these effects at other reporter loci, as well as test whether mutagenesis at these loci is affected by the ESR. However, the lethality of CAN1 cells in the high concentration of canavanine in plates makes it difficult to ask whether mutation at loci other than CAN1 is also induced after plating. Accordingly, we identified a low concentration of canavanine (2.5 µg/ml) that elicited a stress response, as judged by increased Msn2 nuclear localization (Figure S3A) and slower growth (Figure S3B) but did not affect cell viability (Figure 5A). Then, under conditions of chronic sub-lethal canavanine stress in culture, we measured the rate of forward mutation resulting in resistance to 5-fluoroorotic acid (FOA). In yeast, FOA resistance (FOAR) is associated with mutations in the URA3 gene; however, mutations at other loci can also cause FOA resistance in URA3 cells [25]. Indeed, we observed that mutations in URA3 accounted for only a fraction of spontaneous or canavanine-induced FOAR mutants, the rest being comprised of mutations in other, as yet unidentified loci (Figure S4). As with CAN1, rates of generation of FOAR mutations were similar in WT and msnΔ cells growing in culture in the absence of stress (Figure 5B). The low concentration of canavanine was mutagenic in culture, inducing formation of FOAR mutations by five-fold in WT cells (Figure 5B). Importantly, and as predicted by results obtained with can1, this mutagenesis depended in part on Msn2-Msn4, as canavanine induced FOAR in msnΔ by only 2.6-fold (Figure 5B). These results showed that the mutagenic effect of canavanine was not limited to can1 and could be detected at other loci, where it also depended on the function of Msn2-Msn4.
To ask whether mutagenesis by chronic, low-level canavanine exposure also depended on NHEJ and TLS, we analyzed the involvement of Ku, Rev1, and Polζ in this process by deleting YKU80, REV1, and REV3 and measuring spontaneous and canavanine-induced rates of FOAR in the mutant strains. We observed that deletion of REV1 partially reduced canavanine-induced FOAR, while deletion of REV3 or YKU80 completely abolished it (Figure 5B). Thus, in an otherwise WT background, the functions of Polζ and Ku were required for canavanine-induced mutagenesis. To begin to understand the relationships between the ESR and the DNA repair and DNA damage bypass pathways implicated in canavanine-induced mutagenesis (i.e. TLS and NHEJ), we combined deletions of MSN2 and MSN4 with rev1Δ or rev3Δ or yku80Δ and measured the rates of emergence of FOAR in the triple mutants in culture in the absence and presence of canavanine (Figure 5B). Canavanine induced FOAR mutations in the msnΔ rev1Δ mutant by about two-fold, showing an attenuation of mutagenesis relative to the msnΔ or the rev1Δ mutants (Figure 5B). This result indicated that the msnΔ and rev1Δ mutations had at least partially independent effects on canavanine-induced mutagenesis (Figure 5C). When we measured canavanine-induced mutagenesis in msnΔ rev3Δ and msnΔ yku80Δ strains, we were surprised to find that, unlike the rev3Δ or yku80Δ single mutants, the triple mutants were not defective for canavanine-induced mutagenesis, but instead were able to induce mutagenesis in canavanine at least as well as the msnΔ mutants did (Figure 5B). Thus, msnΔ was epistatic to rev3Δ and yku80Δ for canavanine-induced mutagenesis. This result suggested that in the presence of low level, chronic canavanine stress the ESR functioned upstream of Rev3 and Yku80 and also suppressed another mutagenic pathway, potentially one involving Rev1 (Figure 5C). Together, these results showed that the ESR could either promote or suppress pro-mutagenic pathways and linked the ESR and error-prone DNA repair and DNA damage bypass pathways.
Our results indicated that under conditions of canavanine stress MSN2 and MSN4 could either promote or suppress mutagenesis, depending on the genetic context (e.g. presence or absence of Polζ or Ku). Several other types of environmental stress are mutagenic in yeast, in particular osmotic and DNA replication stresses [20], [21]. Both types of stress also activate the ESR, as evidenced by characteristic gene expression signatures and/or localization of Msn2 to the nucleus [1], [2], [31], [35]. To investigate whether these types of stress-induced mutagenesis also required the function of Msn2 and Msn4, we measured the rates of emergence of resistance to canavanine or 5-FOA in cells growing in the presence of osmotic stress (1M NaCl) or replication stress (100 mM hudroxyurea [HU]). These drug concentrations retarded cellular growth and reduced viability slightly (NaCl) or moderately (HU), and their effects were similar in WT and msnΔ cells (Figure 6A). Consistent with published reports, we observed that both types of stress were mutagenic (Figure 6B). Interestingly, while 100 mM HU induced CAN1 mutagenesis very strongly (consistent with results in [21]), it had a much weaker effect on promoting FOAR mutations (Figure 6B), suggesting that replication stress has different effects on mutagenesis at different genomic loci. Importantly, we observed that neither osmotic stress-induced mutagenesis nor replication stress-induced mutagenesis depended on MSN2 and MSN4 (Figure 6B). This result showed that although various stresses can activate the ESR, its role in mutagenesis is specific to certain types of stress.
Although canavanine toxicity is well documented, the nature of canavanine-induced stress has not been examined. We hypothesized that canavanine induces proteotoxic stress due to accumulation of unfolded and nonfunctional proteins in which canavanines had replaced arginines. Indeed, we observed that Kar2p, an ER chaperone whose protein level is sensitive to levels of unfolded proteins [40], was increased in abundance in the presence of canavanine (Figure 7). This observation raised the possibility that Msn2-Msn4 function impinged specifically on mutagenesis caused by proteotoxic stress. There have been several other reports of proteotoxic stress promoting genome instability in yeast. For example, cells growing in the presence of another amino acid analog, p-Fluorophenylalanine (PFPA), showed increased rates of forward mutation at the CAN1 locus [41]. More recently, Chen et al. demonstrated that various types of stress could induce aneuploidy in yeast, but that aneuploidy was induced most strongly by proteotoxic stress (specifically by an inhibitor of HSP90, radicicol) [6]. In both of these cases, the investigators hypothesized that under conditions of proteotoxic stress, increased mutagenesis and aneuploidy were due to the action of misfolded DNA repair and chromosome segregation proteins, respectively [6], [41].
To further analyze the relationship between proteotoxic stress, the ESR, and mutagenesis, we measured forward mutation rates to canavanine or FOA resistance in WT and msnΔ strains growing in the presence of radicicol, PFPA, or tunicamycin (a drug that inhibits protein glycosylation in the ER). We chose concentrations of the drugs that retarded cell growth and only slightly (or not at all) reduced cell viability (Figure 8A). Interestingly, we observed that each of the proteotoxic agents had a distinct effect on mutagenesis, both in terms of affected loci and in terms of Msn2-Msn4 involvement. For example, treatment with radicicol did not induce mutagenesis of CAN1 (and may have even slightly reduced it) but did induce FOAR in a manner independent of MSN2 and MSN4 (Figure 8B). These results showed that although radicicol is a potent inducer of aneuploidy [6], its effects on mutagenesis were mild and locus-specific. Treatment with PFPA induced mutation of CAN1 by two-fold (consistent with data reported in [41]) in both WT and msnΔ strains (Figure 8C). Interestingly, PFPA induced formation of FOAR mutations only in the msnΔ strain, showing that under these conditions, as in the rev3Δ and yku80Δ mutants, Msn2 and Msn4 must suppress a pathway that promotes mutagenesis. Finally, tunicamycin treatment had no effect on mutation of CAN1 but induced formation of FOAR mutations by over two-fold in the WT strain but not in the msnΔ mutant (Figure 8D). This result showed that tunicamycin-caused ER stress was mutagenic and that this mutagenesis was promoted by Msn2 and Msn4. Together, these results showed that proteotoxic stress could induce genetic instability via multiple pathways and that proteotoxic stress-induced mutagenesis was regulated by the ESR.
In this study, we present evidence that transcriptional activation of the ESR in the yeast S. cerevisiae can regulate mutagenesis elicited by several types of proteotoxic stress, including two amino acid analogs, canavanine and PFPA, and a drug that interferes with protein glycosylation, tunicamycin. The effect of the ESR was specific to proteotoxic stresses, as osmotic and DNA replication stresses elicited mutagenesis that was not affected by deletion of ESR activators, MSN2 and MSN4. Moreover, Msn2 and Msn4 promoted specific types of mutation events at the CAN1 locus, including −1 deletions in simple repeats and altered types of base pair substitutions. Two TLS polymerases, Rev1 and Rev3/Polζ, and NHEJ factor Ku promoted canavanine-induced mutagenesis in culture, and deletion of MSN2 and MSN4 was epistatic to rev3Δ and yku80Δ mutants. Together these results establish a previously unknown connection between a stress response pathway and specific mutagenic DNA repair and DNA damage bypass processes and provide the first example in eukaryotes of the involvement of the general stress response in mutagenesis.
Proteotoxic stress is associated with various forms of genetic instability: radicicol is a potent inducer of aneuploidy and PFPA enhances mutagenesis of CAN1 [6], [41]. In both cases, the effect of proteotoxic stress was explained by invoking a direct role of misfolded, defective chromosome maintenance and DNA repair proteins in creating the genetic instability, without the participation of any intermediate signaling pathways [6], [41]. If this were the sole source of mutagenesis in cells experiencing proteotoxicity, then different proteotoxic agents would be expected to have similar effects. However, in this study, we observed that different types of proteotoxic stress affected different loci differently: for example, two amino acid analogs, canavanine and PFPA, have different effects on forward mutagenesis leading to FOA resistance. Furthermore, radicicol, while being a very potent inducer of aneuploidy, had relatively minor effects on mutagenesis. Even more strikingly, we observed that proteotoxic stress-induced mutagenesis was either promoted or suppressed by the ESR, arguing that proteotoxic stress induces specific signaling pathways that are regulated by Msn2-Msn4 and that culminate in the activation of mutagenic DNA repair pathways. The locus specificity may be due to the fact that mutation rates are not uniform across the genome and are influenced by local parameters such as replication timing and chromatin structure [42]–[45]; thus, proteotoxic stress and the ESR could affect at least one of these parameters. This alternative model of proteotoxic stress-induced mutagenesis is further supported by the dependency of this mutagenesis on specific DNA repair pathways, TLS and NHEJ, and by genetic interactions of the msnΔ mutant with DNA repair mutants, indicating that Msn2 and Msn4 functions affect DNA repair.
While our results implicate the ESR in mutagenesis, ESR-dependent mutation is not a universal consequence of every type of environmental stress. This result is consistent with recent evidence showing that both type and degree of stress affect the dynamics of Msn2 cytoplasmic-nuclear shuttling [46]. Precisely how different Msn2 nuclear dynamics correlate with downstream transcriptional changes is not yet well understood, but it is highly likely that different patterns of Msn2 nuclear entrance and exit may result in activation of different target genes. In a related fashion, different dynamics of p53 induction lead to different types of downstream responses: oscillating p53 levels activate cell cycle and DNA repair genes while constant p53 induction activates pro-apoptotic and pro-senescence genes [47]. The ESR target gene set contains several genes with known roles in chromatin structure and DNA repair that could potentially regulate mutagenesis, as well as over 100 genes with unknown functions. Undoubtedly, ongoing studies of Msn2 and Msn4 behavior in response to various stresses and of downstream effects on global transcript and protein levels will reveal the relevant targets of the ESR that regulate mutagenesis during specific types of stress.
Our results implicate two processes with known roles in mutagenesis – TLS and NHEJ. Deletion of REV1 resulted in a partial reduction in canavanine-induced mutagenesis that was further decreased by deletion of MSN2 and MSN4. In contrast, rev3Δ and yku80Δ were fully defective for canavanine-induced mutagenesis but msnΔ, which partially suppressed canavanine-induced mutation, was fully epistatic to these mutations. Although biochemical evidence indicates that Rev1 and Polζ can function together in TLS, with Rev1 creating a substrate for Polζ, genetic evidence has shown that their functions in vivo can be separable [48]. The different phenotypes of rev1Δ and rev3Δ mutants in canavanine-induced mutagenesis suggest that in this case Rev1 and Rev3 function in different branches of mutagenesis. A simple model consistent with our data is shown in Figure 5C. Canavanine induces mutagenesis through two branches, one of which is mediated by Msn2-Msn4 through Ku and Polζ while the other is inhibited by Msn and promoted by Rev1. Thus, mutation of REV3 or YKU80 results in inactivation of both mutagenic pathways while elimination of Msn2-Msn4 eliminates only one, even if Rev3 or Yku80 are concomitantly inactivated. No evidence currently exists for transcriptional regulation of REV1, REV3, or genes encoding the Ku complex by Msn2 and Msn4. However, our genetic data strongly suggest that the ESR regulates aspects of TLS and NHEJ. Future research will determine whether other factors in these pathways are subject to ESR regulation and/or whether this regulation may be indirect or occur at the post-translational level.
We observed that rev3Δ and yku80Δ exhibited identical phenotypes in these assays: both deletions were fully defective for canavanine-induced mutagenesis in the MSN strain but this defect was significantly rescued by msnΔ. Both Polζ and Ku have been associated with repair of DNA DSBs [30], [49], suggesting that DSBs may be an important intermediate in proteotoxic stress-induced mutagenesis. In yeast DSBs are predominantly repaired by one of two repair pathways: homologous recombination (HR) or by NHEJ [30]. HR, unlike NHEJ, uses an intact homologous sequence as template for repair and thus has been traditionally considered as the error-free DSB repair pathway. However, recent results indicate that DSB repair by HR is associated with increased mutagenesis around the DSB and that this mutagenesis is partially dependent on Polζ [50], [51]. Interestingly, DSB repair-associated mutagenesis is characterized by a distinct mutation spectrum that includes an increase in deletions in mononucleotide runs [50]. Furthermore, Lehner et al. recently reported that defects in NHEJ can also result in mononucleotide run instability [52]. Thus, increased deletions in mononucleotide runs observed during canavanine-induced mutagenesis of CAN1 are consistent with DSB involvement in this mutagenic process. Interestingly, mutagenic repair of DNA DSBs also underlies stress-induced mutagenesis in E. coli [53], [54] and may thus represent a universal mechanism of producing genetic change during environmentally unfavorable conditions.
Our study implicates the ESR in regulating DNA repair pathways in response to proteotoxic stress in a model eukaryote and as such touches on several issues with important implications for human health. First, several lines of evidence have suggested that proteotoxic stress is an important driver of emergence of drug resistance in fungal pathogens [6], [55]. Second, tumor microenvironments are characterized by a variety of stresses, such as nutrient deprivation and hypoxia, that activate the unfolded protein response in tumor cells [56], raising the possibility that unfolded protein responses are implicated in genetic instability of cancer cells. Third, proteotoxic stresses have been recently implicated in the process of aging in worms [57], although a connection between stress and increased genetic instability of aging cells has not yet been established. To develop therapeutic approaches against stress-induced genetic instability it is essential to identify cellular pathways that promote this process. In this study we have identified several factors that promote proteotoxic stress-induced mutagenesis, including Polζ, Ku, and Msn2-Msn4. Further research into this phenomenon will reveal fundamental biological principles that underlie the roles of stress responses and DNA repair pathways in stress-induced mutagenesis, and thereby enhance the development of therapeutic approaches to combat emergence of drug resistance and genetic instability during carcinogenesis.
Strains were constructed and cultured using standard yeast methods. All strains (Table S2) were derived from W303 (leu2-3,112 trp1-1 ade2-1 his3-11,15 URA3 CAN1 RAD5). The carrier strain for the reconstruction experiment (Figure 1C) contained two CAN1 copies: one at the endogenous locus and one at the rDNA locus. The rDNA::CAN1 gene was partially silenced but provided sufficient canavanine sensitivity to easily distinguish it from a fully canavanine-resistant strain that carried no wild type copies of CAN1. The strains used to analyze Msn2 loclization during canavanine stress carried endogenously expressed MSN2-GFP.
We performed the CAN1 mutation assay almost exactly as described in [25] with some minor differences: 100 µl cultures of CAN1 cells were grown at 23°C in synthetic complete medium containing 2% glucose and lacking arginine (SC-arg) in 96-well plates for 2 to 3 days (usually to a final concentration of approximately 107 cells/culture), then spotted onto SC-arg plates containing 600 µg/mL canavanine and incubated at 23°C for 5 days. 23°C was used because originally some of the experiments included a temperature-sensitive (t.s.) mutant and to allow future comparisons to other t.s. mutants. Furthermore, incubating cells at 23°C allowed us to avoid temperature fluctuations and any potential accompanying transcriptional responses that may occur when cells are transferred from room temperature to the incubator.
For analyses of pre-plating and post-plating can1 mutants, after 5 days of growth the plates were scanned and colony sizes were analyzed. To categorize can1 colonies as “large” or “small”, we used Image J software (National Institutes of Health) on scanned plate images. Image J assigned a numerical value to the area of each colony and we applied a uniform threshold to categorize them as “large” or “small”. Several different threshold values were tried and the results consistently indicated that larger colonies were more likely to have arisen in culture (showed a better fit to Luria-Delbrück) while smaller colonies were more likely to have arisen after plating (showed a poorer fit to Luria-Delbrück and a better fit to Poisson). To find the best-fitting L-D distribution for a given set of data we used the MATLAB code of Lang and Murray [25] which is based on the maximum likelihood estimation method. The code was modified appropriately to find the best fitting Poisson distributions. To calculate mutation rates in culture, large colony data were analyzed using the FALCOR tool (http://www.mitochondria.org/protocols/FALCOR.html) to calculate mutation rates and 95% confidence intervals [37]. To calculate post-plating can1 mutant frequencies, for every culture, the number of small colonies was divided by the total number of cells in the culture and these ratios were then averaged over a single experiment (72–80 cultures). For each genotype, averages and standard deviations were calculated for two to three independent experiments.
To measure survival on plates containing 600 µg/ml canavanine, at different times after plating cells were micromanipulated onto canavanine-free SC plates so that viable cells could form colonies. Three to nine biological replicates were examined for each genotype at each time point. To measure the amount of post-plating proliferation, at different times after plating agar plugs containing the entire 100 µl cultures were pulled from canavanine plates, the cells were resuspended in sterile water, sonicated, and counted using a Beckman Coulter Z2 Particle Counter. Three to six biological replicates were examined for each genotype at each time point.
The FOAR mutation assay was performed similarly to that for CAN1 with a few differences. Cells were cultured in 200 µl or 250 µl of SC-arg medium +/− indicated drug concentrations at 23°C. Because the stresses retarded cell growth, the cultures were incubated for 5 to 6 days to reach 106 to 107 cells/well, at which point they were plated on 5-FOA plates. For canavanine-containing cultures, the wells that contained pre-existing can1 mutations or can1 mutations that occurred during the growth of the culture were easily identifiable because those cultures proliferated much faster and reached saturation within two or three days. Accordingly, such cultures were deemed not to be experiencing canavanine stress and were excluded from the analyses. To confirm that the cells in the slow-growing cultures had not accumulated can1 mutations, several of these cultures were spotted onto SC-arg+600 µg/ml canavanine plates and confirmed to have only a few can1 mutants per culture. 5-FOA-resistant colony distribution data were analyzed using the FALCOR tool (http://www.mitochondria.org/protocols/FALCOR.html) to calculate mutation rates and 95% confidence intervals [37].
The URA3 locus was amplified from FOAR colonies using primers URA3-F3 (TGCCCAGTATTCTTAACCCAAC) and URA3-R1 (TGTTACTTGGTTCTGGCGAGG). Primer URA3-F3 was then used for sequencing by Macrogen USA. Analysis of the sequencing data revealed that in many cases the FOAR colonies did not carry mutations at the URA3 locus, suggesting that the FOA resistance was due to mutation of another gene (FigureS4). To verify that these colonies were indeed wild type for URA3, we performed the following phenotypic and genetic tests. We streaked 20 FOAR colonies on SC-ura medium – four containing ura3 mutations (as determined by DNA sequencing), and sixteen without detectable mutations at URA3. Consistent with the sequencing results, the four ura3 mutants were unable to grow in the absence of uracil, while the sixteen URA3 colonies were uracil prototrophs. Also, we crossed four independent URA3 FOAR colonies to a ura3-1 strain and found each of the four FOAR mutations complemented ura3-1 for FOA resistance (the diploids were FOAS) and segregated independently from ura3-1 in the cross. Thus, we concluded that in many cases FOA resistance was not due to mutation of URA3. We are currently investigating the identity of the non-ura3 FOAR mutations.
We examined Msn2-GFP localization using a wide-field inverted microscope (Deltavision; Applied Precision, LLC) with a charge-coupled device camera (CoolSNAP HQ; Roper Scientific), using a 100× oil-immersion objective, at 25°C and a FITC filter set to detect GFP fluorescence (Chroma, Brattleboro, VT). The transmittance was set at 10%, and the exposure time for Msn2-GFP was 200 ms, except when analyzing low fluorescence conditions (e.g. no GFP and the estradiol-inducible Msn2-GFP in the absence of estradiol) when exposure time was increased to 250 ms. To analyze whether canavanine activated the ESR, cells were taken from canavanine plates at different times after plating, resuspended in water on microscope slides, and immediately analyzed by fluorescent microscopy. To subject cells to glucose starvation, cells from the SC-arg cultures were briefly centrifuged and resuspended in SC-arg medium without glucose, incubated for one hour, then analyzed by fluorescent microscopy. Four to six z-stacks of every field were taken and projected into one image using the average pixel intensity method.
To measure the viability of cells in growing culture in the absence or presence of stress agents, the corresponding cultures were briefly sonicated, appropriately diluted in SC-arg medium, and plated onto YPD plates. Cell concentrations in the original cultures were obtained by using a Beckman Coulter Z2 Particle Counter.
Yeast cells were grown to mid-log phase in SC-arg medium at 23°C, then canavanine was added to a concentration of 100 µg/ml or cells were collected by filtration and transferred to SC-arg medium lacking glucose. After 6 hours, approximately 6×107 cells were collected by filtration and snap-frozen at −80°C. Protein lysates were prepared form the cell pellets as described in [58]. Briefly, cells were lysed in 20% TCA using glass beads and the beads were washed twice in 5% TCA. The lysates were centrifuged, pellets resuspended in Laemmli buffer, and their pH neutralized by 2M unbuffered Tris. The resulting protein lysates were separated using 12% SDS-PAGE and probed using antibodies against Kar2 [40] and an anti-MYC antibody (Clontech) to detect Msh2-13xMyc.
DNA sequencing of the CAN1 ORF was performed by Macrogen USA using primers CAN1-R1 (TGAGAATGCGAAATGGCGTG) and CAN1-R2 (TTTTGATGGCTCTTGGAACG). Statistical analyses of mutational spectra were performed and all the Fisher Exact Test (FET) p-values calculated using R open software (www.r-project.org).
|
10.1371/journal.pcbi.1002476 | On Conduction in a Bacterial Sodium Channel | Voltage-gated Na+-channels are transmembrane proteins that are responsible for the fast depolarizing phase of the action potential in nerve and muscular cells. Selective permeability of Na+ over Ca2+ or K+ ions is essential for the biological function of Na+-channels. After the emergence of the first high-resolution structure of a Na+-channel, an anionic coordination site was proposed to confer Na+ selectivity through partial dehydration of Na+ via its direct interaction with conserved glutamate side chains. By combining molecular dynamics simulations and free-energy calculations, a low-energy permeation pathway for Na+ ion translocation through the selectivity filter of the recently determined crystal structure of a prokaryotic sodium channel from Arcobacter butzleri is characterised. The picture that emerges is that of a pore preferentially occupied by two ions, which can switch between different configurations by crossing low free-energy barriers. In contrast to K+-channels, the movements of the ions appear to be weakly coupled in Na+-channels. When the free-energy maps for Na+ and K+ ions are compared, a selective site is characterised in the narrowest region of the filter, where a hydrated Na+ ion, and not a hydrated K+ ion, is energetically stable.
| Ion channels are integral membrane proteins that control the passive diffusion of ions down their electrochemical gradient. According to the most permeating ion species, ion channels are classified in three categories: K+-channels, Na+-channels, and Ca2+-channels. The atomic structure of a K+-channel was the first to be solved experimentally more than 10 years ago. This structure inspired numerous computational studies, which revealed the mechanisms of conduction and selectivity in K+-channels. Recently, the first atomic structure of a Na+ selective channel has been solved. Here, molecular dynamics simulations and free-energy calculations are described and a possible mechanism for Na+ conduction is identified. In contrast to what it is observed in K+-channels, ion movements through Na+-channels appeared highly uncorrelated.
| Sodium channels allow the passive diffusion of Na+ ions down their electrochemical gradient. They were first discovered, together with potassium selective channels, in nerve fibres where they mediate the fast depolarizing phase of action potentials [1]. Na+-channels are also involved in the initiation of action potentials in cardiac myocytes and in general, in the propagation of electrical impulses in cardiac, muscle and nerve cells. The cell membrane is exposed to a high-Na+/low-K+ concentration on the extracellular side, and to a low-Na+/high-K+ concentration on the intracellular side. It is the selectivity of Na+ and K+ channels, in the presence of these concentration gradients what makes possible the control of the membrane potential. Several atomic structures of K+-selective channels have been solved since 1998 [2], [3], [4], [5], [6]. Recently, the atomic structure of a Na+-selective channel, NavAb from the bacterium Arcobacter butzleri, has also been determined [7], and provides the opportunity to investigate how selectivity and conduction are realized at the atomic level in Na+-channels.
Na+ and K+-channels share the same general architecture, with four transmembrane domains contributing to a central pore. This central pore, where ion permeation occurs, is delimited by two transmembrane helices, S5 and S6, linked by an intervening loop. The Na+-channel NavAb was crystallized in the closed state, where the four S6 helices are arranged in a conical shape, defining a bundle-crossing at the intracellular side and a water-filled cavity above it. In the open state, the S6 helices are thought to bend, and the water-filled cavity becomes a continuous with the intracellular solution, as observed in K+-channel [8], [9]. The region responsible for selective permeation is located in the loop between S5 and S6. In K+-channels, each chain contributes to the selectivity filter with the signature sequence T/S-x-G-Y/F-G. The oxygen atoms of the residues from this signature sequence define a series of four binding sites where dehydrated K+ ions bind [2]. The selectivity filter of the bacterial Na+-channel NavAb is much wider than that of known K+-channels (Figure 1). Although Na+ ions were not observed in the crystallographic structure, the presence of three Na+ binding sites was hypothesized [7]. Residues Thr175 and Leu176 define two rings of carbonyl oxygen atoms at the intracellular entrance of the filter. Water molecules from the hydration shell of a Na+ ion may interact through hydrogen bonds with these two layers of carbonyl oxygen atoms, thus defining two binding sites for hydrated Na+ ions. Following the convention introduced by Payandeh et al [7], these binding sites will be referred as SIN and SCEN. A third Na+ binding site, SHFS, composed by side chains of four glutamate residues, Glu177, from each of the four protein chains was also proposed [7]. The side chain of Glu177 points to the lumen of the filter, and they could contribute to the hydration shell of cations.
Molecular Dynamics (MD) simulations were crucial for understanding the mechanisms of conduction and selectivity in K+-channels. The knock-on mechanism, first hypothesized in the fifties by Hodgkin and Keynes [10], was described at atomic detail by computational studies [11], [12], and free-energy calculations revealed energy barriers of the order of 2–4 kcal/mol for the concerted motion of three K+ ions through the filter of K+-channels [13], [14], [15]. These free-energy barriers along a multi-ion permeation pathway were shown to increase if one of the permeating ions was Na+ instead of K+, offering some insight about selectivity in K+-channels [16], [17], [18]. The recent crystal structure of a prokaryotic voltage gated sodium channel NavAb from Arcobacter butzleri in combination with all-atom MD simulations can also provide important information concerning conduction and selectivity in Na+-channels. Here, simulation strategies similar to the ones used for K+-channels were adopted to analyze conduction and selectivity in the NavAb Na+-channel, with the purpose to get insights into: (i) the number of Na+ binding sites, (ii) the nature of the conduction mechanism, and (iii) selectivity.
Root mean square deviation of the protein backbone atoms below 1.4 Å, and 1.0 Å in the case of the backbone atoms of the filter, residue Thr175 to Ser178, were observed in a 40-ns MD trajectory, which suggests that the system is stable. The selectivity filter was initially occupied by a single Na+ ion in SCEN, which remained at this position for the entire trajectory. A second Na+ ion approached the filter from the extracellular side during the course of the simulation, and it stayed close to SHFS for more than 10 ns. The intracellular cavity below the selectivity filter was initially occupied by 15 water molecules. The number of water molecules inside the cavity reached equilibrium value of ∼45 in the course of the simulation, due to the incorporation of water molecules from the extracellular solution across the selectivity filter. This contrasts with the situation in K+-channels, where the movement of water molecules through the selectivity filter was highly unlikely [19]. Similar observations where described by Klein et al [20].
MD simulations in the nanosecond time scale cannot reveal the mechanisms of conduction and selectivity in Na+-channels alone. In order to get further insights into the permeation process, the permeation free-energy profiles for Na+ and K+ ions in the Na+-channel were calculated by the umbrella sampling technique (see the Methods section for the computational details). The free-energy profile of a single Na+ ion moving from the intracellular cavity to the extracellular solution displays two minima with similar energies (Figure 2). In the innermost minimum, the ion lies in the plane defined by the carbonyl oxygen atoms of residues Leu176, binding site SCEN. Figure S1 shows the positions of the permeating ion in a plane perpendicular to the permeation axis. At SCEN, a Na+ ion preserves its entire hydration shell, i.e. it is coordinated only by oxygen atoms from the surrounding water molecules. The nature of the atoms of the coordination shell of a Na+ ion at different positions along the permeation pore is shown in Figure S2. In the region above SCEN, the permeating ion may move ∼2 Å away from the pore axis, where it interacts with a carbonyl oxygen atom of Leu176 from one of the chains. This position does not correspond to a minimum in the free-energy profile. Proceeding toward the extracellular side, a local free-energy minimum is encountered with the ion close to the pore axis, in a region 1–2 Å below the oxygen atoms of the side chain of Glu177. At this position, the Na+ ion is fully solvated by water molecules. This local minimum is followed by the second global free-energy minimum, where the ion occupies SHFS. The ion moves away from the centre of the pore axis (Figure S1) and two of the water molecules from the hydration shell are lost, and substituted by oxygen atoms from the side chain of Glu177 and Ser178 (Figure S2).
The free-energy profile of a single K+ ion permeating the selectivity filter of the Na+-channel is markedly different from the free-energy profile of single Na+ ion (Figure 2). In the most stable minimum, a K+ ion occupies a position similar to the one occupied by Na+ in its innermost minimum. As observed with Na+, a K+ ion at this position preserves intact its hydration shell. The second minimum is analogous to the outermost minimum encountered by Na+, with oxygen atoms from the side chains of Glu177 and Ser178 contributing to the coordination shell of the ion. In contrast to Na+, no local minimum is present in the region between the carbonyl oxygen atoms of Leu176 and the side chain oxygen atoms of Glu177 in the permeation pathway of K+. Overall, the composition of the coordination shell and the position of the ion in a plane perpendicular to the pore axis are analogous for K+ and Na+ ions (Figure S1 and S2). Likewise, a free-energy barrier higher than 8 kcal/mol hampers the movement of both Na+ and K+ ions from SHFS to the extracellular solution. These high free-energy barriers associated with the exit of Na+ and K+ from the selectivity filter may originate from the binding affinity for cations of this region of the pore, where the carboxyl groups of the side chain of Glu177 line the pore on the extracellular side. Therefore, under these circumstances, to enhance conduction, it seems possible that the selectivity filter is occupied on average by one or more positive ions, and that conduction occurs when an incoming ion displaces the ion(s) already inside the pore, like described in K+-channels [14]. To test this hypothesis, free-energy maps for conduction events with two Na+ ions were calculated (Figure 3). It was found that in the most stable configuration, both SCEN and SHFS are occupied by Na+ ions. Starting from this position and moving the innermost ion toward the intracellular side, a local minimum is observed with the bottom ion inside the cavity and the top ion in SHFS. The free energy of this local minimum is 4.3±0.4 kcal/mol higher than the free-energy of the minimum with ions in SCEN and SHFS. The differences in free energy and the energy barriers were evaluated along the minimum energy path that connects the local energy minima, and the errors were estimated dividing the trajectories into three data sets (see the Methods section for further details). A second free-energy minimum is observed with both ions close to Glu177 and Ser178. The free energy of this configuration is 1.2±0.6 kcal/mol higher than the free energy of the minimum with ions in SCEN and SHFS. The free-energy barriers between the two basins are 3.5±0.5 kcal/mol and 2.4±0.3 kcal/mol respectively in the outward and inward directions. Regardless of the position of the bottom ion (SCEN or SHFS), the movement between the extracellular solution and the pore of the outermost ion does not encounter any free-energy barriers higher than 3 kcal/mol.
In order to understand if the low free-energy barriers for ion translocation across the pore were exclusive of Na+ ions, free-energy maps for conduction events with two K+ ions were also calculated (Figure 4). Like in the case of Na+ ions, the lowest free-energy configuration has two ions occupying binding sites SCEN and SHFS, but in contrast to a pore occupied by Na+ ions, this free-energy minimum is much broader. A local minimum with both ions close to Glu117 and Ser178 can be also described. The free energy of this configuration is 2.7±0.5 kcal/mol higher than the free energy of the lowest minimum, and the barriers between the two basins are 5.0±0.4 kcal/mol and 2.3±0.5 kcal/mol respectively in the outward and inward directions. The inward movement of the ion at SCEN toward the intracellular cavity while the second ion is at SHFS causes an energy increase of 1.8±0.2 kcal/mol. The change in free energy is lower than 1 kcal/mol if the ion in SHFS leaves the filter toward the extracellular solution and the barriers for the entry/exit of K+ in SIN and SHFS are below 2 kcal/mol.
Under physiological conditions, the cellular membrane is exposed to high-Na+/low-K+ concentration on the extracellular side, and to a low-Na+/high-K+ concentration on the intracellular side. As a consequence, the selectivity filter of Na+-channels is more likely to be approached by Na+ ions on the extracellular side and by K+ ions on the intracellular side. To characterize this situation, the free-energy map for a mixture of K+ and Na+ ions in the selectivity filter was calculated, with the K+ ion in the innermost position and the Na+ ion in the outermost position (Figure 5). The free-energy profile characterising the outward movement of K+ through the selectivity filter is similar to that when both ions are K+. In the lowest free-energy configuration, the Na+ ion is in site SHFS, while the K+ ion can move in a wide region around SCEN, experiencing low-energy barriers (<2 kcal/mol). A second local minimum exists with both ions close to Glu177 and Ser178 (2.6±0.8 kcal/mol higher in energy). The barriers between the two minima are similar to those observed in the case of two K+ ions (4.6±1.1 in the outward direction, 2.0±0.5 in the inward direction).
Since the number and position of Na+ binding sites in the NavAb channel were only speculated using the crystallographic structure, insight from atomistic simulations is essential. Two Na+ binding sites have been described in the selectivity filter in simulations with either one or two permeating ions. These binding positions correspond to sites SCEN and SHFS, which were hypothesized experimentally. Computations confirm that a Na+ ion remains close to the centre of the pore axis and it is fully hydrated in the innermost binding site. In contrast, it occupies a position slightly off the centre of the pore axis, while it interacts directly with the side chain of Glu177 in the outermost site. The side chain oxygen atom of Ser178 also contributes to the coordination shell of a Na+ ion in SHFS, which was not an obvious experimental observation. The experimentally speculated binding site SIN does not appear as a well defined minimum in the free-energy maps from atomistic simulations. A configuration with two ions in SCEN and SHFS is found to be the most stable during conduction events with two permeating ions. Surprisingly, the energetic cost of having two Na+ ions in close proximity of binding site SHFS is extremely low (∼1 kcal/mol). Therefore, the picture that emerges from the free-energy profiles is that of a pore preferentially occupied by two ions, which can switch between different configurations by crossing low energy-barriers. In contrast to K+-channels, the movements of the ions appear to be weakly coupled.
The nature of the conduction mechanism for K+ ions in this Na+-channel was also tested, in order to understand selectivity. In the case of Na+, the strongest impediment to conduction is the raise in free energy associated with the inward movement of an ion from SCEN to the intracellular cavity. Similarly, the free energy also increases when a K+ ion moves from SCEN to the intracellular cavity, but to a lower extent (∼2 kcal/mol for K+; ∼4 kcal/mol for Na+). The free energy of an ion in the cavity is likely to be affected by the conformation of the intracellular gate in NavAb, which is closed. And this effect may be different for K+ and Na+, as already observed in K+-channels [21].
The biological function of Na+-channels is to allow Na+ ions inside the cell at nearly the rate of free diffusion, while at the same time blocking the flux of K+ ions in the outward direction. During permeation in the selectivity filter, the highest free-energy barrier that a Na+ ion experiences in the inward direction is ∼2 kcal/mol (independently of the nature of the innermost ion, Na+ or K+), while the highest energy barrier for the outward movement of K+ ions is ∼5 kcal/mol (independently of the nature of the outermost ion, Na+ or K+). Considering that discrimination of ions is performed with an efficiency of one K+ ion every 10–100 Na+ ions [7], [22], the differences in free-energy barriers of 1–2 kcal/mol described before are perfectly in line with the experimental observations. The highest barrier for the passage of ions through the filter is localized between SCEN and SHFS, both for K+ and Na+. In the case of Na+, a local free-energy minimum exists in this region, with the ion fully hydrated and aligned at the centre of the channel axis. In contrast, the same configuration is not a free-energy minimum for K+. This difference is already evident in the free-energy profiles with a single permeating ion. The instability of a hydrated K+ ion in the middle of the selectivity filter is a possible cause of the preference of the channel for Na+ permeation. The shortcomings of the adopted computational methods should be clearly kept in mind when interpreting the present results. The convergence of the PMF calculations in complex molecular systems is already a well reported issue [23], [24]. Here, the magnitude of the errors of the PMF was calculated by comparing different parts of the simulated trajectories. This approach is reliable if the simulated trajectories are extensive enough to sample the relevant ion configurations in the selectivity filter. However, the possibility that structural changes of the protein on a longer time scale may affect the calculated energy profiles cannot be ruled out. In contrast to K+-channels, the movements of the ions appear to be weakly coupled in Na+-channels, and the ions maintain their hydration shell. Therefore, the issue of the absolute binding energy values of the ions [11], that is, the energy of the ions in the protein relative to their energy in water should be less crucial than in the case of K+-channels. In the later, the water molecules of the ion's hydration shell are replaced with the carbonyl oxygen atoms of the protein. A particular critical point is the presence of ionisable residues in the close proximity of the permeation pathway (Glu177). The ionization state of these residues may have an important effect on the permeation properties, and need to be closely analyzed. In this respect, it is also important to remember that a non-polarisable force field has been used and polarization effects may be important in the present context [25], [26], [27]. Such calculations will be reported in the future.
Another crucial point in this type of studies that was clearly pointed out by Warshel et al. [28] is that ‘the issue of ion selectivity cannot be resolved quantitatively without calculating the corresponding ion current, which is the actual direct observable of the system’. As they explained in their paper ‘such a calculation should be able to convert the free energy profile of the system to corresponding time dependence of the ion permeation process’. In order to do this the present study should be extended, and the free energy obtained here should be coupled to Brownian dynamics simulations as those reported by Warshel et al. [28]. In such a way, under certain conditions that allow replicating experimental data and using the calculated free energy profiles as starting point, a detailed understanding of ion selectivity may then be possible.
Atomic coordinates of the NavAb channel were taken from the Protein Data Bank entry 3RVY [7]. Only residues 116 to 221 that correspond to the pore region of the channel were included in the model. Default protonation states were used for all the ionisable residues. N- and C-terminals were amidated and acetylated respectively. The channel was centred in the x-y plane and the permeation pathway was aligned with the z-axis. The aromatic belt defined by the amphipathic residues Trp195 was aligned with the upper layer of a pre-equilibrated bilayer of DOPC molecules. All lipid molecules closer than 2.0 Å to the protein atoms were removed. The system was solvated by ∼20.000 water molecules, and 32 Na+ ions and 24 Cl− ions were added. A Na+ ion was positioned in site SCEN. Harmonic restraints were initially applied to the backbone atoms and to the Na+ ion in the selectivity filter, and gradually removed during a one nanosecond period. The total production run was 40 ns.
MD trajectories were simulated with the version 2.7 of NAMD [29], using the CHARMM27 force field with CMAP corrections [30], and the TIP3P model for water molecules [31]. Standard parameters for K+ and Na+ in CHARMM27 force field were adopted. Simulations were performed in the NpT ensamble. Pressure was kept at 1atm by the Nose-Hoover Langevin piston method [32], [33], with a damping time constant of 100 ps and a period of 200 ps. Temperature was kept at 300 K by coupling to a Langevin thermostat, with a damping coefficient of 5 ps−1 [33]. Electrostatic interactions were treated by the Particle Mesh Ewald algorithm, with grid spacing below 1 Å [34]. Van der Waals interactions were truncated at 12 Å, and smoothed at 10 Å. Hydrogen atoms were restrained by the SETTLE algorithm [35], which allowed a 2 fs time-step.
Free-energy profiles for one and two ion conduction events were calculated by umbrella sampling [36]. The reaction coordinate for one ion conduction events was the distance along the z-axis between the permeating ion and the centre of the carbonyl oxygen atoms of Thr175. The same reaction coordinate was used for the bottom ion in the analyses of conduction events considering two ions. The reaction coordinate for the upper ion was the distance along the z-axis between the ion and the centre of the carbonyl oxygen atoms of Leu176. Harmonic potentials (force constant 10 kcal*mol−1*Å−2) were applied to the reaction coordinates. Each umbrella sampling simulation consisted of 0.5 ns, and the first 50 ps were considered as equilibration period and discarded. 20 and 270 umbrella sampling simulations were performed respectively for conduction events considering one or two ions, moving the centres of the harmonic potentials in 1.0 Å steps. The starting configurations for the umbrella sampling simulations were defined using the final snapshot of the normal MD trajectory. The water oxygen atoms or the Na+ ions closer to the centres of the reaction coordinates were selected. Then, the positions of the ions subjected to the harmonic potentials were switched with the positions of these molecules/ions closer to the centre of the harmonic potentials. The ions restrained in the umbrella sampling simulations were the ion already inside the selectivity filter in the normal MD simulation, plus an ion randomly taken from the extracellular solution in the case of two-dimensional umbrella sampling simulations. For simulations with K+ ions, all the Na+ ions were alchemically transformed to K+ ions in the first snapshot, while for the calculation of the energy map for Na+/K+ mixtures, only one ion (the innermost of the two restrained ions) was transformed to K+. Free-energy profiles were calculated by the Weighted Histogram Analysis Method [37]. The string method [38] was used to calculate the minimum energy path in free-energy profiles of conduction events with two ions (Figure S3). The initial guess for the minimum energy path was the segment that connects the free-energy minimum with the innermost ion in the cavity, and the free-energy minimum with the outermost ion in the extracellular solution (200 equally spaced points were used for the discretization). The path evolved in the direction opposite to the free-energy gradient until the root mean square distance between two successive iterations was below 10−3 Å. The relative energies of the local minima and the energy barriers between different free-energy basins provided in the text are evaluated along this minimum energy path. In order to estimate the errors affecting the free-energy values, the umbrella sampling trajectories were divided into 3 separate sets of 150 ps each. Free-energy profiles and minimum energy paths were computed separately for each data set. The values reported in the text are the average and the standard deviation among these 3 data sets. The free-energy profiles and the minimum energy paths shown in Figures 2, 3, 4, 5 and S3 were calculated using the whole umbrella sampling trajectories after the equilibration period.
Note Added In Proof. While this paper was in revision similar results were reported by B. Corry and M. Thomas in J. Am. Chem. Soc., 2012, 134 (3), pp 1840–1846.
|
10.1371/journal.pntd.0001379 | Detection of Echinococcus multilocularis in Carnivores in Razavi Khorasan Province, Iran Using Mitochondrial DNA | Echinococcus multilocularis is the source of alveolar echinococcosis, a potentially fatal zoonotic disease. This investigation assessed the presence of E. multilocularis infection in definitive hosts in the Chenaran region of Razavi Khorasan Province, northeastern Iran.
Fecal samples from 77 domestic and stray dogs and 14 wild carnivores were examined using the flotation/sieving method followed by multiplex PCR of mitochondrial genes. The intestinal scraping technique (IST) and the sedimentation and counting technique (SCT) revealed adult Echinococcus in the intestines of five of 10 jackals and of the single wolf examined. Three jackals were infected only with E. multilocularis but two, and the wolf, were infected with both E. multilocularis and E. granulosus. Multiplex PCR revealed E. multilocularis, E. granulosus, and Taenia spp. in 19, 24, and 28 fecal samples, respectively. Echinococcus multilocularis infection was detected in the feces of all wild carnivores sampled including nine jackals, three foxes, one wolf, one hyena, and five dogs (6.5%). Echinococcus granulosus was found in the fecal samples of 16.9% of dogs, 66.7% of jackals, and all of the foxes, the wolf, and the hyena. The feces of 16 (21.8%) dogs, 7 of 9 (77.8%) jackals, and all three foxes, one wolf and one hyena were infected with Taenia spp.
The prevalence of E. multilocularis in wild carnivores of rural areas of the Chenaran region is high, indicating that the life cycle is being maintained in northeastern Iran with the red fox, jackal, wolf, hyena, and dog as definitive hosts.
| Echinococcus multilocularis causes alveolar echinococcosis, a serious zoonotic disease present in many areas of the world. The parasite is maintained in nature through a life cycle in which adult worms in the intestine of carnivores transmit infection to small mammals, predominantly rodents, via eggs in the feces. Humans may accidentally ingest eggs of E. multilocularis through contact with the definitive host or by direct ingestion of contaminated water or foods, causing development of a multivesicular cyst in the viscera, especially liver and lung. We found adult E. multilocularis in the intestine and/or eggs in feces of all wild carnivores examined and in some stray and domestic dogs in villages of Chenaran region, northeastern Iran. The life cycle of E. multilocularis is being maintained in this area by wild carnivores, and the local population and visitors are at risk of infection with alveolar echinococcosis. Intensive health initiatives for control of the parasite and diagnosis of this potentially fatal disease in humans, in this area of Iran, are needed.
| Echinococcus multilocularis is the agent of alveolar echinococcosis, a potentially fatal zoonotic disease [1], [2]. The life cycle of E. multilocularis is sylvatic; adult worms are found in wild carnivores, principally foxes, and in the raccoon dog, wolf, coyote, and jackal, while their metacestodes develop in small mammals, predominantly rodents such as Cricetidae, Arvicolidae, and Muridae [3], [4], [5], [6], [7]. In some rural areas domestic dogs and sometimes cats can be definitive hosts after acquiring the infection from wild rodents, and thus become a major zoonotic risk for infecting humans [3], [8], [9]. Humans can serve as an aberrant intermediate host for E. multilocularis, with transmission occurring through direct contact with the definitive host or by ingestion of contaminated water, vegetables, or other foods [10]. Human alveolar echinococcosis is a lethal zoonotic disease caused by infection with the multivesiculated metacestode of E. multilocularis [11], [12].
The geographical distribution of the parasite is restricted to the northern hemisphere. The cestode has been reported in areas of central Europe, the Near East, Russia, central Asian republics, northern Japan, and Alaska [11], [13], [14], [15]. In the Middle East, cystic echinococcosis is prevalent in most countries, although a low prevalence of alveolar echinococcosis is reported in Iran, Iraq, and Tunisia [16]. In Asia, canine infections have been recorded in dogs in China, Kazakhstan and Kyrgyzstan [17], [18], [19]. Echinococcus multilocularis in canids was reported in northwestern Iran for the first time in 1971 [20], [21]. Further investigation in 1992 found its infection in 22.9% of red foxes (Vulpes vulpes) and 16% of jackals (Canis aureus) [22]. The latest research in 2009 reported no evidence of E. multilocularis infection in canids of the Moghan Plain in northwest Iran [23]. The majority of previous research focused only on the northwestern part of the country [20], [21], [22], [23]. Alveolar echinococcosis, based on histopathological and clinical data, was first reported in a village in Chenaran County of Razavi Khorasan Province in 2007 [24]. The disease was subsequently confirmed by molecular evaluation, and a second case reported (E. Razmjou, unpublished data). Razavi Khorasan Province is located in northeastern Iran, near the border with Turkmenistan, where E. multilocularis is endemic. A pilot study revealed that suitable hosts such as foxes, jackals, dogs, wolves, and rodents are frequently present near villages in the mountains of Chenaran region. The presence in Razavi Khorasan of suitable conditions for completing the life cycle of this parasite, such as presence of definitive and intermediate hosts in mountainous areas and proximity to other countries where the parasite is endemic, led to the present study to assess the prevalence of E. multilocularis infection in carnivores, and to identify natural definitive hosts of this life-threatening parasite in the Chenaran region of Razavi Khorasan Province, Iran.
Animals were shot under license from the Iran Environment Protection Organization, solely for the purpose of investigating the presence of Echinococcus multilocularis in wild carnivores. The Protocol of this investigation was reviewed and approved by the Ethics Committee of Tehran University of Medical Sciences.
The study area, the Chenaran region, covers approximately 2400 km2 in northeastern Iran, 55 kilometers northwest of Mashhad (Figure 1) (36°4′N, 59°7′E). It lies between the Binalood Heights and the Hezar Masjed Mountains. Chenaran city is surrounded by rural areas, mainly consisting of human habitations, gardens, farms, and moorland. The region has cold and snowy winters and mild summers. The average annual temperature is 13.4°C with variable rainfall; mean annual precipitation of 212.6 mm. It is rich in wildlife, including carnivores and small rodents appropriate for supporting the life cycle of E. multilocularis.
From November 2009 to January 2010, fecal samples from 77domestic and stray dogs from 17 villages and the entire gut of three foxes, ten jackals, and one wolf (Canis lupus pallipes), either shot or killed accidentally, were collected. In addition, during October and November 2010, the intestines of one fox and one hyena (Hyena hyena) killed on roads were added to our samples. A standard form including place and date of killing was completed. Fecal samples were collected from the rectum of each wild carnivore. The intestine and fecal samples were placed in labeled ziploc bags, stored at −80°C for at least seven days [25] to reduce the risk of laboratory infection by inactivating any Echinococcus oncospheres and other infective materials, and subsequently stored at −20°C until further examination.
The intestinal scraping technique was performed as described by Deplazes and Eckert [25]. The intestine was opened full length and after removal of undigested food and visible parasites from the proximal, middle, and posterior parts of the small and large intestine, 15 deep mucosal scrapings were taken using microscope slides. Material adhering to the slides was transferred to plastic Petri dishes and examined stereomicroscopically at 120× magnification. Echinococcus worms were isolated and stored in 85% ethanol for molecular examination and in 10% formalin for morphological diagnosis.
The sedimentation and counting technique was done as previously described [26]. The intestine was cut into 10 cm pieces and each was placed in a flask containing one liter of 0.9% saline. After vigorous shaking for a few seconds, the pieces of intestine were pressed firmly between the fingers using gloves and with care to avoid contamination with eggs to remove attached worms. The supernatant was decanted and the procedure was repeated several times with physiological saline. The sediment was placed in plastic Petri dishes and examined stereomicroscopically at 120× magnification. All isolated worms, including Echinococcus, were stored in 85% ethanol and 10% formalin for molecular and microscopic identification, respectively.
Fecal samples were submitted to flotation with zinc chloride for isolating parasite eggs. Each sample (4–5 g) was stirred into 50 ml distilled water until completely dispersed. The suspension was passed through four layers of gauze and large particles removed. The suspension was transferred into a 50 ml Falcon tube and centrifuged at 1000×g for 5 min. For isolating eggs, zinc chloride solution (specific gravity 1.45 g ml−1) was added to sediment up to a final volume of 12 ml and, after complete mixing, centrifuged at 1000×g for 30 min [27]. The supernatant was passed through sequential sieves on 50 ml falcon tubes with metal and polystyrene screens of mesh sizes 37 and 20 µm, respectively [27]. The sieves were inverted and washed thoroughly with distilled water containing 0.2% Tween 20. After adding phosphate-buffered saline (PBS; pH 7.2) to a final volume of 50 ml, suspensions were centrifuged at 1000×g for 30 min, the supernatant fraction was aspirated, and sediment (approximately 400 µl) was transferred to 1.5 ml tubes and stored at −20°C until further examination.
Detection of Echinococcus spp. was based on morphological characteristics. Adult worms of E. multilocularis and E. granulosus were differentiated using morphological characteristics including size, length of gravid proglottids, shape of the uterus, number of eggs per proglottid, and position of genital pore after acetic acid alum carmine staining and mounting in Canada balsam.
DNA of adult worms and taeniid eggs was extracted using the QIAamp DNA Mini kit (Qiagen, Germany) according to the protocol of Verweij et al. [28] with slight modifications. Briefly, one Echinococcus worm was removed from 85% ethanol and washed in sterile PBS buffer three times. The worm was then placed in 200 µl of PBS buffer and, after 10 min boiling at 100°C, an equal volume of ATL buffer plus 10% proteinase K was added and completely mixed and incubated two hours at 55°C in heat block. DNA extraction was continued according to manufacturer's instructions with the minor modification of increasing incubation time to five minutes to increase the yield of DNA in the final step. DNA was stored at −20°C until analysis. Before submitting the eggs from fecal samples to the DNA extraction procedure described for adult worms, they were subjected to seven freeze/thaw cycles, using liquid nitrogen and boiling water, to disrupt the egg wall. Then, 200 µl of the sample was heated at 100°C for 10 min as in Verweij's procedure [28]. The concentration of extracted DNA was measured spectrophotometrically by Biophotometer (Biophotometer Plus, Eppendorf, Germany).
Multiplex PCR of adult worms and eggs was performed as described [29]. The mitochondrial multiplex reaction was designed to amplify a 395 bp fragment of NADH dehydrogenase subunit 1 (nad1) of E. multilocularis and 117 bp and 267 bp of a small subunit of ribosomal RNA (rrnS) of E. granulosus and other Taenia spp., respectively. Primers, conditions, and parameters for PCR were as previously described [29]. All samples were tested in 25 µl amplification reaction mixtures with 12.5 µl of the master mix (QIAGEN Multiplex PCR, Germany), 2.5 µl of primers (2 µM of primers Cest1, Cest2, Cest3, Cest4 and 16 µM of primer Cest5 in H2O), 8 µl H2O, and 2 µl of template DNA. Initially, multiplex PCR was confirmed with standard DNA of E. multilocularis, E. granulosus, Taenia multiceps, and T. hydatigena provided by Professor Deplazes of the Institute of Parasitology of Zurich, Switzerland. Additionally, for all PCR reactions one negative control without DNA and one positive control with standard DNA were included to confirm the results of multiplex PCR. Finally, 10 µl of the PCR products were loaded on 2% (W/V) agarose gels, and stained with ethidium bromide to visualize by electrophoresis.
The results of multiplex PCR were confirmed by single PCR using the primer pair Cest1/Cest2 and Cest4/Cest5 for E. multilocularis and E. granulosus, respectively [29], and EM-H15/EM-H17 [9] for E. multilocularis.
All E. multilocularis PCR-positive samples were confirmed by sequencing of a 395 bp amplified fragment. PCR products were excised from agarose gels and purified using the QIAquick Gel Extraction Kit (QIAgen, Germany), according to the manufacturer's instructions. Products were sequenced in both directions using the Cest1/Cest2 primers by MilleGen Company (France). Sequences were read by CHROMAS (Technelysium Pty Ltd., Queensland, Australia) and aligned using the DNASIS MAX (version 2.09; Hitachi, Yokohama, Japan) software program.
The entire small intestine of 16 wild carnivores comprising 10 jackals, four foxes, one hyena, and one wolf were examined by both IST and SCT. These techniques found 6 of 16 (37.5%; 95% CI: 18.5%–61.4%) wild canids to be infected with Echinococcus spp. The worms were isolated from five of 10 (50%, 95% CI: 23.7–76.3%) jackals and one wolf, while the remaining jackals, the foxes, and the hyena tested negative. The intensity of infection was classified as low (1–100), medium (101–1000), or high (>1000) worm burden [30]. All positive jackals showed a high Echinococcus worm burden, while the wolf had a low burden.
We differentiated all Echinococcus worms by microscopic examination (Figure 2). Among six Echinococcus positive samples, three jackals (30%) had a single species infection with E. multilocularis but two jackals (20%) and the wolf were infected with both E. multilocularis and E. granulosus.
To detect eggs, fecal samples of dogs were investigated by direct microscopic examination and the flotation method. Eggs were observed in 13 of 77 (16.9%, 95% CI: 10.1–26.8%) dog fecal samples.
The result of multiplex PCR by amplification of 395 bp fragment of nad1 indicated that 19 of the carnivores were infected with E. multilocularis. The 117 bp fragment of rrnS identified E. granulosus in 24, and the 267 bp fragment found Taenia spp. in 28 of the fecal samples. Echinococcus multilocularis infection was detected in the feces of all wild carnivores (100%; 95% CI: 78.5–100%) including nine jackals, three foxes, one wolf and one hyena, and five dogs (6.5%; 95% CI: 2.8–14.3%). Echinococcus granulosus was found in the fecal sample of 16.9% (95% CI: 10.1–26.8%) of dogs, 66.7% (95% CI: 35.4–88.0%) of jackals, and all of the foxes, the wolf, and the hyena. The feces of 16 of 77 (21.8%; 95% CI: 13.2–31.1%) dogs, 7 of 9 (77.8%; 95% CI: 45.3–94.0%) jackals, and all three foxes, one wolf and one hyena were infected with Taenia spp. (Table 1) (Figures 3, 4).
Among 26 PCR positive dog samples, a single DNA fragment amplified in 18 samples indicated two E. multilocularis (2.6%; 95% CI: 0.7–9.0%), six E. granulosus (7.8%; 95% CI: 3.6–16.0%), and ten Taenia spp.(13%; 95% CI: 7.2–22.3%) infected dogs. Two species-specific fragments in seven cases showed five dogs (6.5%; 95% CI: 2.8–14.3%) to be co-infected with E. granulosus and Taenia spp., and two E. multilocularis infected dogs were also infected with Taenia spp. (1.3%; 95% CI: 0.2–7.0%) or E. granulosus (1.3%; 95% CI: 0.2–7.0%). Three amplicons revealed one dog (1.3%; 95% CI: 0.2–7.0%) infected simultaneously with Taenia spp. and two species of Echinococcus. Of 14 E. multilocularis infections in wild carnivorous, 11 showed triple infections including six jackals (66.7%), three foxes (100%), the hyena and the wolf. One jackal was co-infected with Taenia spp., and two jackals were infected with E. multilocularis (Table 2; Figure 4). The prevalence of E. multilocularis infection was high in wild carnivores (100%; 95% CI: 78.5–100%), whereas the rate of infection in domestic and stray dogs was low (6.5%; 95% CI: 2.8–14.3%). The results of multiplex PCR were confirmed with single PCR. Sequencing results of amplicons obtained from worms and eggs were identified as E. multilocularis. The nucleotide sequences of the amplified nad1 were equivalent to positions 7645 to 8040 of the published E. multilocularis mitochondrion complete genome (accession no. AB018440). Sequences were aligned using DNASIS MAX (version 2.09; Hitachi, Yokohama, Japan) with the published reference sequences. Analysis revealed 100% identity between our isolates and the corresponding published reference sequences for E. multilocularis. The nucleotide sequence data reported in this paper will appear in the DDBJ/EMBL/GenBank nucleotide sequence databases with the accession numbers AB617846–AB617855 and AB621793–AB621801.
Iran is located in the Middle East and central Eurasia. It is bordered on the north by Armenia, Azerbaijan, Turkmenistan and Caspian Sea. Afghanistan and Pakistan are Iran's direct neighbors in the east. The country borders the Persian Gulf and the Gulf of Oman in the south, Iraq in the west, and Turkey to the north-west (Figure 1).
In the Middle East, echinococcosis is one of the most important zoonotic diseases [16]. Cystic echinococcosis (CE) and alveolar echinococcosis (AE) have been reported in the Mediterranean region, but CE is more prevalent [31], [32]. Although Iran is an endemic area for echinococcosis, most studies have been carried out only on E. granulosus. Mobedi and Sadighian in 1971 reported E. multilocularis for the first time in three of 30 red foxes tested from northwest Ardebil Province [20], [21]. A study of E. multilocularis in carnivores of that area was followed in 1992 by Zariffard and Massoud [22] showing comparable results, infection in 22.9% (16/70) of red foxes and 16% (4/25) of the jackals. In a study in 2009 Zare-Bidaki et al. did not observe E. multilocularis in the investigated canids [23]. With the exception of these three studies, E. multilocularis infection has not been investigated outside of the northwest part of the country. Although there is limited information about AE in Iran, Torgerson et al. [10] suggested that Iran, since it is bordered by highly endemic countries, is an endemic area for E. multilocularis, and that the estimated annual incidence of eleven cases of AE are likely underreported. Molecular confirmation of some AE reports from inhabitants of a village in Chenaran County of northern Razavi Khorasan Province [24] (E. Razmjou, unpublished data), which neighbors a hyperendemic country, Turkmenistan, led to our investigation of the establishment of E. multilocularis' life cycle in this area.
Assessment of the occurrence of E. multilocularis in definitive hosts showed that this cestode has a high prevalence in the wild carnivores of the Chenaran area in northeastern Iran (100%; 95% CI: 78.5–100%). In comparison to the prevalence of infection in foxes in Belgium (24.55%) [33], Switzerland (47%–67%) [26], Ukraine (36%) [34], Kyrgyzstan (64%) [35], and Japan (49%) [36], it is assumed that Razavi Khorasan Province is hyperendemic for this tapeworm.
The significantly lower prevalence of E. multilocularis infection in dogs (6.5%) than in other carnivores (100%) may be due to decreased ingestion of metacestode infected intermediate hosts through controlled and limited diet of domestic dogs. As the multiplex PCR is subject to amplification of DNA from taeniid eggs, the results could also be associated with overlooking some positive cases due to prepatent infections, the intermittent egg excretion during the patent period [29], or degradation of taeniid eggs in the environment and unsuitable conditions such as solar radiation [37].
The rate of infected dogs in our study (6.5%) was less than reported in China (12%) [38] and Kyrgyzstan (18%) [19]. On the other hand, the infection rate observed in our study was considerably greater than prevalence reported in Germany (0.24%) [4] and Lithuania (0.8) [39]. The low rates found there may have been a consequence of conducting PCR only on positive taeniid egg samples by the sieving/flotation method. Our experiment showed Taenia eggs in only 16.9% of the 77 dog feces by sieving/flotation and microscopic examination. This increased to 33.9% with multiplex PCR on the flotation material. It assumed that results might be related to difficulty in detecting small numbers of eggs by microscopic examination, while DNA of a single taeniid egg from a sieved fecal sample can be amplified by the multiplex PCR [29].
While the prevalence of infection in dogs is low, the large population of domestic and stray dogs in the villages that are in close contact with inhabitants must be considered a potential source and risk factor for transmission of E. multilocularis to humans. The results of multiplex PCR in the current study showed that most of the wild (78.6%) and some domestic (2.6%) canids were co-infected with E. granulosus and E. multilocularis. Previous studies have reported a high rate of CE in livestock and humans in this province [32]. Consequently, Razavi Khorasan should be considered an endemic area for both E. granulosus and E. multilocularis infection. Razavi Khorasan Province is a tourist area, and many travelers are at the risk of exposure to these zoonotic diseases. The province has extensive agriculture and export of fruits to other parts of the country. For these reasons and because of close contact of humans with infected domestic dogs and other definitive hosts that forage for food on farms and gardens of this region, it is important to initiate intensive health initiatives.
In conclusion, our study confirms the presence of the life cycle of E. multilocularis in the Chenaran region of northeastern Iran. In this cycle the red fox, jackal, wolf, hyena, and dog play the role of the definitive host, and efforts are underway to elucidate the intermediate host of this parasite.
Since the current survey is the first evidence of existence of E. multilocularis in domestic and wild animals in northeastern Iran, further studies should be conducted to investigate the presence of E. multilocularis in other parts of the country.
|
10.1371/journal.pntd.0007590 | Expression of Bacillus thuringiensis toxin Cyt2Ba in the entomopathogenic fungus Beauveria bassiana increases its virulence towards Aedes mosquitoes | The entomopathogenic fungus Beauveria bassiana has been widely used to kill mosquito larvae and adults in the laboratory and field. However, its slow action of killing has hampered its widespread application. In our study, the B. bassiana fungus was genetically modified to express the Bacillus thuringiensis (Bt) toxin Cyt2Ba to improve its efficacy in killing mosquitoes.
The efficacy of the wild type (WT) of B. bassiana and a transgenic strain expressing Cyt2Ba toxin (Bb-Cyt2Ba) was evaluated against larval and adult Aedes mosquitoes (Aedes aegypti and Aedes albopictus) using insect bioassays. The Bb-Cyt2Ba displayed increased virulence against larval and adult Aedes mosquitoes compared with the WT: for Ae. aegypti adults, the median lethal time (LT50) was decreased by 33% at the concentration of 1× 108 conidia/ml, 19% at 1× 107 conidia/ml and 47% at 1× 106 conidia/ml. The LT50 for Ae. albopictus adults was reduced by 20%, 23% and 29% at the same concentrations, respectively. The LT50 for Ae. aegypti larvae was decreased by 42% at 1× 107 conidia/ml and 25% at 1× 106 conidia/ml, and that for Ae. albopictus larvae was reduced by 33% and 31% at the same concentrations, respectively. In addition, infection with Bb-Cyt2Ba resulted in a dramatic reduction in the fecundity of Aedes mosquitoes.
In conclusion, our study demonstrated that the virulence of B. bassiana against mosquitoes can be significantly improved by introducing the Bt toxin gene Cyt2Ba into the genome to express the exogenous toxin in the fungus. The transgenic strain Bb-Cyt2Ba significantly reduced the survival and fecundity of Ae. aegypti and Ae. albopictus compared with the WT strain, which suggested that this recombinant B. bassiana has great potential for use in mosquito control.
| Mosquito vectors transmit many diseases to humans and animals, causing illness and death and resulting in substantial socio-economic burdens in endemic countries. The control of mosquitoes has almost exclusively relied on the use of chemical insecticides, which has recently led to the broad resistance of important mosquito vectors worldwide. Entomopathogenic fungi, such as Beauveria bassiana, are an important alternative or complement to chemical insecticides. However, the relatively slow action of fungal pathogens in killing mosquitoes, compared with chemical insecticides, has hampered their widespread application. To improve the insecticidal efficacy of the entomopathogen B. bassiana, the fungus was genetically modified to express the Bacillus thuringiensis toxin Cyt2Ba. The mitotically stable transformant (Bb-Cyt2Ba) successfully expressed Cyt2Ba toxin, and the virulence of this strain against adults and larvae of Aedes aegypti and Aedes albopictus mosquitoes was significantly improved. In addition, egg laying was significantly affected by Bb-Cyt2Ba infection. Infection with this fungus resulted in a dramatic reduction in fecundity of the target mosquitoes. Therefore, this recombinant B. bassiana has great potential for use in mosquito control.
| Mosquito vectors transmit many diseases to humans and animals, causing illness and death that result in considerable socio-economic burdens in endemic countries [1]. Aedes mosquitoes (primarily Aedes aegypti and Aedes albopictus) are the primary vectors of dengue, Zika, chikungunya and yellow fever in tropical and subtropical zones, which have a devastating impact on human health [2, 3]. Vector control via chemical insecticides is a major method for vector-borne disease control, but the extensive use of chemical insecticides poses toxicity risks to humans as well as the environment and creates intensive pressure for mosquitoes to develop resistance [4].
Biological control agents such as entomopathogenic fungi are important alternatives or complements to chemical insecticides [5]. Many studies have shown the potential of entomopathogenic fungi, such as Beauveria bassiana, for the control of agricultural pests [6] and the vectors of human diseases, including mosquitoes [7]. One of the main advantages of using entomopathogenic fungi in insect control is that these fungi can infect all stages of the insects, including larvae and adults [8–10]. Furthermore, no cases of resistance of insects to entomopathogenic fungal infections have been reported to date [11]. However, the relatively slow action of fungal pathogens in killing vectors, compared with chemical insecticides, has hampered their widespread application in the field.
Genetic engineering is essential for introducing desirable traits into entomopathogenic fungi [12], which has resulted in a wide range of feasible types of genetic manipulation, from the expression of UV protectants, heat shock factors, immune modulators, cuticle-degrading enzymes and exogenous toxins to targeting insect vectors, disease transmission, and even arthropod behaviors [12]. It has been reported that transgenic B. bassiana expressing Ae. aegypti trypsin modulating oostatic factor (TMOF) exhibited increased virulence against Anopheles gambiae compared with the wild type strain [13]. Our previous study also found that the expression of the insect-specific toxin scorpion neurotoxin AaIT in B. bassiana enhanced virulence against Ae. albopictus mosquitoes [14]. The virulence of B. bassiana for mosquitoes could be increased via the expression of insecticidal toxins.
Bacillus thuringiensis (Bt), an entomopathogenic bacterium, is also an important bioinsecticide used for control of insects, including mosquitoes [15]. Based on the last updated data in the especial database for Bt toxins (June 2018), about 323 holotype toxins have been identified and characterized in the Bt strains isolated from different regions of the world [16]. The cytolytic δ-endotoxin (Cyt2Ba) containing 263 amino acids is found at very low concentrations in Bt crystals [17]. It has been reported that Cyt2Ba is toxic to Culex, Aedes and Anopheles larvae [18, 19]. In our study, the Cyt2Ba gene was introduced into the B. bassiana genome by genetic engineering to improve fungal virulence. We measured the virulence of this recombinant B. bassiana to the adults and larvae of Aedes mosquitoes (Ae. aegypti and Ae. albopictus). The effect of this strain on mosquito fecundity was also determined.
Our only experimental animals are Aedes aegypti and Aedes albopictus, which do not involve animal ethical issues.
The Guangdong Provincial Center for Disease Control and Prevention collected Ae. albopictus and Ae. aegypti mosquitoes from different sites in the cities of Foshan and Zhanjiang in Guangdong Province. All mosquitoes were reared in standard insectary conditions at (28 ± 1) °C and (80 ± 5) % relative humidity with a light:dark cycle of 16 h:8 h. The larvae were fed daily with turtle food (Shenzhen INCH-GOLD Fish Food,.LTD, Shenzhen, CHA). Mosquito adults were provided with 10% glucose solution ad libitum. All collection was done on public land.
The B. bassiana GIM3.428 strain (wild type) was purchased from the Guangdong Microbiology Culture Center and maintained on Czapek’s agar (CDA) at 4°C for preservation. Conidia were obtained by growing the fungus on CDA for 7 days at 25°C. Blastospores were obtained by the growth of the fungus in Sabouraud dextrose broth for 48 h and glucose-mineral medium for 24 h at 25°C under 120 rpm on a rotatory shaker.
The coding sequence of Cyt2Ba (GenBank: GQ919041.1) was synthesized (Generay Biotech, Shanghai, CHA) with the B. bassiana preferred coding usage and cloned between the BamHI and EcoRI sites of pBARGPE1 to generate the plasmid pBARGPE1-Cyt2Ba. This plasmid retains a strong gpdA promoter to drive the insert’s gene expression and has the phosphinothricin (PPT) resistance gene (Bar) as a selectable marker for fungal transformation.
The plasmid pBARGPE1-Cyt2Ba was linearized with ScaI and then transformed into B.bassiana blastospores by electroporation as described previously [14]. Transformants were grown on CDA plates containing 150 μg/ml PPT at 25°C. After single spore isolation and subsequent subculturing for three generations on CDA with 150 μg/ml PPT at 25°C (7 days each generation), the putative transformants were verified by polymerase chain reaction (PCR) using primers for the bar gene (Bar-F, TCGTCAACCACTACATCGAGAC and Bar-R, GAAGTCCAGCTGCCAGAAAC) and Cyt2Ba gene (Cyt-F TATGGATCCGCCACCATGGAAC, and Cyt-R, TATGAATTCCTAGGACTTGATGGG). Furthermore, to verify the mitotic stability of their PPT resistance, the recombinants were subcultured for three generations on CDA without PPT at 25°C, and finally, they were subcultured on CDA with 400 μg/ml PPT at 25°C. Then, the transformants were analyzed by PCR. A stable transformant named Bb-Cyt2Ba was selected for the subsequent experiments.
The wild type (WT) and Bb-Cyt2Ba strains were grown on CDA for 4 days. The conidia were collected with cotton swabs. Mosquitoes were infected by contact with WT or Bb-Cyt2Ba conidia on the swab. Dead female adult Ae. albopictus mosquitoes were maintained at 25°C under saturated humidity for 4 days. To verify the transcription of the Cyt2Ba gene, total RNA was extracted from the CDA culture supernatant or infected mosquitoes by using an RNeasy mini plant kit (Qiagen, Duesseldorf, GER). Reverse-transcription PCR (RT-PCR) was performed using the primer pair Cyt-F and Cyt-R.
The expression of Cyt2Ba was detected by western blot analyses. Cyt2Ba polyclonal antibodies were raised in rabbits (BGI Genomics, Shenzhen, CHA), and an alkaline phosphatase-conjugated anti-rabbit IgG secondary antibody (Santa Cruz Biotech, Newport, USA) was used for detection. The total proteins were extracted from the CDA culture supernatant and the infected mosquitoes as previously described (14). Fifty micrograms of each sample was separated on a 12% polyacrylamide gel by sodium dodecyl sulfate (SDS) polyacrylamide gel electrophoresis, and western blotting was then performed.
B. bassiana strains were grown and maintained on CDA at 25°C. Conidial suspensions in 0.02% (vol/vol) Tween 80 were prepared from 7-d-old cultures. To conduct fungal infection, adult female Aedes mosquitoes were transferred to the filter (placed on a 300 ml plastic cup covered with a net), which had absorbed 10 ml of conidial suspension at 1 × 106 conidia/ml (low), 1 × 107 conidia/ml (middle) or 1 × 108 conidia/ml (high) or 0.02% Tween 80 (control) for 3 h. Then, the mosquitoes were transferred separately to different plastic containers (30 mosquitoes/cup). The contaminated or control mosquitoes were maintained on a 10% glucose diet in plastic containers at 28°C. The dead adult mosquitoes in each treatment group were counted and removed every 12 h until the last mosquito died. The dead mosquitoes were washed twice in phosphate-buffered saline and placed in humid filter papers for conidiation. To infect the larvae, conidial suspensions of 1 × 105 conidia/ml (low), 1 × 106 conidia/ml (middle) and 1 × 107 conidia/ml (high) of B. bassiana were added to the cups containing 30 second-instar Aedes larvae in 30 ml of double-distilled water (ddH2O). Thirty individuals in ddH2O without fungi were used as the control. The larvae were provided with turtle food at a rate of 0.2–0.3 mg/larva per day and maintained at 28°C. The larvae were monitored every 12 h for survival (the percentage surviving) until day 10. Each treatment for the larvae or adults contained 90 (30 × 3) mosquitoes. The survival of the adults or larvae in each group was calculated every day. The bioassay was repeated three times for statistical analysis.
To examine the effect of fungal infection on the fecundity of mosquitoes, Aedes females were infected by contact with the filter paper that had absorbed 10 ml of the conidial suspension with 1 × 108 conidia/ml of the wild type or Bb-Cyt2Ba strain. After fungal infection, the females were fed with defibrinated sheep blood for 1 h. Then, the engorged mosquitoes were placed individually into a 250-ml, gauze-covered paper cup containing a small amount of water and funnel-shaped filter paper to serve as a repository for the eggs. The eggs of each mosquito were counted one week after blood feeding. Sixty female mosquitoes per treatment were used to assess the impact of fungal infection on fecundity.
For survival assays, data were analyzed using Probit and Kaplan–Meier survival tests; a log-rank test was used to assess the equality of survival distributions during the Kaplan–Meier analysis. Furthermore, cross-tabulation analyses using Pearson’s chi-square test were applied to compare the percentages of ovipositing females among the different groups. In addition, Mann-Whitney tests were applied to compare the fecundity among the groups. All analyses were performed with SPSS (v. 20.0), and significance was defined by P<0.05.
Transformation of the competent blastopores of B. bassiana with linearized plasmid pBARGPE1-Cyt2Ba produced one transgenic colony on the CDA plate containing 150 μg/ml PPT. The transformant was able to grow on a CDA plate containing 150 μg/ml PPT for three generations. Then, after three rounds of subculturing on PPT-free CDA plates, the transformant was capable of growing on the CDA plate containing 400 μg/ml PPT. The expected PCR fragments from all the Bb-Cyt2Ba samples appeared on the agarose gel, but not the WT (S1 Fig), which confirmed the consistent heredity of the Cyt2Ba gene in the Bb-Cyt2Ba genome. Transformation of the pBARGPE1-Cyt2Ba construct into B. bassiana did not affect the germination rate of the fungus after 24 h, and we did not find any morphological differences between the germ tubes of WT and Bb-Cyt2Ba.
The WT and Bb-Cyt2Ba strains were harvested from the CDA culture after incubation for 4 days at 25°C. The fungal outgrowths were those of a typical B. bassiana, and they appeared on the dead female Aedes mosquitoes that were infected through cuticle penetration after being maintained at 25°C in saturated humidity for 4 days. RT-PCR and western blot analyses confirmed that the recombinant strain expressed the Cyt2Ba toxin stably, but not the WT strain (Fig 1).
The mosquito mortalities generally increased with the elevated conidial concentration and prolonged post-treatment time for both Bb-Cyt2Ba and Bb-WT (Figs 2 and 3). Significant differences were found among the groups treated with different concentrations of Bb-Cyt2Ba or the Bb-WT (S1 Table). In addition, significant differences were found between the mortalities of Bb-Cyt2Ba and the WT-treated Aedes adults at each concentration (S2 Table). Thus, the adult mosquitoes that were treated with Bb-Cyt2Ba tended to die faster than those treated with the WT at the same concentration (Fig 2), and the difference was significant (S2 Table). Similarly, except for Ae. aegypti larvae infected with fungi at the low concentration, Bb-Cyt2Ba-treated larvae tended to die faster than those treated with the WT at each concentration (Fig 3), and the difference was significant (S2 Table).
The survival curves of Ae. aegypti (A) and Ae. albopictus (B) females when treated with (C1) 1 × 105, (C2) 1 × 106 and (C3) 1 × 107 conidia/ml suspensions of Bb-Cyt2Ba and WT. Each treatment contained 90 larvae, performed three times. The LT50s of the female Ae. aegypti for Bb-Cyt2Ba treatment were reduced by 33%, 19% and 47% compared with those for WT treatment at the high, middle and low concentrations, respectively (Table 1). Moreover, the LT50s of Ae. albopictus adults for Bb-Cyt2Ba treatment were reduced by 20%, 23% and 29% at these three concentrations compared with those for WT treatment. In addition, compared with the WT, the LT50 of the Bb-Cyt2Ba infection was reduced by 42% and 25% for Ae. aegypti larvae and 33% and 31% for Ae. albopictus larvae at the high and middle concentrations, respectively (Table 1). These results indicated that the expressed Cyt2Ba toxin in the fungus reduced the survival of the tested larvae and adult mosquitoes.
Egg laying was significantly affected by fungal infection. The percentage of Bb-Cyt2Ba-infected Ae. aegypti females that oviposited within 7 days after a blood meal was only 45% (27/60), which was significantly lower compared to that of the WT (65%, 39/60) and the control (95%, 57/60) (S3 Table). Similarly, the percentage of Ae. albopictus mosquitoes producing eggs was 50% (30/60), which was significantly lower than that of the WT (70%, 42/60) and the control (97%, 58/60) (S3 Table).
The mean number of eggs produced per female within 7 days for Ae. aegypti was 33.9 (95% CI: 29.2–38.6) in the Bb-Cyt2Ba-treated group, 47.8 (95% CI: 42.5–53.1) in the WT-treated group and 76.4 (95% CI: 68.4–84.4) in the control group (Fig 4A). For Ae. albopictus mosquitoes, the mean number of eggs produced per female was 37.3 (95% CI: 31.8–42.8) in the Bb-Cyt2Ba-treated group, 53.3 (95% CI: 48.2–58.4) in the WT-treated group and 72.6 (95% CI: 64.3–81.0) in the control group (Fig 4B). Thus, infection of mosquitoes with the wild type strain resulted in a significant reduction (37% for Ae. aegypti and 27% for Ae. albopictus) in fecundity compared to noninfected controls. However, expression of Cyt2Ba toxin in the fungi resulted in a dramatic reduction (56% for Ae. aegypti, 49% for Ae. albopictus) in fecundity compared to the controls.
The slow action of insect-pathogenic fungi on target pests is one of the limitations to their commercialization and large-scale application as biocontrol agents [4]. Some studies have shown that the insect-killing efficacy of B. bassiana is considerably improved by recombination with the exogenous Ae. aegypti TMOF gene, scorpion neurotoxin AaIT gene [8, 20], cuticle degradation protease PR1A gene [21], or vegetative insecticidal proteins of Bacillus thuringiensis (Bt) [22].
Cytolytic (Cyt) toxins are produced by a minor group of Bt, mostly in subspecies that are toxic to Dipteran insects [23, 24]. These toxins do not need a specific receptor but directly interact with membrane lipids and insert into the membrane to form pores [25–27] or destroy the membrane by a detergent-like interaction [28]. In our study, the cytolytic toxin Cyt2Ba gene was genetically introduced into the genome of the wild type of B. bassiana to enhance its virulence against mosquitoes. The mitotically stable transformant Bb-Cyt2Ba can successfully express this exogenous toxin, and the virulence of Bb-Cyt2Ba for Ae. aegypti and Ae. albopictus mosquitoes (including adults and larvae) was significantly increased compared with that of the WT.
When adult mosquitoes were infected by the fungus through cuticle contact, the fungus quickly invaded the mosquito tissues and cells. It has been reported that approximately 24 h was required for the fungus to penetrate the cuticle and reach the hemocoel of insects [29, 30]. In our study, the fungus may have released Cyt2Ba toxin into the mosquito hemocoel after penetration, which resulted in quicker death for the Aedes adult mosquitoes. Furthermore, it has been reported that when fungal spores are applied to an aquatic habitat, typical for mosquito larvae, the nutrients in the water are usually sufficient to stimulate the germination of the spores following water intake [31, 32]. For mosquito larvae, the main infection routes are through feeding and respiration [33]. The Cyt2Ba toxin is supposed to be delivered to the insect circulatory system after Bb-Cyt2Ba infects the mosquito larvae, which results in faster death of the larvae compared with the WT infection. The various commercial products of Bt have been used in the control of mosquito larvae. Formulations include a variety of granules, flowable concentrates, wettable powders, and slow-release tablets and briquettes [34]. The efficacy of Bt formulations has been demonstrated in a diversity of habitats against a multitude of species of mosquitoes [34]. Like Bt, the Bb-Cyt2Ba strain will increase its efficiency in mosquito larvae control with a suitable formulation, which is helpful to the development of Bb-Cyt2Ba strain as a commercial product.
In addition, the effects of Bb-Cyt2Ba and WT infection on mosquito fecundity were also assessed in our study. Bb-Cyt2Ba infection resulted in a dramatic reduction in the fecundity of target mosquitoes. Despite the LT50 of Bb-Cyt2Ba-infected mosquitoes suggesting that death was not quick enough to completely inhibit the females from going through their gonotrophic cycles, fecundity reduction might be a byproduct of rapid fungal invasion of Bb-Cyt2Ba, which is hypothesized to be an adaptive strategy of the host [35]. These data suggested that strain Bb-Cyt2Ba has the potential to reduce the size of Aedes mosquito populations by severely compromising fecundity.
Compared to our previous study, the virulence of Bb-Cyt2Ba to Ae. albopictus mosquitoes is a little bit higher than that of the B. bassiana expressing scorpion neurotoxin AaIT toxin [14]. Due to their high specificity against insects and low toxicity to vertebrates and plants of these proteins [36–38], AaIT and Cyt2Ba genes are both good choices for introduction into the genome of B. bassiana to increase the fungal pathogenicity.
Entomopathogenic fungi show considerable potential to be developed as biopesticides [39–41]. The production and application of fungi both involve relatively simple infrastructures and processes [40, 42], which can be readily adopted in mosquito-borne disease-endemic countries. However, before gene-modified fungi can be used and integrated into control programs, more data on the environmental safety, their effect on nontarget insect hosts and the possibility of gene flow are required.
In conclusion, our data showed that expression of the Bacillus thuringiensis toxin Cyt2Ba in B. bassiana increased its effectiveness against two important mosquito vectors, Ae. aegypti and Ae. albopictus. The median lethal times were shorter in the mosquito adult and larval groups infected with the Bb-Cyt2Ba strain compared with the groups infected with the wild type strain. In addition, the fecundity of the females was dramatically reduced by Bb-Cyt2Ba infection compared with WT infection. This recombinant B. bassiana strain Bb-Cyt2Ba was valuable in development as a bioinsecticide for mosquito control and even for other types of pest control.
|
10.1371/journal.pgen.1006453 | Investigating Conservation of the Cell-Cycle-Regulated Transcriptional Program in the Fungal Pathogen, Cryptococcus neoformans | The pathogenic yeast Cryptococcus neoformans causes fungal meningitis in immune-compromised patients. Cell proliferation in the budding yeast form is required for C. neoformans to infect human hosts, and virulence factors such as capsule formation and melanin production are affected by cell-cycle perturbation. Thus, understanding cell-cycle regulation is critical for a full understanding of virulence factors for disease. Our group and others have demonstrated that a large fraction of genes in Saccharomyces cerevisiae is expressed periodically during the cell cycle, and that proper regulation of this transcriptional program is important for proper cell division. Despite the evolutionary divergence of the two budding yeasts, we found that a similar percentage of all genes (~20%) is periodically expressed during the cell cycle in both yeasts. However, the temporal ordering of periodic expression has diverged for some orthologous cell-cycle genes, especially those related to bud emergence and bud growth. Genes regulating DNA replication and mitosis exhibited a conserved ordering in both yeasts, suggesting that essential cell-cycle processes are conserved in periodicity and in timing of expression (i.e. duplication before division). In S. cerevisiae cells, we have proposed that an interconnected network of periodic transcription factors (TFs) controls the bulk of the cell-cycle transcriptional program. We found that temporal ordering of orthologous network TFs was not always maintained; however, the TF network topology at cell-cycle commitment appears to be conserved in C. neoformans. During the C. neoformans cell cycle, DNA replication genes, mitosis genes, and 40 genes involved in virulence are periodically expressed. Future work toward understanding the gene regulatory network that controls cell-cycle genes is critical for developing novel antifungals to inhibit pathogen proliferation.
| The opportunistic fungal pathogen Cryptococcus neoformans infects immune-compromised humans and causes fungal meningitis by proliferating in the central nervous system. The cell cycle has not been studied at the whole transcriptome level in C. neoformans. Here, we present the expression dynamics of all genes from a synchronous population of C. neoformans cells over multiple cell cycles. Our study shows that almost 20% of all C. neoformans genes are periodically expressed during the cell cycle. We also compare the program of cell-cycle-regulated transcription in C. neoformans to the well-studied but evolutionary distant yeast, Saccharomyces cerevisiae. We find that many orthologous cell-cycle genes are highly conserved in expression pattern (e.g. DNA replication and mitosis genes), while others, notably budding genes, have diverged in expression ordering. We also identify 40 virulence genes from previous studies that are periodically expressed during the C. neoformans cell cycle in rich media. Our findings indicate that a conserved set of fungal transcription factors (TFs) controls the expression of conserved cell-cycle genes, while other periodic transcripts are likely controlled by species-specific TFs.
| About 500 million years of evolution separate the fungal phyla Ascomycota and Basidiomycota [1,2]. The cell cycle is an essential biological process driving cell division of these distantly related yeasts, and therefore may be under strong selective pressure for conservation. Both Saccharomyces cerevisiae (Ascomycota) and Cryptococcus neoformans (Basidiomycota) can grow and divide asymmetrically in a budding yeast form. C. neoformans is a causative agent of deadly fungal meningitis, primarily in immune-compromised patients [3,4]. Many groups studying C. neoformans focus on virulence factors for human infection, such as the yeast’s polysaccharide capsule, melanin production, Titan cell formation, and others [5–9]. We propose that the function of cell-cycle regulators, which are essential for proliferation in the host, merit further investigation as virulence factors. Furthermore, there is evidence that virulence pathways are perturbed when cell-cycle progression is slowed, which suggests direct connections between cell-cycle regulators and virulence pathways [10,11].
The cell cycle is the process by which a cell duplicates its contents and faithfully divides into two genetically identical cells. In eukaryotes, a biochemical oscillator drives sequential cell-cycle events, where the cyclin-dependent kinase (CDK) and its variety of cyclin binding partners initiate events by phosphorylation, followed by destruction of kinase activity in mitosis by the anaphase-promoting complex (APC). Another common feature of the eukaryotic cell cycle is a temporally regulated program of transcription, which has been demonstrated in S. cerevisiae, Schizosaccharomyces pombe, Arabidopsis thaliana, mouse fibroblasts, and human tissue culture cells [12–22]. These programs of periodic genes include cyclin mRNAs, DNA replication factors, APC activators, and other cellular components that are utilized at specific times during the cell cycle. Our group and others have proposed that this “just-in-time transcription” mechanism is an important aspect of energy-efficient and faithful cell divisions [23,24]. In S. cerevisiae, an interconnected network of periodic transcription factors (TFs) is capable of driving the periodic program of cell-cycle gene expression [15,25–27]. Aspects of this yeast TF network are conserved in human cells; for example, G2/M genes are activated by a periodic forkhead domain-containing TF in both eukaryotes [22,28]. The topology of cell-cycle entry is also functionally conserved, where a repressor (S.c. WHI5, H.s. RB1) is removed by G1 cyclin/CDK phosphorylation to activate a G1/S transcription factor complex (S.c. SBF/MBF, H.s. E2F-TFDP1) [29]. However, the genes involved in cell-cycle entry are not conserved at the sequence level between fungi and mammals [30], suggesting that the fungal pathway could be targeted with drugs without affecting mammalian host cells.
Sequence-specific DNA-binding TFs have been identified in C. neoformans and phenotypically profiled by single gene knockouts [6,31,32]. This TF deletion collection was profiled over many virulence factor-inducing conditions to discover pathways that regulate disease and drug response genes [32]. Serial activation of TFs during capsule production has also been studied to elucidate the order in which TFs control virulence gene products [31]. However, the cell cycle has not been investigated in synchronous populations of cells to date. Although the phenotypes of some single mutant cell-cycle TFs have been examined from asynchronous populations, these studies offer limited understanding of temporal aspects of gene expression during the cell cycle.
Here we investigate transcriptional dynamics of the pathogenic yeast C. neoformans using cells synchronized in the cell cycle. We compare our findings to the cell-cycle transcriptional program in S. cerevisiae. We find that a similar percentage of all genes (~20%) are periodically transcribed during the cell cycle, and we present a comprehensive periodicity analysis for all expressed genes in both yeasts. We show that S-phase gene orthologs are highly conserved and temporally precede M-phase gene orthologs in both yeasts. Additionally, we find that many TFs in the cell-cycle entry pathway are conserved in sequence homology, periodicity, and timing of expression in C. neoformans, while others, notably genes involved in budding, are not. We also identify 40 virulence genes that appear to be cell-cycle-regulated, along with nearly 100 orthologous fungal genes that are periodic in the same cell-cycle phase. Taken together, these cell-cycle genes represent candidates for further study and for novel antifungal drug development.
Identifying approaches for synchronizing populations of C. neoformans has been challenging. We succeeded in synchronizing by centrifugal elutriation, a method that has been very successful for S. cerevisiae cells [15,27,33]. For C. neoformans, we isolated early G1 daughter cells by centrifugal elutriation and released the population into rich media (YEPD) at 30°C to monitor cell-cycle progression, as described previously [34]. This size-gradient synchrony procedure is conceptually similar to the C. neoformans synchrony procedure presented by Raclavsky and colleagues [35]. For S. cerevisiae, we isolated G1 cells by alpha-factor mating pheromone treatment [36]. We utilized this synchrony technique to isolate larger S. cerevisiae cells and to offset some loss of synchrony over time due to asymmetric cell divisions. A functional mating pheromone peptide for C. neoformans has been described but is difficult to synthesize in suitable quantities [37]. After release from synchronization, bud formation and population doubling were counted for at least 200 cells over time (Fig 1). The period of bud emergence was about 75 minutes in both budding yeasts grown in rich media, although the synchrony of bud emergence after the first bud in C. neoformans appeared to be less robust (Fig 1A and 1B). Each yeast population completed more than two population doublings over the course of the experiments.
Total RNA was extracted from yeast cells at each time point (every 5 minutes for S. cerevisiae, or every 10 minutes for C. neoformans) and multiplexed for stranded RNA-Sequencing. Between 87–92% of reads mapped uniquely to the respective yeast genomes (S1 File). To identify periodic genes, we applied periodicity algorithms to the time series gene expression datasets. Four algorithms were used to determine periodicity rankings for all genes in each yeast: de Lichtenberg, JTK-CYCLE, Lomb-Scargle, and persistent homology [38–42]. Since each algorithm favors slightly different periodic curve shapes [43], we summed the periodicity rankings from each algorithm and ranked all yeast genes by cumulative scores for S. cerevisiae and for C. neoformans (S1 Table and S2 Table, respectively). By visual inspection, the top 1600 ranked genes in both yeasts appeared periodically transcribed during the cell cycle (S1 Fig). There was no clear “threshold” between periodic and non-periodic genes during the cell cycle—rather, we observed a distribution of gene expression shapes and signatures over time (S1 Fig). Previous work on the S. cerevisiae cell cycle has reported lists ranging from 400–1200 periodic genes. To validate our RNA-Sequencing time series dataset for the S. cerevisiae cell cycle, we compared the top-ranked 1600 periodic genes to previously published cell-cycle gene lists and found a 57–89% range of overlap with previous periodic gene lists (S2 Fig) [12–15,33,41,44,45].
Three filters were applied to each budding yeast dataset to estimate and compare the number of periodic genes (S1 File). First, we pruned noisy, low-expression genes from each dataset, leaving 5913 expressed genes in S. cerevisiae (S1 Table) and 6182 expressed genes in C. neoformans (S2 Table). Next, we took the top 1600 expressed genes from the cumulative ranking of the four periodicity algorithms described above. Finally, we applied a score cutoff to each list of top 1600 genes using the Lomb-Scargle algorithm (see S1 File) [39,40,43]. We estimated that there are 1246 periodic genes in S. cerevisiae (~21% expressed genes) and 1134 periodic genes in C. neoformans (~18% expressed genes) (Fig 2). We also provided multiple criteria for evaluating the cell-cycle expression patterns of individual genes in each yeast (S1 Table, S2 Table, S1 Fig).
Cellular processes that contribute to virulence are a major focus of work in the C. neoformans field. We took advantage of the partial C. neoformans deletion collection and genetic screens for virulence factors [6] and searched for periodic virulence genes. We found that 40 genes (about 16% of the virulence genes characterized by the Madhani group and many previous studies) were periodically expressed in C. neoformans during the cell cycle (S3 Table). These virulence genes are periodic during normal cycles in rich media, which suggests that some virulence processes are directly cell-cycle-regulated. For example, budding and cell wall synthesis are coupled to cell-cycle progression in S. cerevisiae. A subset of 14 periodic virulence genes in C. neoformans had capsule and/or cell wall phenotypes reported in previous studies (S3 Table). We then asked if the 40 periodic virulence genes might be co-regulated during the C. neoformans cell cycle (S3 Fig). Over half of the periodic virulence genes clustered together and peaked in a similar cell-cycle phase (20–30 minutes into cycle 1). 11 of the 14 capsule / cell wall genes were contained in this cluster (S3 Fig, S3 Table).
Next, we wanted to ask if periodicity and temporal ordering of orthologous genes is evolutionarily conserved between the two budding yeasts. We compiled the largest list to date of putative sequence orthologs between C. neoformans and S. cerevisiae from the literature, databases, and additional BLAST searches (S1 File, S4 Table) [32,46–48]. About half of the periodic genes from each yeast (Fig 2) had at least one sequence ortholog in the other species. However, there were only about 230 pairs of orthologous genes that were labeled periodic in both yeasts. Those pairs of periodic orthologs have diverged in temporal ordering between C. neoformans and S. cerevisiae (Fig 3, S5 Table). These results indicated that the programs of periodic gene expression, and possibly the regulatory pathway, have diverged to some degree between the two budding yeasts. This altered temporal ordering between S. cerevisiae and C. neoformans periodic orthologous genes was likely not due to the experimental synchrony procedure. We obtained transcriptome data from two previous studies on S. cerevisiae cell-cycle-regulated transcription (which applied a different cell-cycle synchrony procedure, used different lab strains of S. cerevisiae, and/or measured gene expression on different platforms), and our list of periodic S. cerevisiae genes maintained temporal ordering during the cell cycle in all three datasets (S4 Fig).
Cell-cycle regulated gene expression has also been investigated in a species of pathogenic Ascomycota, Candida albicans [49]. To ask about common periodic gene expression in an evolutionarily intermediate budding yeast species, we further identified putative periodic orthologous genes shared between S. cerevisiae, C. neoformans, and C. albicans. A core set of almost 100 orthologs appeared to have both conserved periodicity and temporal ordering between all three budding yeasts (S5 Fig, S5 Table). This fungal gene set was enriched for functions in mitotic cell cycle and cell-cycle processes, which suggested that core cell-cycle regulators are under strong selection for conservation at the sequence level and by timing of periodic gene expression.
We reasoned that some cell-cycle events must be invariable in temporal ordering between fungi (S5 Fig). DNA replication (S-phase) should be highly conserved across organisms because duplication of genetic material is essential for successful division. Segregation of genomic content during mitosis (M-phase) is also essential for division, and duplication must precede division. Using annotations for S. cerevisiae [50] we identified lists of genes known to be involved in regulating events in various cell-cycle phases including bud formation and growth [51,52], DNA replication [53,54], and spindle formation, mitosis, and mitotic exit [55–58]. We filtered the resulting gene lists by periodicity in S. cerevisiae (Fig 2A, S6 Table). We then identified orthologous genes in C. neoformans without enforcing a periodicity filter.
We have previously shown that expression timing of canonical cell-cycle orthologs in S. cerevisiae and S. pombe can vary—some gene pairs shared expression patterns while others diverged [59]. To temporally align orthologous gene plots between S. cerevisiae and C. neoformans, we used the algorithmic approach described previously with S. cerevisiae and S. pombe time series transcriptome data [59]. The first, most synchronous cycle of budding data from each yeast was fit using the CLOCCS algorithm (Fig 1, S6 Fig) [59,60]. Time points in minutes were then transformed into cell-cycle lifeline points to visualize the data (see S1 File).
As observed previously, S. cerevisiae genes that regulate budding, S-phase, and mitosis were largely transcribed periodically in the proper phases (Fig 4A, 4D and 4G) [12–15]. Cell-cycle gene expression peak time patterns were examined to quantitatively compare cell-cycle phases (S7 Fig). Bud assembly and growth genes peaked throughout the cell-cycle transcription program, and the temporal ordering of these genes repeated across cell cycles (Fig 4A, S7A and S7B Fig). Similarly, spindle assembly and mitosis genes peaked in the mid-to-late phases of the transcription program (Fig 4G). DNA replication genes peaked in a defined window in the middle phase of the transcription program (Fig 4D). We observed analogous expression patterns for C. neoformans orthologs associated with S-phase and mitosis (Fig 4E and 4H), but orthologs associated with budding appeared to be expressed with less restriction to a discrete cell-cycle phase or strict temporal order (S7 Fig). This budding gene pattern can be observed qualitatively where the unrestricted expression timing creates a more “speckled” appearance in the C. neoformans heatmap (Fig 4B) and differentially timed gene expression peaks (Fig 4C).
We hypothesize that bud emergence and bud growth are not as tightly coordinated with cell-cycle progression in C. neoformans cells. Unlike S. cerevisiae where bud emergence occurs primarily at the G1/S transition, C. neoformans bud emergence can occur in a broad interval from G1 to G2 phases [61,62]. The difference in budding transcript behaviors between S. cerevisiae and C. neoformans orthologs could therefore reflect the difference in the cell biology of bud emergence and growth (Fig 4A and 4B). Only about 33% of the orthologous budding gene pairs were periodically expressed in C. neoformans, compared to 53% DNA replication and 61% mitosis orthologs (Fig 4B, 4E and 4H). Furthermore, budding orthologs that were periodic in both C. neoformans and S. cerevisiae showed some divergence in expression timing (Fig 4C). We also observed that bud emergence of C. neoformans cells during the time series appeared less synchronous in second and third cycles than S. cerevisiae cells (Fig 1A and 1B). Bud emergence in C. neoformans could be controlled by both stress pathways and TF inputs because the first budding cycle is highly synchronous after elutriation synchrony, which causes a transient stress response in released cells (Fig 1B). However, our data do not rule out a model where some budding genes in C. neoformans are controlled post-transcriptionally by localization, phosphorylation, or other periodic mechanisms. It is also possible that budding orthologs are more difficult to identify than other cell-cycle genes due to sequence divergence or that novel budding genes have evolved in the C. neoformans lineage.
We have previously shown that a network of periodically expressed TFs is capable of driving the program of periodic genes during the S. cerevisiae cell cycle [15,27]. We hypothesized that a network of periodic TFs could also function in C. neoformans to drive a similar fraction of cell-cycle genes. Thus, the temporal re-ordering of part of the C. neoformans gene expression program (Fig 3) could be explained by two models: evolutionary re-wiring of shared network TFs with S. cerevisiae or novel TF network components arising in C. neoformans to drive cell-cycle genes. First, we asked if network TFs were conserved from S. cerevisiae to C. neoformans. Indeed, a majority of network TFs and key cell-cycle regulators have putative orthology between the two yeasts (Table 1) [30]. As observed for other cell-cycle genes (Fig 4), orthologs of some network TFs were expressed in the same phase in both yeasts, while others were expressed at different times (Table 1).
Second, we asked if there were any novel periodic TFs in C. neoformans (i.e. TFs with no predicted ortholog in S. cerevisiae, or TFs with an ortholog in S. cerevisiae that is not known to function in the TF network). We constructed a list of periodic C. neoformans TFs by filtering a previously annotated transcription factor list [32] with our list of periodic genes (Fig 5, S7 Table). Indeed, 30 novel TF genes were periodic during the C. neoformans cell cycle (Fig 5A). Taken together, results from Table 1 and Fig 5 suggested that both network TF re-wiring and novel periodic TFs in C. neoformans could explain the differential ordering of periodic genes during the cell cycle (Fig 3).
Putative S-phase regulators in C. neoformans exhibited transcript behaviors that were very similar in periodicity and in ordering to their S. cerevisiae orthologs (Fig 4D–4F). Thus, we predicted that the network motifs and TFs controlling the transcription of periodic S-phase genes could be conserved. Orthologous genes in the G1/S topology were largely conserved in periodic expression dynamics at cell-cycle entry (Fig 6). The expression timing of some genes had shifted earlier in the C. neoformans cell cycle (Fig 6C and 6F, Table 1), but this result does not refute the hypothesis that these genes are activated and functional at G1/S phase. Therefore, the network topology of cell-cycle entry appeared largely conserved in C. neoformans both by sequence and by gene expression dynamics. The prediction of this model is that a common G1/S transcriptional network drives a common set of S-phase periodic genes. To test this model, we examined promoter sequences from TF network genes in S. cerevisiae and C. neoformans, as well as the promoters of 38 periodic DNA replication ortholog pairs, and did an unbiased search for enriched TF binding sequences. The core motif “ACGCGT” for SBF/MBF transcription factors [63–65] was identified in both S. cerevisiae and C. neoformans promoters. The motif was not enriched in randomly selected periodic gene promoters, suggesting that SBF/MBF is functionally conserved in C. neoformans to drive TF network oscillations and DNA replication gene expression (S8 Fig).
Here, we present the first RNA-Sequencing dataset of transcription dynamics during the cell cycle of C. neoformans. Despite evolutionary distance between Basidiomycota and Ascomycota, S. cerevisiae and its extensive genome annotation provided an excellent analytical benchmark to compare to cell-cycle transcription in C. neoformans. RNA-Sequencing has been shown to be more quantitative than microarray technology for lowly- and highly-expressed genes using asynchronous S. cerevisiae cells due to microarray background fluorescence and saturation of fluorescence, respectively [66].
We demonstrate that 20% or more of all genes in the budding yeast genomes are periodically transcribed during the cell cycle. A ranking of periodicity for transcript dynamics in C. neoformans is provided (S2 Table). For the sake of comparison, we have presented gene sets of 1100–1200 periodic genes with the highest relative periodicity scores as “cell-cycle-regulated”; however, there is a continuum of periodic gene expression dynamics during the cell cycle in both yeasts (S1 Fig). The four periodicity algorithms applied here yielded a range of periodicity scores with no clear distinction between “periodic” and “non-periodic” gene sets (S1 and S2 Tables). These results suggest that yeast mRNAs fluctuate in expression with various degrees of cell-cycle periodicity. We propose that the top 20% periodic genes presented in this study are directly regulated by periodic cell-cycle TFs in C. neoformans and in S. cerevisiae. We also posit that some of the remaining 80% genes are weakly cell-cycle regulated. For example, some genes could be subject to complex regulation with one regulatory input from a cell-cycle periodic TF and another input from a constitutively expressed TF.
We raise two important questions about the yeast periodic gene expression programs: is periodic expression of a core set(s) of genes required for the fungal cell cycle, and how are periodic gene dynamics controlled in each yeast?
In both yeasts, periodic transcription is a high dimensional cell-cycle phenotype because transcriptional state reflects the phase-specific biology of the cell cycle over repeated cycles (Fig 2 and Fig 4). In other words, G1-, S-, and M-phase genes follow a defined temporal ordering pattern. S. cerevisiae cells synchronized by different methods and/or grown in different conditions display similar ordering of periodic cell-cycle genes, despite different cell-cycle period lengths (S4 Fig). Here, we examined the transcriptome of cycling C. neoformans cells at 30°C. Other groups have shown that C. neoformans cells spend more time in G1 phase at 24°C [67]. We predict that future studies examining cell-cycle transcription of C. neoformans cells grown in different conditions (i.e. non-rich media or 37°C infection temperature) would continue to display a similar temporal ordering of cell-cycle genes. These findings provide more evidence that “just-in-time transcription” is a conserved feature of eukaryotic cell cycles [23].
We show that some orthologous periodic genes have diverged in temporal ordering during the cell cycles of S. cerevisiae and C. neoformans over evolutionary time (Fig 3). We specifically investigated genes that play a role in bud emergence and bud growth, and we find that many budding gene orthologs are not controlled in a defined temporal order during the C. neoformans cell cycle (Figs 1A, 1B, 4A and 4B). On the other hand, DNA replication and mitosis genes do appear to be conserved by sequence homology, periodic expression, and temporal ordering (Fig 4D–4I). Lastly, we find that a set of about 100 orthologous genes is both periodic and expressed in proper cell-cycle phase in the budding yeasts S. cerevisiae, C. neoformans, and C. albicans (S5 Fig) [49]. These findings suggest that there may be a conserved set of fungal cell-cycle-control genes, which represent novel therapeutic targets for fungal infections.
We posit that a network of periodic transcription factors (TFs) could control the periodic gene expression program in C. neoformans, which has been shown in S. cerevisiae and suggested in human cells [15,22,25,27]. Many orthologous genes to S. cerevisiae TF network components have diverged in expression timing in C. neoformans cells (Table 1). However, we show that the G1/S network topology is likely conserved between S. cerevisiae and C. neoformans because orthologous genes display similar expression dynamics (Fig 6). Furthermore, we find that the promoters of G1/S TF network orthologs and promoters of periodic DNA replication orthologs are enriched for an “ACGCGT” sequence motif, which matches the SBF/MBF binding site consensus in S. cerevisiae (S8 Fig) [63–65]. Therefore, we propose that the G1/S transcriptional motif—where a co-repressor is removed by G1 cyclin/CDK phosphorylation and a TF activator complex is de-repressed—is also conserved in C. neoformans (Fig 6B–6D and 6G) [29,30]. Downstream of the G1/S activator complex, the C. neoformans TF network may also contain a common forkhead domain S-phase activator and homeobox domain G1/S repressor (Fig 6E, Table 1) [14,68,69]. This partially conserved TF network model in C. neoformans explains the common G1/S topology, on-time DNA replication gene transcription, as well as differential expression of budding and other cell-cycle genes by divergent parts of the TF network.
The regulation of periodic transcription and the function of a putative TF network warrant further investigation as virulence factors of fungal meningitis caused by C. neoformans. It has been previously shown that fluconazole drug treatment can affect cell ploidy in C. neoformans [70]. More recently, polyploid Titan cells were shown to produce haploid and aneuploid daughter cells during C. neoformans infection [71]. Therefore, future work on proper regulation of DNA replication and the contribution of periodic gene products could greatly benefit our understanding of genome stability in C. neoformans. The C. neoformans TF deletion collection was recently phenotyped, and the potential of targeted TF therapies was discussed [32,72]. We have added to the C. neoformans genotype/phenotype map by documenting the functional outputs of cell-cycle TFs over synchronized cell cycles. We also propose that a conserved G1/S topology of cell-cycle TFs may initiate the cell-cycle transcription network in C. neoformans. It is possible that a multi-drug combination targeting cell-cycle regulators and previously characterized virulence pathways could yield more successful antifungal therapies [72]. For example, a combination therapy could target TFs at the conserved G1/S topology to slow cell-cycle entry and also target fungal cell wall or capsule growth. In the circadian rhythm field, it has been shown that drugs targeting Clock Controlled Genes are most potent when administered at the time of the target gene’s peak expression [73].
Interestingly, deletion of the known SBF/MBF ortholog, Mbs1 (CNAG_07464), is viable in C. neoformans [32,74]. These genetic results do not match S. cerevisiae, where swi4 mbp1 double mutants are inviable [75]. In fact, deletion of the single known G1 cyclin ortholog, CNAG_06092, is also viable in C. neoformans [10]. Mbs1 and the G1 cyclin are likely important for cell-cycle progression in C. neoformans because mutant phenotypes are highly defective in capsule formation in G1 phase, melanin production, and response to Hydroxyurea treatment during S phase [10,11,32,74]. However, the genetics are inconsistent with findings in S. cerevisiae and warrant further investigation to characterize the G1/S TF network topology of C. neoformans. It is possible that uncharacterized, redundant genes exist in the C. neoformans G1/S network motif.
We find that 40 candidate virulence genes are periodically expressed during the C. neoformans cell cycle (S3 Table, S3 Fig). An important direction for future work is to identify the mechanistic links between cell-cycle regulators and virulence pathways. 14 periodic virulence genes have annotated phenotypes in capsule formation and/or cell wall secretion. Fungal cells must secrete new cell wall and capsule during growth, and the direct links between cell cycle and these virulence factors in C. neoformans warrants further study because the cell wall and capsule are not present in host cells. The ultimate goal of this work is to identify the regulatory mechanism of periodic gene expression in C. neoformans and to find optimal drug targets and combination therapies for disrupting the fungal cell cycle.
The wild-type Saccharomyces cerevisiae strain is a derivative of BF264-15D MATa bar1 [76,77]. The wild-type Cryptococcus neoformans var. grubii serotype A strain is a derivative of H99F [47]. Yeast cultures were grown in standard YEP medium (1% yeast extract, 2% peptone, 0.012% adenine, 0.006% uracil supplemented with 2% dextrose sugar). For centrifugal elutriation, cultures were grown in YEP-dextrose (YEPD) medium at 30°C overnight. Elutriated early G1 cells were then resuspended in fresh YEPD medium at 30°C for time series experiments. For α-factor arrest, cultures were grown in YEPD medium at 30°C and incubated with 30 ng/ml α-factor for about 110 minutes. Synchronized cultures were then resuspended in fresh YEPD medium at 30°C. Aliquots were taken at each time point and subsequently assayed by RNA-Sequencing.
Total RNA was isolated by acid phenol extraction as described previously [34]. Samples were submitted to the Duke Sequencing Facility (https://www.genome.duke.edu/cores-and-services/sequencing-and-genomic-technologies) for stranded library preparation and sequencing. mRNA was amplified and barcoded (Illumina TruSeq Stranded mRNA Library Preparation Kit for S. cerevisiae and KAPA Stranded mRNA-Seq Library Preparation Kit for C. neoformans) and reads were sequenced in accordance with standard Illumina HiSeq protocols. For S. cerevisiae, libraries of 50 base-pair single-end reads were prepared, and 10 samples were multiplexed and sequenced together in each single lane. For C. neoformans, libraries of 125 base-pair paired-end reads were prepared (due to larger and more complex yeast transcriptome with introns), and 12 samples were multiplexed and sequenced together in each single lane. Raw FASTQ files were aligned to the respective yeast genomes using STAR [78]. Aligned reads were assembled into transcripts, quantified, and normalized using Cufflinks2 [79]. Samples from each yeast time series were normalized together using the CuffNorm feature. The normalized output FPKM gene expression levels were used in the analyses presented. A detailed description of each analysis pipeline is presented in the S1 File.
RNA-Sequencing gene expression data from this manuscript have been submitted to the NCBI Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo/) under accession number GSE80474.
|
10.1371/journal.pcbi.1003971 | Yielding Elastic Tethers Stabilize Robust Cell Adhesion | Many bacteria and eukaryotic cells express adhesive proteins at the end of tethers that elongate reversibly at constant or near constant force, which we refer to as yielding elasticity. Here we address the function of yielding elastic adhesive tethers with Escherichia coli bacteria as a model for cell adhesion, using a combination of experiments and simulations. The adhesive bond kinetics and tether elasticity was modeled in the simulations with realistic biophysical models that were fit to new and previously published single molecule force spectroscopy data. The simulations were validated by comparison to experiments measuring the adhesive behavior of E. coli in flowing fluid. Analysis of the simulations demonstrated that yielding elasticity is required for the bacteria to remain bound in high and variable flow conditions, because it allows the force to be distributed evenly between multiple bonds. In contrast, strain-hardening and linear elastic tethers concentrate force on the most vulnerable bonds, which leads to failure of the entire adhesive contact. Load distribution is especially important to noncovalent receptor-ligand bonds, because they become exponentially shorter lived at higher force above a critical force, even if they form catch bonds. The advantage of yielding is likely to extend to any blood cells or pathogens adhering in flow, or to any situation where bonds are stretched unequally due to surface roughness, unequal native bond lengths, or conditions that act to unzip the bonds.
| Cells adhere to surfaces and each other in the presence of forces that would easily overpower the individual noncovalent receptor-ligand bonds that mediate this adhesion, raising the question as to how these bonds cooperate to withstand such high forces. Here we show that cooperation and robust adhesion depends on the elastic properties of the bonds. A type of nonlinear elasticity referred to as elastic yielding ensures that the total force is distributed equally across the individual bonds regardless of geometry. In contrast, with linear or strain-hardening elasticity, the bonds that are stretched the most are exposed to higher forces, which cause them to fail sequentially. This work explains why elastic yielding is found in structurally and evolutionarily diverse adhesive complexes.
| Bacteria and Eukaryotic cells must resist mechanical forces when they bind to their surroundings. For example, bacteria and blood cells adhere to other cells or tissues in the presence of fluid flow that applies a drag force on the cell, while many other cells apply force to each other or to solid surfaces via cytoskeletal contraction. These mechanical forces affect the lifetime of the individual noncovalent receptor-ligand bonds that mediate cell adhesion. Some receptors form slip bonds, which are shorter-lived with applied force. However, it is now understood that many adhesive receptors form catch bonds, which are longer-lived at higher force [1], [2], [3], [4], [5]. Still others form ideal bonds, which have a constant lifetime over a range of force [5]. However, all bonds transition to slip bonds above a critical force, which is generally much less than the total force involved in cell adhesion. Thus, strong and stable cell adhesion requires clusters with multiple receptor-ligand bonds. This raises the question of whether cells have evolved mechanisms of stabilizing bond clusters.
Multivalent receptor-ligand adhesion is affected not just by the properties of the receptors, but by how they are incorporated into a cluster or cell [6]. For example, a receptor-coated and a ligand-coated surface can be easily separated by peeling forces, which stretch bonds to unequal lengths, but resist much higher forces if all bonds are stretched to the same length by shearing between two parallel surfaces [7], or if multiple bonds are stretched in parallel [8]. However, many surfaces are rough or curved, or tethers have unequal equilibrium lengths, so that bond strains are unequal regardless of force direction. When bond strains are unequal, the elastic properties of the tethers anchoring each receptor or ligand to the cell or surface affect how force is distributed among bonds. For example, longer tethers increase the rupture force of the clusters [9].In most studies of clusters of bonds, it is assumed that tethers are either stretched equally [7], [10], or are Hookean springs [11], [12], for which force increases linearly with extension (Fig. 1A).
However, many biological tethers anchoring adhesive molecules exhibit nonlinear elasticity. While entropic polymers and tissues often exhibit strain-hardening elasticity (Fig. 1A), many bond tethers exhibit yielding elasticity, where the force plateaus at a critical force, allowing long extensions at a constant force (Fig. 1A). Yielding elasticity is observed for biological macromolecules and organelles as structurally and evolutionarily divergent as membrane microvilli [13], [14], [15], [16], alpha helical proteins [17], or quaternary helices in bacterial fimbriae [18], [19], [20], [21], [22]. Moreover, the ‘saw tooth’ pattern (Fig. 1A) caused by the sequential unfolding of multiple globular domains in proteins like fibronectin [23] approaches yielding elasticity when pulled more slowly. This raises the question as to why yielding elasticity is so common in cell adhesion.
Escherichia coli bacteria with type 1 fimbriae provide an ideal model system for studying the role of nonlinear elasticity in cell adhesion because the adhesive structures are well characterized. Type 1 fimbriae exhibit nonlinear elastic extension due to uncoiling of a quaternary helix of a linear polymer that consists of hundreds to thousands of subunits [19], [24], [25]. Each type 1 fimbriae has a single FimH adhesin at the tip [26], [27], that form bonds with well-characterized properties [2], [28]. Depending on the FimH sequence, FimH can form either catch bonds that require force to be activated or strong slip bonds that do not require any activation [2]. Type 1 fimbriae thus provide and ideal system for understanding the role of tether elasticity in dynamic cell adhesion. This would require methods to probe the forces on single fimbriae during dynamic cell adhesion, and to change fimbrial elastic properties. Fluorescent methods don’t provide high simultaneous temporal and spatial resolution, in spite of recent advancements in fluorescent force sensors [29] and single molecule fluorescence [30], while other methods of measuring tether forces [31] disrupt adhesion. We also lack methods of genetically or chemically altering fimbriae to dramatically change elastic properties. Fortunately, computational simulations [28], can be used to probe bond forces and control elastic properties. Adhesive dynamics simulations were applied to E. coli binding via type 1 fimbriae, but yielding elasticity was not incorporated because simulated forces were too low to uncoil the fimbriae at the flow conditions studied [28]. On the other hand, computational models of uncoiling fimbriae have been fit to data [19], [24], [35], [36], and used to predict functional advantages [19], [25], [37], [38], [39], [40], but have never been incorporated into experimentally validated models of whole cell adhesion, so the importance of fimbrial yielding to cell adhesion remains unclear.
Here we use type 1 fimbrial E. coli adhesion as a model system to investigate the role of yielding elasticity in biological adhesion. We develop a complete model for fimbrial coiling and uncoiling in dynamic conditions by fitting a biophysical model to new elongation and contraction data obtained from Atomic Force Microscope (AFM) experiments. We introduce this model into a previously validated adhesive dynamics model for bacterial adhesion without fitting any additional parameters. We validate the complete model with new experimental data on bacterial adhesion; both model and experiments showed that bacteria crept forward but did not detach with large increases in shear stress. We showed that robust adhesion at high flow requires yielding elasticity since it could not be reproduced when other elastic tether properties were used in the simulations. We analyzed the underlying mechanisms to determine that yielding elasticity allows a nearly perfect distribution of load between bonds that were stretched to varying lengths. Finally, we predicted based on the simulations that bacteria binding via catch-bonds can withstand low flow only if exposed previously to sufficiently high flow to induce elastic yielding, which we validated experimentally. These observations demonstrate that yielding elasticity is critical for robust cell adhesion in dynamic conditions via noncovalent bonds.
In order to characterize the elastic behavior of fimbriae in experiments, we stretched and relaxed single fimbriae in a back-and-forth manner with an AFM (Fig. 1C). During initial extension, the force ramped up rapidly, then plateaued suddenly (Fig. 1D), showing the instantaneous switch from linear to yielding elasticity that has been observed previously for many fimbriae and pili [19], [24], [35], [36]. During the back-and-forth movement, the fimbriae demonstrated hysteresis since the force cycled between a higher force during extension and a lower force during retraction (Fig. 1D). The force levels during extension and retraction were calculated from pulls on several fimbriae at several velocities, (diamonds and triangles, Fig. 1E) and this was used to fit the parameters that determine the transition between the uncoiled and coiled states (xAB, xBA, k0AB, k0BA).
After a long extension, the hysteresis ended, so the extension and retraction phases converged into a single S-shaped force-extension curve that ended at about 150 pN with detachment of the cantilever from the fimbriae (Fig. 1D). For several fimbriae at several forces, the extension was measured from this curve and normalized to maximum extension (squares, Fig. 1E). This data was used to fit the parameters for the Worm-Like Chain (WLC) extension of the uncoiled state B and the stretched state C as well as the stretch transition between these states (keq, xeq, lpB, lpC, x0B, x°C).
Together this fitting resulted in the parameters shown in the table of Text S1 and dynamic elastic behavior as shown in Fig. 1E. Previous models that did not address the cooperative nature of the uncoiling transition were unable to reproduce the flatness of the main plateau [24]. Previous models that did not include the stretch transition [24] or did not allow for different persistence lengths before and after the stretch transition [25] could not fit the shape of the S-shaped curve (not shown). Thus, our new model was necessary to accurately reproduce the entire range of dynamic stretching data for type 1 fimbriae.
We incorporated this fimbrial elasticity model into previously developed adhesive dynamics simulations of E. coli, and validated the complete simulations by comparison to experimental data. Specifically, the shear stress was stepped up from 1 to 25 Pa in both simulations and experiments, and then dropped to 0.01 Pa. In both cases, the bacteria crept forward as shear increased, and then relaxed backwards when shear decreased, but not back to the original position (Fig. 2). There were small quantitative differences between the bacteria in simulations and experiments; in the simulations, the bacteria moved twice as far, and required slightly higher shear stress to begin moving. However, the relatively close fit is remarkable since there were no free fit parameters for this validation; all 33 simulation parameters were determined independently (Fig. 1 and reference [28]).
The most important observation, observed in both experiments and simulations, is that bacteria never detached, even at 25 Pa, which is higher than most physiological niches. Visual inspection of a typical simulation (e.g. Video S1) revealed the bacterium is anchored in place via one activated FimH bond at low shear, but creeps forward at increased shear, as the anchoring fimbria uncoils, until a second FimH bond is activated, and so on. We analyzed the simulations to quantify these observations. Each time the flow rate was stepped up, the mean force per bond increased suddenly (Fig. 3A), but relaxed back to about 50 pN per bond within seconds, if it had increased above this range (Fig. 3A). This drop in force corresponded to an increase in the number of uncoiled fimbriae and activated FimH bonds (Fig. 3A). Not only did the average force remain at 50 pN as shear increased further, but the distribution of bond forces was narrow (Fig. 3B). It was shown previously that FimH bonds are long-lived between 30 and 70 pN, but break within seconds above 90 pN [41] because of the exponential effects of force, so we consider bonds exposed to over 90 pN force as vulnerable to dissociation. There were almost no vulnerable bonds in these simulations, as indicated by the presence of only one symbol above the dotted line at 90 pN in Fig. 3A. Thus, bacteria in the simulations withstand high shear stress by recruiting more activated bonds and by distributing the force evenly across these bonds.
We next asked whether the nonlinear elasticity of the fimbriae was necessary for bacteria to resist high shear stresses. In the simulations, we changed the elastic properties of the fimbriae to model native yielding elasticity, strain-hardening elasticity, or linear elasticity. Shear stress was increased at 1 Pa/s until 100 Pa, or until the bacteria detached. If only one fimbria was allowed to attach (by setting the bond association rate to zero for the unbound fimbriae), then bacteria detached between 10 and 12 Pa for all regardless of the type of tether (Fig. 4A, dashed lines). Allowing multiple fimbriae to bind provided a small improvement for bacteria with strain-hardening tethers, which all detached between 10 and 18 Pa (Video S2), and slightly more improvement to bacteria with linear elastic tethers (Video S3), which detached between 15 and 25 Pa (Fig. 4A). In contrast, bacteria with multiple native yielding tethers withstood much higher shear stress, with very few detaching by 30 Pa, and over 50% remaining bound through 100 Pa (Video S4). Thus, multiple tethers with yielding elasticity were necessary in the simulations to reproduce the ability of bacteria to withstand over 25 Pa, which was observed experimentally.
To understand why the linear and strain hardening elastic tethers were unable to maintain adhesion at high shear stress, we calculated the number of activated bonds per bacterium (Fig. 4B), the mean force per bond (Fig. 4C), and the distribution of bond forces (Fig. 4D). The strain-hardening tethers failed to mediate adhesion at high shear stress because the number of activated bonds remained under two per bacterium, and the average force per bond increased to above 90 pN at and above 10 Pa. In contrast, the linear elastic tethers recruited even more bonds and maintained a similar average force per bond relative to yielding tethers in the same conditions (Fig. 4B). However, the distribution of bond forces for linear elastic tethers was broader, with over one quarter of activated bonds exposed to over 90 pN and thus vulnerable to detachment at and above 10 Pa (Fig. 4D). Since each bacterium had only 3 to 4 activated bonds in these conditions with the linear fimbriae, this means that on average one bond per bacterium breaks rapidly, transferring its load to the remaining bonds, which overloads one of them, and so on. Therefore, linear tethers recruit enough bonds, but fail to protect bacteria from detaching because they do not distribute force evenly between bonds. This demonstrates that the ability to recruit more bonds and distribute force evenly between them, which stabilizes adhesion at high shear stress, requires yielding elastic tethers.
Bacteria in vivo are often exposed to variable shear stress due to intestinal peristalsis or salivary motion. Bacteria binding via FimH catch bonds were shown previously to detach when the flow is turned down from 2 to 0.01 Pa [42], presumably because catch bonds detach at low force. However, in our current study, bacteria relaxed backwards but did not detach when shear stress was dropped from 25 to 0.01 Pa, in both experiments or in simulations (Fig. 2). Surprisingly, the number of activated bonds increased when shear stress dropped to 0.01 Pa (Fig. 3A). Moreover, while the drag force on a bacterium at 0.01 Pa is only 0.2 pN, the force per bond did not drop to near zero, but rather remained tightly distributed around 30 pN (Fig. 3B). Simulations show that the uncoiled fimbriae shorten when shear is decreased, pulling the bacterium backwards, and activating new bonds as the bacterium becomes suspended between partially uncoiled fimbriae pulling in opposite directions (Fig. 5A and Video S1). Since the bacterium is now stationary, the anchoring bond is subjected to the equilibrium uncoiling force (32.2 pN) for all partially uncoiled fimbriae.
We thus predicted that the bacteria would only stay attached at 0.01 Pa in simulations if they were first subjected to enough shear stress to uncoil fimbriae. To test this prediction, bacteria in both simulations and experiments were subjected to 1, 2.5, 5, or 10 Pa, and then dropped to 0.01 Pa. Below 5 Pa, most bacteria had one or fewer uncoiled fimbriae and activated FimH bond (Fig. 5B), and almost none maintained activated bonds after shear was decreased to 0.01 Pa. In contrast, at 10 Pa, most bacteria had 2 or more uncoiled fimbriae and activated bonds (Fig. 5B), and consistently retained activated bonds after shear was decreased. This corresponded to the ability of bacteria to remain attached after shear stress was decreased from 10 Pa but not from 5 Pa or less (Fig. 5C). This supports the idea that uncoiling and recoiling are needed to withstand variable shear stress. Finally, we validated this prediction by performing the same test in experiments (Fig. 5D). Slight quantitative differences were observed, with bacteria in experiments requiring slightly less shear stress for the same behavior, and with a higher fraction failing to detach at low flow. Thus, the experiments validated the prediction that bacteria can withstand a prolonged period at low flow better after being subjected to enough shear stress to uncoil fimbriae.
In this work, we draw our important conclusions from the simulations themselves, so it is essential that they be reliable. We ensure this by using a previously validated model in which almost all parameters were identified independently in cell-free assays, with only two parameters determined by fitting the simulations to cell adhesion data [28]. To add fimbrial uncoiling to this model, we determined all parameters independently by characterizing the elastic properties of individual type 1 fimbriae with atomic force microscopy, and fitting the data with a biophysical model (Fig. 1). Finally, we validated the accuracy of the combined model for cell adhesion by testing predictions of the model (Fig. 2, 5). Since none of the 33 parameters were adjusted to fit the cell adhesion data, minor quantitative discrepancies are expected, such as the higher shear stress required for the same behavior (Fig. 2, 5), and larger distance moved (Fig. 2) in simulations relative to experiments. The first discrepancy suggests that we underestimated the drag coefficient [28]. The second discrepancy suggests that we underestimated the number of fimbriae from the 2D projection in the electron micrographs. These small quantitative differences likely vary from cell to cell and do not affect the conclusions of this paper. The creeping we observe in simulations also resembles the behavior of E coli binding through type 1 fimbriae as flow increased in a recent publication [43]. The ability to reproduce a variety of adhesive behaviors with no adjustable parameters provides a high level of certainty to the conclusions drawn from the simulations.
Our major conclusion is that E. coli can withstand high shear stress because yielding elastic tethers called fimbriae distribute the drag force equally between multiple bonds (Fig. 3B). To understand why this is important, consider that simple mechanics theory dictates that cell adhesion in flow, like many other conditions, occurs in a peeling manner in which tethers at one edge of the adhesive contact zone are stretched the farthest [11]. While many studies have shown that tethers exhibit some sort of strain-softening, or yielding, viscoelastic behavior [13], [14], [15], [16], the theories developed to address the strength of clusters of bonds during rolling or peeling have assumed that the tethers are linearly elastic, so longer tethers apply proportionally higher force [11], [12]. Since bond lifetimes decrease exponentially with force above a critical value even for catch bonds, the bonds under most force break first, transferring force to the remaining bonds, and causing the cluster of bonds to unzip [7]. Indeed, we observe this behavior in our simulations with linear tethers, which apply a wide range of forces on bonds (Fig. 4D) and peel from the rear until they detach as the drag force is increased (Video S3). Most biological polymers and materials demonstrate nonlinear elastic properties. We show here that strain-hardening materials, which concentrate force even more on the most stretched tethers (Fig. 1, 4D), mediate even weaker adhesion in flow (Fig. 4A). However, many cells have evolved yielding elastic tethers, which provide a constant or nearly constant force independent of extension length (Fig. 1). It is well understood that bond clusters are mechanically stronger when oriented relative to force such that all bonds are stretched equally, rather than oriented so that force can unzip the cluster by stretching them one at a time [7], because the former distributes force better. However, we demonstrate here that yielding tethers can ensure equal force distribution even when bonds are stretched unequally, preventing peeling or unzipping in situations such as cell adhesion in flow. In our study, we assume that each tether has only one FimH, because this is dictated by the structure of type 1 fimbriae [26], [27]. Tethers such as microvilli can have multiple receptors per tether. In these cases, the force per tether may be distributed between multiple receptors, but our conclusions about load sharing between tethers should still apply. Thus, adhesion with yielding tethers is much more robust.
Our results demonstrate that robust adhesion requires the perfect load distribution that is unique to yielding elasticity (Fig 4). However, other previously demonstrated properties of yielding elasticity also benefit cell adhesion. Yielding provides a mechanism for creating long tethers, which reduce the force on a bond when a cell is anchored to a surface via a single tether in flowing fluid [44], although this effect is similar for long tethers with any elastic property, as shown in Fig. 4A (dashed lines). Yielding elastic tethers of all sorts are also usually viscoelastic [16], [45], [46], [47], so that they buffer force on a single bond in variable flow conditions [39], [40], or during dynamic single molecule force spectroscopy [14], [15]. Previous studies have shown that yielding forces are optimized for the catch bonds at their tips [19], [40], [48], which suggests that robust cell adhesion requires not just yielding, but yielding at a force that is appropriate for the mechanical properties of the bond supported by the yielding tether. Our conclusion can also explain previous observations about yielding tethers. Theory and simulations showed that elastic yielding tethers allowed clusters of bonds loaded in parallel to survive much longer than single bonds [37] and experiments showed that long yielding tethers greatly increased the adhesive strength of clusters of bonds between two surfaces [49]. In summary, while yielding elasticity provides many advantages to buffering force on single bonds, we demonstrate here that elastic yielding is most critical to robust adhesion of cells or large bond clusters because it uniquely distributes load equally in a complex environment.
This conclusion may apply to many cell types, because many cells have elastic yielding tethers comprised of alpha helical proteins [50], unfolding domains [51], quaternary helices [43], or membrane tethers [52], all of which yield under force. For example, leukocytes and platelets extend membrane microvilli with selectin or GPIb [13], [53], fibroblasts bind to extensible fibronectin [23] through integrin , platelets to extensible fibrin [54], bacteria bind through many types of quaternary helical fimbriae that yield [20], [21], [22], [55], and many adhesion proteins like integrins and cadherins are anchored to the cytoskeleton via alpha-helical adaptor proteins that also unfold [50]. Our simulations and experiments used catch bonds, and many of the proteins anchored to yielding tethers also form catch bonds, including P-selectin [1], L-selectin [56], GPIb [3], integrin [4], and fibrin knob-hole interactions [57]. Catch bonds and yielding tethers appear to co-evolve to provide the ideal force to optimize catch bond lifetime in order to enable robust binding in high force environments [19]. Catch bonds with yielding tethers (Fig. 3A and 4B), but not catch bonds with other elastic anchors (Fig. 4B) allow the number of activated bonds to increase proportionally to the flow rate, finally providing a mechanism for the ‘automated braking system’ observed previously for leukocytes binding via selectins [58]. Nevertheless, it is unlikely that the importance of yielding tethers is unique to catch bonds, since all catch bonds transition to slip bonds above a critical force, and in our simulations, the ability to distribute force evenly was critical to preventing the failure of FimH bonds in the high-force slip regime. This analysis suggests that yielding elastic tethers may be critical for robust adhesion of a wide range of cells binding through a wide range of receptors.
While the importance of yielding tethers to bond force distribution has not been shown previously for cell adhesion, yielding elasticity has been shown to be important in related fields. Adhesives are weak in a peeling mode, where load is applied so that stress concentrates at one edge, which propagates in a crack as the adhesive fails. This is minimized by soft thin film adhesives that can undergo a plastic deformation to form long yielding fibers that distribute stress equally along multiple fibers in spite of their difference in length. Because this deformation is irreversible, the thin film adhesives are weakened by this process, and could be improved by the development of a bio-inspired adhesive material that exhibits fully reversible viscoelastic yielding, like the yielding biological tethers described above. Yielding elasticity has also been demonstrated in fibers that make up certain biological materials, such as fibrin clots [54] and the spectrin network in red blood cell membranes [59]. These materials are resistant to tearing because yielding fibers prevent stress concentration. Thus, thin film adhesives, biological materials and cell adhesion are all strengthened by yielding fibers that prevent stress concentration and crack propagation. However, cell adhesion provides a new level of elegance, as the yielding force of the fibers must be optimized for the lifetime of the adhesive bond.
AFM experiments were conducted with an Asylum MFP3D AFM to determine the dynamic behavior of fimbriae in response to force. Olympus Biolever cantilevers were incubated with RNaseB (a naturally mannosylated protein) and surfaces with type 1 fimbriae using direct nonspecific adsorption, essentially as previously described [49]. Force pulls were controlled with a custom written script that allows back-and-forth movements at speeds from 0.1–10 µm/s. Plateau forces were determined by averaging at least 23 separate pulls from 2–4 experiments performed on different days with different cantilevers, except for the condition of recoiling at 10 µm/s, for which only 8 pulls were performed. All experiments were conducted in Phosphate Buffered Saline with 0.2% Bovine Serum Albumin (PBS/BSA) to prevent nonspecific adhesion. We have shown previously that proteins incubated in this manner remain adherent under much higher forces than 150 pN [2], and that adhesive strength is maintained even after hundreds of pulls on the same surface-immobilized fimbriae [49], so it is safe to assume that the proteins remain attached to the surface and cantilever during our experiments. Thus, the observed yielding behavior is not due to an experimental artifact.
Flow chamber experiments were performed as previously described [42]. Briefly, a bolus of E. coli expressing KB-91 FimH and K12 fimbrial shafts was introduced at a moderate shear stress (0.1–0.3 Pa) to allow bacteria to accumulate and then the shear was increased to 1 Pa to induce predominately stationary adhesion and to wash out unbound bacteria. The shear was then increased or decreased as indicated in each figure. Time-lapse videos were taken at 1 frame per second and analyzed to quantify cell position and detachment. In some experiments, a second syringe pump was used in parallel with the first to deliver a soluble inhibitor with minimal disruption to the system.
Simulations were performed as previously described [28] except that the fimbrial uncoiling model was added. Briefly, the 3D simulations model the interaction of a fimbrial-coated spherical cell with a mannose-coated planar surface in a laminar fluid flow. The tip of each fimbria represents a single FimH adhesin which can stochastically form and break bonds with the surface according to the two-state allosteric catch bond model [60]. Fimbriae can stretch, bend and buckle due to linear elastic properties [28], or uncoil and recoil with higher tension using the model described below. Simulations start with a single fimbria bound to the surface in the high-affinity state representing a bacterium that has just transitioned to stationary adhesion as in the experiments. In simulations with only a single fimbriae, the fimbriae was always set to 1 µm in length to remove differences that result from varied fimbrial lengths. In simulations with multiple fimbriae, bacteria were surrounded by 186 randomly distributed fimbriae with an average length of 0.572 µm around an exponential distribution [61].
All fimbrial models had the same length distribution as the native fimbriae. The native yielding fimbriae were modeled with a three state model in which each subunit can be in state A (fully coiled), state B (uncoiled), or state C (uncoiled and stretched), with transitions allowed between A and B and between B and C, as illustrated in Fig. 1B. The subunits in state A form a contiguous segment, since uncoiling is a cooperative phase transition that only occurs at the edge of the helical coil, as indicated by the flat uncoiling transition in Fig. 1D and 1E. However, the subunits in states B and C form noncontiguous segments because the stretch transition occurs independently for any uncoiled subunit, as indicated by the sloped stretch transition in Fig. 1D and 1E. The number of subunits in each segment is determined by force-dependent transitions between the states, according to the Bell model. The A segment was modeled as a spring, and the B and C segments were modeled as worm-like chains with different persistence lengths. The total length of the fimbriae is the sum of the lengths of the three segments. All parameters were fit to AFM data. The linear elastic tethers were modeled by disallowing all uncoiling, so that the native linear elasticity is always in effect. The strain-hardening tethers were modeled as if the uncoiling occurs at negligible force, so that they elongated beyond their native rod length under force with the WLC model, using the parameters for segment B. A complete description of the uncoiling model is provided in Text S1.
|
10.1371/journal.pcbi.1002937 | Trade-off between Multiple Constraints Enables Simultaneous Formation of Modules and Hubs in Neural Systems | The formation of the complex network architecture of neural systems is subject to multiple structural and functional constraints. Two obvious but apparently contradictory constraints are low wiring cost and high processing efficiency, characterized by short overall wiring length and a small average number of processing steps, respectively. Growing evidence shows that neural networks are results from a trade-off between physical cost and functional value of the topology. However, the relationship between these competing constraints and complex topology is not well understood quantitatively. We explored this relationship systematically by reconstructing two known neural networks, Macaque cortical connectivity and C. elegans neuronal connections, from combinatory optimization of wiring cost and processing efficiency constraints, using a control parameter , and comparing the reconstructed networks to the real networks. We found that in both neural systems, the reconstructed networks derived from the two constraints can reveal some important relations between the spatial layout of nodes and the topological connectivity, and match several properties of the real networks. The reconstructed and real networks had a similar modular organization in a broad range of , resulting from spatial clustering of network nodes. Hubs emerged due to the competition of the two constraints, and their positions were close to, and partly coincided, with the real hubs in a range of values. The degree of nodes was correlated with the density of nodes in their spatial neighborhood in both reconstructed and real networks. Generally, the rebuilt network matched a significant portion of real links, especially short-distant ones. These findings provide clear evidence to support the hypothesis of trade-off between multiple constraints on brain networks. The two constraints of wiring cost and processing efficiency, however, cannot explain all salient features in the real networks. The discrepancy suggests that there are further relevant factors that are not yet captured here.
| What are essential relationships between fundamental physical constraints and the architecture of neural systems? Most existing investigations have considered a single constraint, either wiring cost or processing path efficiency, and little is known about how characteristic neural network features, such as the simultaneous existence of modules and hubs, are related to the constraints from multiple requirements. Here we emphasized the competition between the global wiring cost and an important functional requirement, path efficiency, as factors in forming Macaque cortical connectivity and C. elegans neuronal connections. By comparing real to reconstructed networks using optimization under multiple constraints, we found that several network features are related to the competition of these two constraints, in particular the simultaneous formation of network modules and hubs. However, not all the properties of the real networks could be attributed to these two constraints, suggesting that, likely, there exists additional structural or functional requirements.
| It is widely appreciated that complex neuroanatomical networks are the physiological basis for brain dynamics, information processing and mental function [1]–[5]. In the last years, understanding the organization of neural systems and structural-functional relationship using graph theoretical approaches and methods of complex network research has attracted a great deal of interest [6]–[16]. Neural systems possess several pronounced network features, including the small-world property [17]–[19] characterized by short path lengths and high clustering [7], [12], [13], hub nodes with much larger degree than the average node degree of the network [20]–[22] and network modules broadly coinciding with functional subdivisions of the systems [14], [23]–[27]. While these features of neural networks are similar to those in many other real-world complex networks [28]–[30], the mechanisms underlying the formation of such complex network organization are still poorly understood.
The simultaneous existence of modules and hubs is an ubiquitous mesoscopic structural property in neural networks, and may play a significant role in the information processing and functioning of the systems. It was shown that cortical brain connectivity comprises dense communities, which are more densely linked internally than externally [24], [25], [27]. Such a modular organization was observed in structural networks obtained from tract-tracing studies [14], or diffusion spectrum imaging tractography [21], across various species, such as human [21], [31]–[33], cat [34], rhesus Macaque monkey [35] and C. elegans [36] as well as in functional networks derived from EEG/MEG, fMRI and MEA experiments [31], [37]–[39]. The modules of the cortical network, in broad agreement with functional subdivisions of the cerebral cortex, are spatially segregated — as the areas within the same functional subsystems (visual, auditory etc.) are mostly spatially co-localized [22], [36], [40], [41]. A recent study [42] showed that that modularity similar to that in human functional networks can be obtained based on an objective function combining the number of common nearest neighbors with a power-law decay of connections over distance, implying that modularity may be closely related to local connections.
Highly connected hub nodes have been shown in the structural network of the human brain, constructed by diffusion tensor approaches and based on 70–90 cortical gray matter areas [43]–[45] or in cortical networks based on 998 region-of-interest [21], as well as in functional networks using fMRI [46] or other imaging techniques (EEG, MEG, MEA) [39], also in other species, such as Macaque cortex [20] and C. elegans [47]. Hubs could effectively integrate information that is segregated due to the existence of modules [22]. In agreement with this idea, the identified high-degree hubs were mostly multimodal association regions [16], [22]. Moreover, in a recent study of brain pathology, a MEG study of connectivity provided additional evidence that a degradation of the small-world property in patients with Alzheimer's disease was due to disease-related changes in hubs [48].
What factors influence the formation of modules and hubs is still an open question. Nonetheless, the organization of neural networks is frequently attributed to fundamental constraints, such as metabolic cost [49], signal propagation efficiency [50], evolutionary history [51] and others. It has been speculated that the network organization is the result of an economical trade-off between the physical cost and the functional values of the topology [52]. But it is still not well understood what these functional constraints are and what the relationship is between network properties and functional values, in spite of intensive research on complex brain networks over the last decade.
One of the most extensively discussed aspects is the constraint of wiring cost [53]–[68], which is related to the possible minimization of neural wiring volume [54], [62] or wiring length [55], [58], [59], [61], [63], [65], [69] in the nervous system. Most previous studies investigated whether the actual component placement layout of neural systems has been optimized for wiring minimization, by comparing the actual wiring cost to the perturbed component placement while keeping the network connectivity as in the real systems [55], [59], [63]–[65], [69]. In coarse-grained data sets, it was found that the wiring length of the Macaque prefrontal areas [59] and C. elegans ganglia is optimized [55]. However, in other networks, such as those linking Macaque or cat visual cortical areas, the wiring was found not to be fully optimized, but relatively more optimal than other subsets of the cortex [64]. Moreover the wiring of the whole neural network of the Macaque cortex and C. elegans neuronal network appeared not to be optimized under the single wiring cost constraint– the total wiring could be decreased to 64% of the original length in Macaque and to 52% in C. elegans [63] when applying the component placement optimization (CPO) to minimize the total wiring length while preserving the specific network connectivity. Alternatively, it has been suggested that constraints such as signal propagation efficiency, measured by the global minimization of processing steps across the network, may shape the organization of neural systems [63], [68], [70], [71].
Generally none of the single constraints is likely able to account for all the functional values of the network. However, the processing efficiency perhaps is the most established network measure shown to correlate with functional performance in normal subjects and dysfunction in various brain diseases. As large-scale information processing and communication systems, neural networks favor reducing the number of intermediate transmission steps in order to respond quickly, with the tendency to minimize the average shortest path length (i.e., graph distance) [50], [72]. Minimizing the graph distance has some important functional advantages. First, a small number of intermediate transmission steps might reduce energy consumption during signal processing. In fact, about 50% of the brain's energy is used to drive signal processing, suggested by halved brain energy consumption in deep anesthesia that blocks neural signaling [73]. Second, minimizing the graph distance would increase the speed of signal processing, ultimately leading to faster behavior for decisions and actions [50]. Third, minimizing the graph distance enables neighboring and distant brain areas (or neurons) to receive signals nearly simultaneously to allow synchronous functional processing [74]–[77]. Fourth, since there exists abundant noise in signal transmission (e.g., ionic channel noise or synaptic noise) [50], [78] as well as a high failure rate for signal transmission (between 50% and 90% in individual synapses [50]), minimizing the path length would limit the interference by noise and increase the robustness of neural systems [72]. Indeed, the global efficiency (the inverse of the average shortest paths) of the resting-state brain network has been found to be strongly associated with the intelligence quotient(IQ) [79]–[82]. In disease, it was found that the efficiency of the human cortical network was disrupted in a manner proportional to the extent of white matter lesions [83]–[87]. Therefore, the graph distance/processing efficiency could be taken as a representative measure conferring functional value of brain networks.
To date, however, most studies either considered the influence of just a single constraint (mostly wiring minimization) [53]–[55], [57]–[62], [64]–[66], or evaluated two constraints (such as the metabolic cost constraint and the propagation efficiency constraint) separately [63]. However, this approach does not mean that the two constraints are independent. In fact, they may have partly opposite impact on network organization. The processing efficiency constraint favors network shortcuts that link topologically distant parts of the network, which may take the form of long-distance connections, in which case they would act against the metabolic cost constraint of wiring length minimization. Conversely, wiring minimization favors the creation of links among spatially adjacent network nodes, which may also be topological neighbors (cf. Fig. S3 in [63]). Such networks with mainly local connections typically possess low path efficiency due to a large average number of processing steps. In conclusion, the two constraints need to be considered in combination. The coexistence of modular organization and hubs in networks could be, at least partially, understood by a balance of these two constraints. Indeed, there is growing evidence to support the idea that neural network connectivity is not optimized either to minimize connection costs or to maximize advantageous topological properties, but rather is an economical trade-off between the physical cost and the adaptive value of its topology [88], see [52] for a recent extensive review. It has been proposed that the cost-efficiency balance of the human functional network [89] may be related to the behavioral performance in cognitive tasks [79], but the anatomical mechanism underlying such desirable functional connectivity is not clear.
In the present work we explored the relationship between multiple constraints and complex network architecture by systematically testing the effect of the competition of multiple constraints. We considered the neuronal network of C. elegans [63], [65], [90], [91] and the cortico-cortical network of the non-human primate (Macaque) brain based on tract-tracing studies [63], [92], for which information is available for both the spatial positions of the nodes and the network connectivity. In a previous analysis of the Macaque cortical network [27], [63], the division of the motor areas was not very highly resolved, which feature might induce biased results when analyzing network modules and spatial clustering. Therefore, in the present work we first improved the data set by a more detailed division of the motor areas based on the CoCoMac database [92], extending the former 6 motor areas to 15 areas with an additional 128 connections (see Materials and Method and Fig. S1 for the adjacency matrix). Different from the CPO method, in our scheme we compared the real network connectivity to reconstructed networks derived from multiple constraints by fixing the spatial position of each network node and the total number of (directed) connections as in the real networks. The reconstructed networks were obtained under various balancing conditions of wiring cost and processing efficiency constraints. As in previous studies [53]–[55], [57]–[66], we used the total physical distance of the wiring to represent the effect of the wiring cost constraint, and the total graph distance of the shortest paths to represent the influence of the processing efficiency constraint, and defined an objective function as a combination of both constraints using a weight parameter , namely, , with and appropriately normalized. So , or corresponds to a single constraint of path efficiency or wiring cost, respectively. Then we reconstructed the connections of the network with the help of a simulated annealing approach to minimize the objective function at different values of , starting from 50 random configurations (see Materials and Methods). We studied the general properties of the competition between the two constraints in a 1D model with one-dimensional uniform spatial layout of nodes and directed connections. For the real neural networks, we investigated the modularity and hub properties of the reconstructed networks and studied the relationship to the spatial layout and compared them to those of the real networks. We found that for certain intervals of balancing these two constraints, the reconstructed networks showed a very similar modular structure and similar spatial positions of the hubs as the real networks. These results are also related to the nonuniform layout and clustering of the network nodes (neurons for C. elegans and areas for Macaque cortical brain) in space. Despite the observed agreements, there still exists significant discrepancies between model and real networks, suggesting that there are additional functional requirements to be considered in the future.
The qualitative properties of the competition between the constraints at different values of were found to be quite common for the 1D model (Fig. 1), the real Macaque cortical network (Fig. 2) and C. elegans neuronal network (Fig. 3), the latter two having highly non-uniform spatial layout of nodes. These properties can be most clearly seen in the 1D model. In addition to and , we used several parameters such as the number of hubs , the average degree of the hubs and the fraction of spatially local connections to characterize the reconstructed optimal networks at various values (see Materials and Methods). The results are summarized below.
At , the network is only optimized by the processing efficiency constraint and achieves minimization of the topological distance. The optimal network configuration depends on the density of connections. If the initial network is sparse (roughly, connection probability ), as in the 1D model (, Fig. 1A) or C. elegans neuronal network (, Fig. 3C)), the reconstructed network is characterized by a single hub connecting to all other nodes while the rest of connections appears to be random. The position of the hub node is arbitrary. A single hub is very effective for reducing the graph distance when the network is sparse enough, because the path length between any pair of nodes is either 1 (direct link) or 2 (connected through the hub). When the network becomes denser (), such as the Macaque cortical network (), the pathlength between any pair of nodes in a random network cannot be larger than 2. It is then very unlikely to obtain a single hub when optimizing from an initially random configuration. Thus, the reconstructed network connections appear to be random (Fig. 2C). The dependence of the network configuration on the connection density is also more systematically shown in the 1D model (Fig. S2).
At , the processing efficiency constraint plays no role while the wiring cost constraint is fully dominant, and the network has the minimal total wiring length. Most of the connections are local and there is no hub () in any of the three systems (Figs. 1A, 2F, 3F). For the 1D model or the reconstructed networks of the Macaque, all the connections are local (, Fig. 4J,K); is reduced to of that in the real Macaque network (Fig. 4B). As for the C. elegans network, many neurons in the head and tail are very densely distributed with very small distance among them, while the distance between neighboring neurons in the ventral cord is much larger. Therefore, significant numbers of non-nearest but short-distant connections within the head or the tail let the physical distance become very small, only of the real C. elegans (Fig. 4C), but is clearly smaller than 1.0 (Fig. 4L)). And in the real networks, the spatial layout of the network nodes is non-uniform, forming spatial clusters; these local connections make the adjacency matrix become non-uniform, showing some clustered pattern similar to the original network connectivity (Figs. 2F, 3F). Spatial clustering and module organization in the reconstructed and real networks are studied below.
For , the processing efficiency and wiring cost constraints combine their impact, resulting in two distinct regimes, depending on the networks. When is positive, but not very large, such as for the 1D model (Fig. 1B,C), for Macaque (Fig. 2D) and for C. elegans (Fig. 3D), there is a single hub connecting to all the other nodes in all the three systems (). Most of the other connections are linked to the nearest neighbors, and is close to 1.0 in the 1 D model (Fig. 4J) and the reconstructed networks of Macaque (Fig. 4K), but is clearly smaller in C. elegans for the reason stated above (, Fig. 4L). This single, global hub is a very effective configuration to provide high efficiency, when the other connections are short-distance due to the cost constraint. While the placement of the hub is arbitrary in the 1D model due to a symmetrical spatial layout (Fig. 1B,C), it is unique in the reconstructed networks of Macaque and C. elegans, located close to the global geographical center of the whole network, as will be discussed in detail below. In this regime, , , and are all constant with respect to , since the optimal network configuration does not change with , though the speed of convergence in the optimization process does. While both and in the reconstructed networks are smaller than those in the real networks in this region, is much more smaller (reduced to in Macaque and to in C. elegans, see Fig. 4B,C).
The insensitivity of the optimal configuration to in this regime can be understood from the objective function . The variation is , where and with increasing . The network configuration would change with only if , i.e., . However, for the Macaque network, we found that for any perturbation to the configuration obtained at (a single global hub and all other local links). When , , thus , then , and the configuration change cannot be accepted by the optimization. Only when (), it becomes possible to have for certain perturbations, and a configuration change will happen. For this reason, in all the three networks, the optimal solution is the same in a range of values, depending on the spatial layout of the network.
With further increase of , the influence of the wiring cost constraint becomes stronger. A single hub is no longer found, because very long-distant connections are prohibited and the systems move into a different regime ( for the 1D model, for Macaque and for C. elegans ). Several smaller hubs emerge, with connections extending to nodes in their spatial neighborhood, and the connection range of such regional hubs becomes smaller at larger , as can be most clearly seen in the 1D model (Figs. 1D,E), and is also true in the real networks (Figs. 2E and 3E). Consequently, is further reduced slightly, but increases and is close to that of the real network (Fig. 4B,C). As shown below, the spatial positions of the hubs in the reconstructed networks are close to the real hubs in the two original networks.
When is very close to 1.0, where the efficiency constraint is weak while the wiring cost constraint is almost fully dominant, most of the connections are local and there are no pronounced hubs (). increases quickly and becomes larger than that in real networks.
We also found that for all values, the input and output degrees of the nodes in the reconstructed networks are largely symmetrical (Fig. S3A and Fig. S3C). While the input and output degrees were found to be significantly correlated in the real networks (Fig. S3B and Fig. S3D), the discrepancy between the optimization model and real data is quite large, because this model does not include possible requirements that could generate the asymmetry, for instance, input-output information flowing as in real networks (e.g., from sensory neurons to motor neurons in C. elegans).
The above results show that the coexistence of local connections and hubs in the cortical networks could be a solution to the multiple constraints of wiring cost and processing efficiency. There is a regime ( for Macaque and for C. elegans) in which the competition of the constraints can allow the formation of several hubs connecting to many of the nodes in the neighborhood. Here, in the reconstructed networks is very close that of the real networks, but can be much smaller. In the following sections, we show that both the modular structure and the positions of hubs are quite similar to the real networks in this regime.
The above results (Figs. 2D–F and 3D–F) showed the emergence of a modular organization in the reconstructed networks similar to the real ones, which is derived from the non-uniform spatial distribution of the nodes and the local connections due to the constraints. In order to further explore these relations, we examined the relation between the spatial clustering of nodes and the modules in network connectivity both in the real and reconstructed networks, and compared the similarity between the modules in the real and reconstructed networks.
The results in Section I. showed that the emergence of hubs could be the result of the combination of the wiring cost and path efficiency constraints, since hubs connecting to many nodes are very effective for improving path efficiency, while allowing most of the other connections to remain local to satisfy the cost constraint. If the hub is a global one, connecting to all the nodes, then it is reasonable to expect that the position of the hub is not arbitrary, but is near the geographical center (the node with minimal total distance to all the other nodes) of the whole network in order to maximally reduce the total wiring length of the connections from the hub. If the hub is a regional one, connecting to most of the nodes within the local region (e.g., within one of the spatial clusters), then the position should be close to the geographical center of this region. Thus, a node with many other nodes densely distributed in the neighborhood (having a high neighborhood density, see Materials and Methods), could be an candidate for a hub under the two constraints.
In the following section, we present findings regarding the location of the hubs in the reconstructed as well as the real networks. In our analysis, network nodes with z-score of total degrees (input and output) larger than 2 (see Material and Methods) are considered as hubs. The locations of the hubs in all 50 realizations of the reconstructed networks at each were identified. We would like to point out that identifying hubs in this way is heuristic and may introduce some ambiguity when comparing different networks (real or reconstructed at different ). However, hubs that were defined in this way indeed provided a plausible way to describe the variation of the nodes with the largest degrees in the network as changes.
The results in the above sections suggested that the mesoscopic properties of networks, the simultaneous formation of modules and hubs, can be partially explained by the combination of the wiring cost and processing efficiency constraints. Now we examine the degree of nodes in the original and reconstructed networks.
When constructing networks under the wiring cost constraint, the nodes tend to connect to their nearest spatial neighbors, which is confirmed by a high value of . Therefore, the number of connections of a node (degree) is expected to be related to the neighborhood density of the node in a certain spatial range. We calculated the density of nodes for various radii and evaluated its correlation with the degree of nodes. The correlation between the degrees and density in reconstructed network with strong enough wiring cost constraint ( close to 1) is quite large in a range of for both the Macaque and C. elegans, as shown by the black dashed curve in Fig. 9A,B.
Interestingly, the correlation between degrees and density is also significantly present in the real network (colored curves in Fig. 9A,B). For example, the correlation can reach 0.49 at for Macaque and 0.29 at for C. elegans, much larger than the of the significance level in the corresponding surrogate data. The results are consistent with the observation above that hubs in the real network are ranked high in term of the neighborhood density.
Although the correlation between degree and neighborhood density is significant in both real and reconstructed networks, the discrepancy is also quite large. In particular, the degree distribution in the real network was not well reproduced in the reconstructed network (see Fig. S9 for a comparison). In both neural systems, the real networks have higher probabilities to have large-degree nodes, but the reconstructed networks have higher probabilities to generate intermediate-degree nodes, because the connections in the reconstructed networks are much more strongly determined by the neighborhood density. It has been shown that both the in- and out- degree of the C. elegans neuronal network obey a power-law distribution [47]. Our test of the significance of the power-law fitting to the distribution (see Materials and Methods) confirmed this statement (, , for in-degree and , , for the out-degree, Fig. S9). The real Macaque cortical network and the reconstructed networks of both systems do not show scale-free features. For all these networks, the largest value is 0.020.02 for the out-degree distribution of C. elegans obtained at .
These results indicate that degrees of the individual nodes reflect the impact of the wiring cost constraint, however there are still some other unknown factors in addition to the wiring cost and processing efficiency constraints that may strongly influence the node degrees in the original network.
In the above sections we explored the similarity between reconstructed and real networks in global network measures (e.g., physical and graphical distance), and in mesoscopic properties (e.g., modules, hubs, degree and degree distribution). Next we analyzed how well the reconstruct networks recovered the connections in the real network. We calculated the recovery rate by comparing the overlap of the adjacency matrices of the original and reconstructed networks entry by entry, taking both asymmetry of 0's and 1's entries and the directionality of connections into consideration (see Material and Methods). is maximal at 1.0 when two networks are identical in all connected and un-connected pairs.
The overall recovery rate of the whole network as a function of is shown in Fig. 10A,B for the two neural systems. For the Macaque cortical network, when , the recovery rate is nearly 60% (Fig. 10 A), which is clearly larger than the recovery rate of random benchmark networks obtained by rewiring the original network while retaining the input and output degrees. For the C. elegans neuronal network, when , the recovery rate is more than 30%, which is significantly larger than 19.14% of random benchmark networks (Fig. 10 B). The recovery rate is much smaller than that of the Macaque networks, partially due to much sparser connectivity in C. elegans.
Importantly, the recovery rate is not uniform in the networks. It is higher within the spatial clusters, but lower between the clusters, as shown as green bars in Figs. 10C, D for Macaque and C. elegans respectively (both for ). Notably, for C. elegans, the recovery rate of the connections within clusters could approach 40% (Fig. 10 D), almost doubling that of the random networks (blue bars). For the connections between the clusters, the recovery rate is even smaller than that in the random benchmark networks in both systems, especially significant for C. elegans, because the connections between clusters become too sparse in the reconstructed networks (see Figs. 5C, 6C). The real neural network has more long-range connections between spatial clusters, likely due to additional functional requirements.
Although the recovery rate within modules (red bars) is also larger than that of the random networks, particularly within spatial clusters, the recovery rate between modules is much higher than for random benchmarks for both systems, in contrast to the case between spatial clusters. The different results between clusters and between modules originate from the mismatching of modules and clusters.
In the combination of wiring cost and efficiency constraints, most of the connections are local. Therefore, the recovery rate is expected to be higher for the local connections. Our more detailed analysis of the recovery rate for the connections of different physical distance confirmed this expectation (Fig. 10E,F). In both the reconstructed and real networks, the probability of two nodes to be connected decreases with distance (insets, Fig. 10E,F). However, in real networks, the connection probability for very short distance is much smaller than in the reconstructed network, but it decays more slowly with the distance; therefore there is much higher probability of long-range connections, especially in C. elegans, likely due to other functional requirements. Nevertheless, the recovery rate is much higher for short-distant links, around 50% in both systems.
There are some differences between the C. elegans neuronal network and Macaque cortical network. (1) The C. elegans neuronal network is quite sparse (connection probability ), while the Macaque cortical network is rather dense (). Therefore the recovery rate of the random benchmark networks is smaller in C. elegans () than Macaque (). Furthermore, when (efficiency constraint only), the reconstructed networks of C. elegans possess a global hub, but for the Macaque network they are quite similar to random networks (Fig. 2C and Fig. 3 C). Thus for C. elegans, at is even smaller than that of the random networks (). (2) About of the neurons in C. elegans gather in the tiny head, and the others scatter in the ventral cord and tail. This highly heterogeneous spatial distribution of neurons requires more long-distant connections in the real network. Indeed, C. elegans has a smaller fraction of local connections () when compared to Macaque () (Fig. 4K, L, dashed lines). These properties are perhaps the main reasons that the recovery rate in C. elegans is low, because too many short-distant links are put into the small head in reconstructed networks when compared to the real networks (Fig. 6 and Fig. 10E,F).
The formation of complex neural networks is subject to many structural and functional constraints. In particular, it has been speculated that the network organization is the result of an economical trade-off between the physical cost and the functional value of the topology [52]. Among various graph theoretical measures, the average shortest pathlength representing the processing efficiency can be taken as an essential representative of functional constraints. In this study, we showed in a systematic and quantitative manner that the competition between these two straightforward constraints, the overall wiring cost and signal propagation efficiency, plays an important role in the network organization of Macaque cortical connectivity and C. elegans neuronal connections. By reconstructing network connections using optimization under multiple constraints, while fixing the component layout as in the real network and comparing the properties the reconstructed to the real networks at different topological scales, we revealed that the connectivity in both neural networks is closely related to the spatial arrangement of the nodes. The main findings are as follows. i) The combination of the wiring cost and processing efficiency constraints can lead to the simultaneous formation of local connections and hubs. ii) When the spatial layout of the nodes is not uniform, but clustered as seen in the real networks, this combination will lead to the formation of modules and hubs. The modules have strong overlap with the spatial clusters and the hubs are located at global or regional geographical centers. iii) In certain regimes of competition between the two constraints, the reconstructed networks have a modular organization quite similar to that in real networks, and the positions of the hubs coincide or are close to the actual hubs. iv) The analysis also revealed that the degrees of nodes in the real networks are correlated with their neighborhood density, and the reconstructed networks can recover a significant portion of the individual links of the real networks, especially for short-distant connections. These observations support the idea of a trade-off between cost and functional values. The two constraints, however, cannot fully explain all the properties in the real networks. There are discrepancies in several important aspects: (1) There are significantly more long-range connections in both real systems when compared to the model. (2) The correlation between the individual degree and the neighborhood density in the real networks is much lower than in the optimization model. (3) The degree distribution in the real network is different from that in the reconstructed network. (4) The model cannot generate asymmetry in the input and output degrees as in the real networks. All these differences suggest that there are additional important factors influencing the formation of real neural networks. In the following sections we discuss these points in more detail.
It is clear that real networks are not solely optimized by topological efficiency, due to a large number of local connections. Neither are they optimized only for minimal wiring cost, even though this cost constraint is playing a significant role. Previous studies investigated whether the component placement layouts in the real neural networks satisfy the concept of wiring length minimization, by comparing the wiring cost of real networks to perturbed component placement while keeping the network connectivity fixed. The accumulated evidence showed that the wiring cost constraint indeed plays an important role [55], [58], [59], [61], [63], [65], [69]. However, for the Macaque cortical network and C. elegans neuronal network, previous work using CPO while fixing the network connectivity showed that the wiring length is not minimized [63]. Here we reconstructed the network connectivity while keeping the component placement layout as in the real network. Under only the wiring cost constraint at , the reconstructed network differed significantly from the real networks: all the connections were short-distant and there were no hubs and no long-range connections as in the real networks.
The following comparison between real networks and reconstructed networks at showed more clearly that the wiring cost constraint is at work, but the real neural networks are not solely optimized by minimal wiring length. (1) The wiring cost in the reconstructed networks is clearly lower than that of the real networks (53.5% for Macaque but only 1.25% for C. elegans), although the link recovery rate is still significant (60.8% for Macaque and 27% for C. elegans), since the wiring cost constraint does impose many short-range connections in the real networks. In the CPO solutions [63], by fixing the network connections as in the real networks, the wiring lengths were 68% and 52% of the real values of the networks in the Macaque and C. elegans, respectively. This finding shows that allowing hubs as in the real network is costly in wiring, especially for C. elegans. (2) The processing efficiency is much lower than in the real networks. As seen in Fig. 4B,C, (the reverse of the processing efficiency) in the real networks is much smaller than the reconstructed networks at . When is slightly smaller than 1.0, the regional hubs in the reconstructed networks reduce the path length significantly and is close to the values in real networks; interestingly these values are not much larger than the optimal at . This finding is consistent with the previous observation that biological neural networks appear to be optimized for processing efficiency [63]. (3) The recovery rate is quite high within the clusters. In particular, the recovery rate of the connections within the Macaque visual system (covering areas of the red cluster of Fig. 5) is , and it is for the rest of the connections of the network, which is consistent with previous observations that the wiring among the Macaque visual cortex is relatively more wiring cost optimal than other cortical subsets [64]. (4) In C. elegans, solely minimizing the wiring cost at cannot produce strong modular structure in reconstructed networks (, Fig. S6B) (as explained in the next section). However, similar modular organization as in the real network can be obtained under the balancing of two constraints (Fig. 3).
We showed that the wiring cost constraint can predominantly determine the modular organization in real neural networks, and the partition of the modules is not very sensitive to the effect of the processing efficiency constraint at different . More precisely, the spatial clustering governs the connectivity modules in the reconstructed networks where most of the connections are extended to the spatial neighbors under the wiring cost constraint. Modules in the reconstructed network overlap largely with the spatial clusters. In the Macaque cortical network, the mismatched areas (mainly of the somatosensory system) are near the boundary of two spatial clusters (see Fig. 5E). This arrangement reduces the wiring cost when these areas make more connections to the other cluster to form a module, likely due to some further functional requirements.
Notably, the modular organization in the C. elegans neuronal network is most likely determined by the combination of the wiring cost and path efficiency constraints. With only the wiring cost constraint at , the reconstructed network of C. elegans did not show strong modularity as seen in the real network (Fig. S6B, at ). The main reason is as follows. About of the neurons gather in the tiny head, while other neurons scatter in the ventral cord and the tail with much longer physical distance between them. At , almost all the links in the reconstructed networks were put in the head to form a highly connected core. The rest of the nodes in the body and tail form an approximately one-dimensional array with a minimal number of necessary links in order to avoid disconnection from the main core, but do not form dense modules. However, when the path efficiency constraint becomes effective for slightly smaller than 1.0, some long-range connections are forced between the hubs in the head and the other parts of the network. These long-range connections from the hubs now can also take the role of avoiding disconnection of the nodes into subsets. Now the remainder of the connections is allowed to be more short-ranged in the tail and in the ventral cord, which form modules that coincide with the spatial clusters. With suitable combinatorial influence of the two constraints at , the fraction of mismatched neurons is reduced to about 23% (Fig. 6F).
In both systems, the best overlap between the modules in the reconstructed networks and real networks appears for a region , suggesting that the combination of the two constraints plays an important role in the formation of real neural networks which possess both modules and hubs.
The emergence of hubs can be attributed to the combination of the wiring cost and path efficiency constraints, because a single constraint on its own, either efficiency () or wiring cost (), does not support hubs in networks as dense as the Macaque cortical network. For sparse networks, such as the one of C. elegans, the efficiency constraint on its own can generate a global hub, but the hub position is arbitrary. The wiring cost constraint is effective even for small values. In both systems, there is a broad range of the competing parameter ( for Macaque and for C. elegans) where the reconstructed networks are composed of local connections and just one dominant hub. For larger , the wiring cost constraint becomes stronger and several regional hubs emerge, similar to the organization in real networks. However, the wiring cost constraint on its own () cannot enforce hubs in the reconstructed networks.
The nodes chosen as the hubs in the reconstructed networks occupy the highest ranks of the regional geographical centrality and neighborhood density in order to be wiring economic. Interestingly, real hubs are found close to these nodes with high centrality. Quite strikingly, in the Macaque cortical network, the location of the biggest hub at area V5/MT and the output hub of area 46 in the real network of the Macaque cortical network can be reproduced in the reconstructed networks by the combination of the wiring cost and path efficiency constraint at . The strong impact from the wiring cost constraint in real networks is further reflected by the economic wiring of these hubs nodes: (1) the areas V5/MT and area 46 stay near the geographical centers of the red and blue clusters, ranked No. 3 and No. 2 of the regional geographical centrality of the red and blue clusters, respectively; these areas have most of their connections within their respective cluster. (2) V5/MT and area 46 are ranked No. 1 and No. 3 in terms of high density of nodes in the neighborhood within a certain range of radius. In C. elegans, the locations of hubs in the reconstructed network were also found to be very close to or overlapping with the position of real hubs, and there is also a regime ( which best matches modules and hubs to those of the real network. In this regime, is close to the real networks, but is much smaller, especially in the C. elegans neural network.
It appears in both systems that there are more nodes with larger degrees in the real networks than the reconstructed networks. However, importantly, the nodes with the largest degrees are close to the regional geographical centers similar to the model. Although one cannot expect that only two constraints can recover all the features of the real networks, the observation of the overlapping and close spatial locations of the reconstructed and real hubs (i.e., economic wiring of the hubs) provided clear evidence that probably both cost and efficiency constraints are at work in the real neural systems, supporting the idea of a trade-off between cost and functional values of the networks.
In a broad range of the balance parameter ( for Macaque and for C. elegans), the reconstructed networks have a single dominant hub linking to most of the nodes. Such a configuration is due to a strong impact of the path efficiency constraint, however, it is not functionally robust. Failure of the single dominant hub node greatly degrades the efficiency in information processing, because without the central hub, the network has mainly local connections. We performed a systematic analysis of the impact of removing the node with the largest degree in the reconstructed networks and measured the increase of [94]. The results are shown in Fig. S10. It is seen that in the regime with a single, global hub, in the reconstructed networks increases significantly when the hub is removed. The degrading effect is much more serious in C. elegans, because the reconstructed networks without the hub contain an almost one dimensional array for the neurons in the body that separates head and tail. In the next regime with several regional hubs, ( for Macaque and for C. elegans), the reconstructed networks are much more robust against the removal of the node with the largest degree (Fig. S10), very close to that in the real network. Thus, while hubs enhance the processing efficiency significantly, they are also the points of vulnerability to pathological damage [52]. The high energy consumption of the brain puts it under high vulnerability for energy undersupply, and the metabolically most expensive node is particularly vulnerable in pathological circumstances. There is evidence that metabolic costs of a node are proportional to its degree [95]; thus some brain disorders may be closely related to hubs' abnormalities. For example, this effect has been speculated to be a network mechanism for Alzheimer's disease [48], [52], [96]. In the C. elegans neuronal network, laser ablation of hub neurons (AVA and AVB interneurons) generated uncoordinated phenotypes [97], [98]. In particular, laser ablation experiments have demonstrated that AVA neurons are required for normal spontaneous and evoked backward locomotion [99]–[101], while AVB neurons are responsible for forward locomotion [100], [102]. Therefore, avoiding the configuration of a single dominant hub could indeed be functionally significant. In conclusion, the robustness requirement might be an additional important factor that generates the evolutionary pressure for the real networks to have more hubs, so that both in Macaque cortical network and C. elegans neuronal networks, the competition between the two constraints is settled down to a regime with smaller, but very close to 1.0.
We have shown that there is a regime () of the competition between the two constraints of wiring cost and processing efficiency where the reconstructed networks can reproduce some major properties of the real networks, including the simultaneous formation of hubs and modules, locating hubs or other large-degree notes close to the regional centers of the spatial clusters and similar resilience to node failure. However, there are still significant differences. The reconstructed networks have much smaller wiring length (about 54% of the real wiring length for Macaque and nearly 5% for C. elegans). This difference stems from the fact that the real networks have more long-range connections: there are more large-degree nodes or hubs (Figs. 4E,F, Fig. S3), the fraction of spatially local connections is much smaller and the probability of long-range connections is much larger (insets in Figs. 10E,F). Moreover, the number of connections is not strongly determined by the neighborhood density (Fig. 9).
Another significant difference is the asymmetry in the in- and out-degrees in real networks. Although directed links are used in our model influencing the calculation of processing efficiency, the objective function did not contain possible constraints that can reflect the functional role of the asymmetry in the input and output links; the reconstructed networks are thus highly symmetric. In fact, the asymmetry in the biological neural networks may be closely related to functional requirements of specific signal processing flow. For instance, in C. elegans, there are more connections from the sensory neurons to motor neurons than vice versa. Also from the connection matrix of the Macaque cortical network, there are more connections from some visual areas to the motor system, and fewer along the opposite direction. Therefore, in future work, one may include additional constraints to enforce information flow between different types of nodes, such as input nodes (the sensory neurons or primary sensory areas) and output nodes (the motor neurons or motor areas).
In our view, the differences imply that most likely real neural networks are influenced by additional functional requirements and constraints. The trade-off between the wiring cost and the processing efficiency with these new constraints could give a better account of the asymmetry and other properties of the real neural networks.
We studied the formation of complex network connectivity derived from multiple constraints, in particular the competition between wiring cost and path efficiency requirements. By reconstructing networks while preserving the spatial layout of the components, we obtained an understanding of the relationship between the spatial layout and network connectivity derived from the multiple constraints. This understanding guided us to investigate the relationship between the spatial layout and network connectivity of the real Macaque cortical network and C. elegans neuronal network. The results are consistent with previous observations that wiring cost and efficiency constraints are playing an important role in shaping the network organization and provide evidence to support the idea of a trade-off between them. While significant, the wiring cost and path efficiency constraints cannot completely explain all the features in the connectivity patterns of the real Macaque cortical network and C. elegans neuronal network. Other factors, such as robustness of networks against node failure, more long-range connections and asymmetric input and output information flow appear to be important. Moreover, with our combinatory optimization model, we have mainly discussed the question of which constraints are at work to influence the neural network properties, but have not yet addressed possible mechanisms underlying the biological implementation of these constraints. Exploring additional functional constraints and incorporating them into growth or generative models [42], [103] may be worthwhile directions in the future research.
The present study on the impact of multiple constraints on the architecture of neural systems might also provide an anatomical foundation for the cost-efficiency balance in human functional networks [79]. A more detailed understanding of the relationship between the cost-efficiency balance in anatomical and functional neural networks will require the analysis and modeling of the dynamical interdependence in the reconstructed networks under different competition conditions relative to that in real networks.
|
10.1371/journal.pgen.1007897 | A specialized MreB-dependent cell wall biosynthetic complex mediates the formation of stalk-specific peptidoglycan in Caulobacter crescentus | Many bacteria have complex cell shapes, but the mechanisms producing their distinctive morphologies are still poorly understood. Caulobacter crescentus, for instance, exhibits a stalk-like extension that carries an adhesive holdfast mediating surface attachment. This structure forms through zonal peptidoglycan biosynthesis at the old cell pole and elongates extensively under phosphate-limiting conditions. We analyzed the composition of cell body and stalk peptidoglycan and identified significant differences in the nature and proportion of peptide crosslinks, indicating that the stalk represents a distinct subcellular domain with specific mechanical properties. To identify factors that participate in stalk formation, we systematically inactivated and localized predicted components of the cell wall biosynthetic machinery of C. crescentus. Our results show that the biosynthesis of stalk peptidoglycan involves a dedicated peptidoglycan biosynthetic complex that combines specific components of the divisome and elongasome, suggesting that the repurposing of preexisting machinery provides a straightforward means to evolve new morphological traits.
| Bacteria show a variety of different cell shapes that are critical for survival in the environmental niche they inhabit. While the mechanisms generating the prototypic rod-shaped and coccoid morphologies have been studied intensively, only little is known about the processes that underlie the formation of more complex morphological features. The model organism Caulobacter crescentus is characterized by a polar stalk, which carries an adhesive organelle mediating surface attachment at its tip. This structure forms through the insertion of new cell wall material at its base and elongates considerably in phosphate-limited conditions. Our work reveals significant differences in the architecture of cell walls isolated from stalks and cell bodies, respectively, hinting at the existence of a stalk-specific cell wall biosynthetic apparatus. To identify components of this machinery, we systematically inactivated and localized proteins with a predicted enzymatic or regulatory function in cell wall biosynthesis in C. crescentus. Our results show that stalk formation is mediated by a pole-associated complex composed of proteins that have previously been identified as components of the cell elongation and cell division machineries. The stalk biosynthetic apparatus may thus have evolved through the repurposing of preexisting machinery, indicating that even complex morphological traits can emerge without the need for extensive changes to the complement of morphogenetic factors.
| The shape of most bacteria is determined by a cell wall made of peptidoglycan (PG), a mesh-like heteropolymer that surrounds the cytoplasmic membrane and provides resistance against the internal osmotic pressure [1, 2]. The backbone of PG is formed by strands of alternating N-acetylglucosamine (GlcNAc) and N-acetylmuramic acid (MurNAc) subunits. These glycan chains are connected by short peptides that are attached to the MurNAc moieties, giving rise to a single elastic macromolecule known as the PG sacculus [3].
The PG meshwork needs to be continuously remodeled to allow for cell growth and division [4]. In Gram-negative bacteria, this task is achieved by a large and seemingly redundant set of PG synthesizing and degrading enzymes. Insertion of new cell wall material is initiated by the translocation of lipid-linked GlcNAc-MurNAc-pentapeptide precursors across the cytoplasmic membrane to the periplasm [5–7]. Glycosyltransferases (GTases) then incorporate the disaccharide units into preexisting glycan strands, while the L-Ala–D-Glu–L-Lys/meso-DAP–D-Ala–D-Ala pentapeptides of adjacent glycan strands are crosslinked by transpeptidases (TPases) [2, 8]. Depending on their domain structure, PG synthases can be classified as bifunctional GTases/TPases (class A PBPs), monofunctional TPases (class B PBPs) and monofunctional GTases [8].
The majority of TPases are DD-TPases, also known as penicillin-binding proteins (PBPs) [9]. These proteins catalyze the formation of D-Ala4–meso-DAP3 (4-3) crosslinks, in a reaction that releases the D-Ala5 moiety of the donor molecule [10]. Alternatively, crosslinks can also be formed between two meso-DAP3 residues (3–3 crosslinks), catalyzed by specific LD-TPases that use tetrapeptide side chains as donor moieties and release their terminal D-Ala4 residue to gain energy for the crosslinking reaction [11]. For PG to grow, cells require not only synthetic but also lytic enzymes that cleave bonds in the PG meshwork and thus make space for the insertion of new material [12]. Depending on their cleavage specificity, these so-called autolysins can be typically sorted into three main categories. Lytic transglycosylases act on the glycan strands and cleave the β-1,4-glycosidic bond between MurNAc and GlcNAc, leaving 1,6-anhydro-MurNAc as the terminal residue [13]. Amidases, by contrast, hydrolyze the amide bond between the peptide and the MurNAc moiety [12], whereas endo- and carboxypeptidases hydrolyze specific amide bonds within the peptides [8, 14].
The formation and degradation of PG need to be closely coordinated to prevent cell lysis [2, 15], a task that is presumably achieved by the assembly of synthetic and lytic enzymes into dynamic multi-protein complexes [16]. In the majority of rod-shaped bacteria, two of these complexes have been identified to date. The first one, called the elongasome (or Rod complex), mediates the dispersed incorporation of new PG along the lateral walls of the cell during the elongation phase. Its positioning is controlled by the actin-like protein MreB [17–19], which forms patch- or arc-like filaments that are attached to the inner face of the cytoplasmic membrane [20–24]. These structures move around the circumference of the cell and, thus, ensure even growth of the rod-shaped sacculus. In Gram-negative bacteria, their effect on the PG biosynthetic machinery is mediated by the transmembrane protein RodZ [25–27], which links MreB to a periplasmic complex containing the elongation-specific monofunctional TPase PBP2 [2, 28, 29]. Towards the end of the elongation phase, PG synthesis is taken over by a second complex, called the divisome [2, 30], which mediates pre-septal elongation and subsequent constriction of the PG sacculus at midcell. Its positioning and activity are regulated by FtsZ, a tubulin homolog that assembles into a dynamic ring-like structure at the future division site. This so-called Z-ring then recruits, directly or indirectly, all other components of the cell division machinery. The divisome includes a variety of PG synthases and hydrolases, among them the division-specific monofunctional TPase PBP3 [31], which act together to coordinately remodel the PG layer during the division process. Of note, in some species, MreB relocalizes to the division site before the onset of cell constriction but then moves back to the lateral walls as cytokinesis progresses, suggesting that the elongasome and divisome cooperate during certain stages of the division cycle [32, 33].
While the function of the elongasome and divisome and their roles in establishment of generic rod and coccoid morphologies have been studied intensively [2, 34], the mechanisms generating more complex cell shapes are still poorly understood. A model organism known for its distinctive morphological features is the alphaproteobacterium Caulobacter crescentus (henceforth Caulobacter) [35]. This species is characterized by a biphasic life cycle that involves two morphologically and physiologically distinct cell types. One of them, the swarmer cell, possesses a single polar flagellum mediating swimming motility. The stalked cell, by contrast, displays a tubular extension (stalk) whose tip carries an adhesive holdfast mediating surface attachment. Whereas the stalked cell undergoes repeated cycles of chromosome replication and cell division, the swarmer cell is arrested in G1 phase, searching its environment for nutrients. However, at a defined point in the cell cycle, it sheds its flagellum, starts to establish a stalk at the previously flagellated pole, and enters S phase. The cell then elongates, forms a new flagellum at the pole opposite the stalk, and finally divides asymmetrically to produce a stalked cell and a new swarmer cell [36]. The biological role of the Caulobacter stalk is still controversial, but it may serve as a spacer to elevate the cell above the substratum and thus enhance its access to nutrients [37]. Consistent with this idea, its length increases up to 20-fold under conditions of phosphate limitation [38].
In Caulobacter species, the stalk consists almost exclusively of the three cell envelope layers (inner membrane, cell wall and outer membrane) and does not contain any cytoplasm [35, 39]. Moreover, it is compartmentalized by large disc-like protein complexes, so-called crossbands, which are deposited at irregular intervals along its length, serving as non-selective diffusion barriers that physiologically separate the stalk envelope from the cell body [35, 40]. Formation of the stalk is driven by zonal incorporation of new cell wall material at the stalk base, as detected by the labeling of newly synthesized PG with tritiated glucose [38], radiolabeled D-cysteine [41], or fluorescently labeled D-alanine derivatives [42]. To date, various mutants have been identified that lack stalks under standard growth conditions [43–46]. However, in all cases, cells regained the ability to form stalks after transfer into phosphate-limited media, indicating that they suffered from a block in the cell cycle-regulated initiation of stalk formation rather than a defect in the underlying biosynthetic machinery. Depletion of MreB or the elongasome-specific GTase RodA [47], by contrast, impaired stalk elongation under all growth conditions [48]. Similar results were obtained after inhibition of the elongasome-specific TPase PBP2 [49] with the β-lactam antibiotic mecillinam. However, because of the severe global cell shape defects observed in these cases, it was difficult to conclude on a specific role of MreB, RodA and PBP2 in the stalk biosynthetic pathway. Finally, a moderate reduction in stalk length was observed for mutants lacking the cytoskeletal protein bactofilin A (BacA) or the BacA-associated class A PBP PbpC [50]. Together, these results suggest that components of the generic PG biosynthetic apparatus may be critical for stalk formation, but the precise composition of the machinery responsible for this process still remains elusive.
In the present study, we comprehensively investigate the mechanism of stalk formation, focusing on phosphate-limiting conditions to obtain a sensitive readout of the contributions that individual factors make to this process. We show that phosphate starvation induces a G0-like resting state that is characterized by the absence of key cell cycle regulators, including FtsZ. Comparing the muropeptide profiles of isolated stalk and cell body PG, we then identify significant differences in the composition of cell walls from these two compartments, suggesting that stalks are formed by specialized machinery with distinct biosynthetic properties. Systematic deletion and localization studies of cytoskeletal and PG biosynthetic proteins then indeed reveal a distinct set of factors involved in stalk elongation, which we characterize in detail with respect to their impact on PG composition and the spatial regulation of PG biosynthesis. Morphometric analysis of the corresponding mutants shows that these factors make varying and, in part specific, contributions to stalk and cell body elongation, indicating that these two modes of growth a mechanistically distinct. Finally, we identify MreB as a key component of the stalk biosynthetic complex and pinpoint a region on its surface that appears to be required for stalk formation but largely dispensable for elongasome-mediated lateral growth. Collectively, our results show that stalk formation represents a specialized growth process that is mediated by a composite complex including components of both the elongasome and divisome, with distinctive properties that clearly differentiate it from other PG biosynthetic machineries.
Although the stimulatory effect of phosphate starvation on Caulobacter stalk elongation has been known for decades [38], the underlying regulatory mechanisms are still poorly understood. Prompted by the fact that stalk formation is tightly linked to cell cycle progression, we set out to investigate the effects of phosphate deprivation on central cellular processes such as DNA replication and cell division. First, flow cytometry was used to assess the replicational state of cells after transfer from standard to phosphate-free (M2G-P) medium. To this end, replication initiation was blocked with rifampicin and ongoing rounds of replication were allowed to finish. Previous work has shown that Caulobacter cells contain a single chromosome that is replicated only once per division cycle [51, 52]. Consistent with this finding, we observed that cells accumulated either one or two chromosome equivalents when grown in standard conditions, indicating that a large fraction of the population was in S-phase (Fig 1A). However, upon phosphate deprivation, DNA replication gradually ceased, with most cells arrested in G1 phase after 24 h of incubation. These data suggest that the lack of phosphate leads to a block in the cell cycle prior to S-phase, thereby preventing new rounds of chromosome replication. To support this conclusion, we visualized the number and positions of the chromosomal replication origins. In doing so, we made use of a fluorescently (GFP-) tagged derivative of the chromosome partitioning protein ParB, which interacts with specific motifs (parS) in the origin region [53, 54]. The expression of GFP-ParB thus typically results in the detection of either one or two foci, depending on the number of origin copies in the cell. Microscopic analysis revealed that most (~ 85%) cells exhibited a single ParB focus at the stalked pole when subjected to 24 h of phosphate starvation, indicating that they are arrested in G1 phase (Fig 1B). To clarify the reason for this G1 arrest, we analyzed the cellular levels of the replication initiator protein DnaA and the cell cycle master regulator CtrA, which act as positive and negative regulators of chromosome replication, respectively [52]. Interestingly, both proteins were rapidly depleted from the cells during phosphate starvation (Fig 1C), indicating that key drivers of the Caulobacter cell cycle are absent under this condition.
To correlate changes in cell cycle progression with the growth behavior of cells, we monitored changes in cell mass and number after a shift to phosphate-limiting conditions. Interestingly, the optical density of cultures kept increasing exponentially for more than 10 h and only leveled off after ~ 50 h of incubation (Fig 1D), suggesting that cells made use of internal phosphate storage compounds to compensate for the lack of an external phosphate source. Consistent with the detection of DNA replication events (Fig 1A), cells still multiplied during the initial exponential phase. However, after longer starvation periods (> 24 h), the viable-cell count started to decline, whereas the cell mass still increased, likely due to continued elongation of the cell bodies and stalks in the absence of cell division events (see Fig 2B). Western blot analysis indeed revealed that the essential cell division protein FtsZ was depleted from the cells upon phosphate starvation (Fig 1C). The same was true for the cell division regulator MipZ, an inhibitor of FtsZ polymerization that limits Z-ring formation to the midcell region [54]. In line with these findings, an FtsZ-YFP fusion induced after prolonged phosphate starvation formed multiple foci in the vicinity of the stalk-distal pole instead of a defined midcell band (Fig 1B), indicating the absence of a functional and properly localized Z-ring [54]. Notably, FtsZ was never observed at the stalk base, supporting the previous notion that it does not play any role in stalk formation [54].
Taken together, our results demonstrate that phosphate starvation arrests the Caulobacter cell cycle in a G1-like phase, thereby stalling DNA replication and cell division until phosphate becomes available again.
Phosphate starvation induces Caulobacter to enter a non-replicative state in which cells continue to elongate their cell body and stalk. To investigate this atypical mode of growth, we set out to visualize sites of active PG biosynthesis using the fluorescent D-amino acid 7-hydroxy-coumarin-amino-D-alanine (HADA) [42, 55] as a tracer. As a control, we initially analyzed the growth dynamics of cells growing in phosphate-replete medium. To this end, cells were synchronized and then pulse-labeled with HADA at different stages of the cell cycle. Consistent with previous results [41], we observed disperse incorporation of new cell wall material before the onset of cell division, followed by zonal growth at midcell during the constriction phase (S1 Fig). Moreover, concurrent with the switch from disperse to zonal growth, an additional intense focus of fluorescence appeared at one of the cell poles, reflecting the establishment and outgrowth of the stalk. This polar signal faded gradually as the cell cycle progressed and was no longer detectable in late pre-divisional cells. Thus, HADA reliably detected all known growth zones in Caulobacter cells.
Next, we used HADA labeling to determine the pattern of PG synthesis under phosphate-limiting conditions (Fig 2A). After 6 h of incubation in phosphate-free medium, most cells showed a bright fluorescent patch at the stalked pole as well as a faint disperse signal extending throughout the rest of the cell body. Cells longer than ~ 4 μm often displayed an additional bright focus at their center, which could reflect FtsZ-dependent zonal growth or cell division, consistent with the observation that the viable-cell counts still increased in the early phase of starvation (Figs 1D and S2A). Interestingly, the intensity of the polar signal decreased considerably upon appearance of a midcell focus, suggesting that the machineries mediating stalk formation and cell division may compete with each other for at least some of their components (Fig 2A). After longer starvation periods (>18 h), midcell foci were almost undetectable, and HADA fluorescence was largely limited to the stalk base, which correlates with the lack of cell division events at this time point. Notably, the intensity of the polar signal decreased slightly during long-term incubation (S2B Fig), although the rate of stalk elongation remained constant at all time points (Fig 2B). The increase in cell body length, by contrast, was most pronounced during the early phases of starvation, when cells still showed midcell HADA foci, suggesting that it may, at least in part result, from FtsZ-mediated zonal growth at the cell center. Collectively, phosphate starvation induces a switch in the pattern of PG synthesis that ultimately limits cell growth to the stalked cell pole.
Our and previous labeling studies suggest that stalk formation is driven by the insertion of new cell wall material at the stalk base [38, 41, 42]. To determine whether stalk PG is still subject to modification or turnover, phosphate-starved Caulobacter cells were incubated with HADA for an extended period of time (1.5 h). After this treatment, staining was observed throughout the entire cell envelope (Fig 2C), including distal segments of the stalk that were clearly formed prior to the start of the labeling procedure (considering a stalk elongation rate of 0.28 ± 0.03 μm/h; see Fig 2B). Given that HADA is likely incorporated by the action of periplasmic DD- or LD-TPases [42, 56, 57], this finding indicates the presence of transpeptidase activity in the stalk compartment that mediated the addition of HADA to preexisting PG independently of the pole-associated biosynthetic complex. After transfer of the cells to HADA-free medium, fluorescence was rapidly lost in the cell bodies and in the basal region of the stalk, whereas it was stably retained in the distal stalk segments, indicating that stalk PG is not turned over at significant rates (Fig 2C and 2D). Notably, the same behavior was observed for a strain lacking crossbands. The differences in the behavior of cell bodies and stalks may thus not result from restrictions in the diffusion of envelope-localized PG biosynthetic enzymes but rather from the retention of these enzymes at the stalk base, possibly due to the lack of cytoplasm in the stalk compartment.
The distinct mode of growth involved in stalk formation opens the possibility that there may be compositional differences between the PG layers encompassing the cell body and stalk compartments. To address this issue, phosphate-starved cells were agitated vigorously to shear off stalks from the cell bodies. After separation of the two compartments by differential centrifugation (S3 Fig), PG was isolated from each of the fractions and subjected to muropeptide analysis. Interestingly, stalk PG contained a high proportion of 3–3 crosslinked peptides and non-crosslinked tripeptides (resulting from the cleavage of 3–3 bonds), whereas these muropeptide species were barely detectable in the cell body samples (Fig 3 and S1 Table). Similarly, the total fraction of crosslinked peptide side chains was significantly higher in stalk PG, mostly because of a higher proportion of trimeric muropeptides. The glycan chain lengths, by contrast, did not vary between the two compartments. Collectively, these findings indicate that the PG layers of stalks and cell bodies differ in both the type and extent of peptide crosslinks.
Previous work has shown that 3–3 crosslinks are generated by LD-TPases, which are characterized by a conserved YkuD domain [11]. The Caulobacter genome contains two so-far uncharacterized open reading frames, CC_1511 and CC_3744, which encode proteins with this signature domain (now referred to as LdtD and LdtX, respectively). To determine how these factors contribute to the distinctive composition of the stalk cell wall, we generated a strain carrying in-frame deletions in both the ldtD and ldtX gene and analyzed the composition of PG purified from its stalk and cell body compartments. In both samples, 3–3 crosslinked peptides and non-crosslinked tripeptides were virtually undetectable (Fig 3 and S1 Table), indicating that the formation of these muropeptide species is linked to the activity of the two predicted LD-TPases. Notably, however, the total fraction of crosslinked peptides barely changed in either of the compartments, because the loss of 3–3 crosslinks was compensated by a proportional increase in the fraction of 4–3 crosslinks. Thus, LD-TPase activity is not the main factor responsible for the elevated degree of crosslinking detected in stalk PG (see Fig 4 for a summary of the key results obtained in this study).
The stalk is physiologically separated from the cell body, because it is devoid of cytoplasm and contains crossband complexes that block the exchange of periplasmic and membrane proteins [40]. It was conceivable that crossbands could help establish the differences in the PG composition observed for the two compartments, for instance by facilitating the establishment of distinct pools of PG biosynthetic enzymes or blocking the diffusion of lipid II into the stalk structure. To test this idea we determined the muropeptide profile of stalk and cell body PG isolated from a crossband-less strain (ΔstpAB, SW51). Notably, we still observed a higher content of 3–3 crosslinks and a higher total proportion of crosslinked peptides in stalk PG (S1 Table). Similar to the differences in PG turnover (Fig 2C), this characteristic thus appears to be independent of the presence of crossbands.
Stalk formation involves a growth process that is distinct from the disperse and zonal incorporation of PG mediated by the elongasome or division complex, respectively. To determine the composition of the underlying machinery, we systematically analyzed all predicted PG biosynthetic proteins encoded in the Caulobacter genome for their contribution to stalk elongation under phosphate-limiting conditions (see Fig 4 for a summary). In doing so, we initially focused on enzymes with PG synthase activity, including RodA, mono- and bifunctional PBPs and LD-TPases. Previous studies have shown that the depletion of RodA or the inhibition of the monofunctional DD-TPase PBP2 with mecillinam largely abolished the synthesis of stalks in phosphate-replete media, although these treatments concomitantly induced severe morphological defects in the cell body [48, 49]. To verify these results, we determined the consequences of RodA depletion on stalk elongation during phosphate limitation (S4 Fig). The absence of RodA indeed led to a severe decrease in both stalk and cell body length. Similar tendencies were observed for mecillinam-treated cells, although the effects were generally less pronounced. These results indicate a global morphogenetic function of RodA and PBP2 in phosphate-starved cells. To further investigate these proteins, we generated functional fluorescent protein fusions [58] and analyzed their localization in phosphate-limited media. Consistent with a global role in PG biosynthesis, both GFP-RodA (S9 Fig) and GFP-PBP2 formed large patches that were distributed throughout the cell bodies. Moreover, both proteins frequently formed a faint but distinct focus at the stalked pole, particularly in cells with clearly elongated (> 4 μm) bodies (Fig 5A). Thus, RodA and PBP2 appear to be partly associated with the stalked pole, where they likely cooperate to mediate stalk formation.
Previous work has also implicated bifunctional PBPs in stalk elongation, largely based on the analysis of mutants lacking one of these proteins [50, 59]. To verify and extend these results, we analyzed strains carrying single or multiple mutations in the PBP-encoding pbpY, pbp1A, pbpC, pbpX, and pbpZ genes [59, 60]. Our results confirm that deletion of pbpC led to a moderate reduction in stalk length, whereas the absence of any other PBP, either alone or in combination, did not have any effect (Fig 5B). However, as observed under standard growth conditions [59, 60], at least one bifunctional PBP was required for viability during phosphate starvation (S5A Fig). In line with the results of the deletion studies, localization analyses revealed that none of the bifunctional PBPs except for PbpC accumulated at the stalked pole, indicating that these proteins may not be specifically associated with the stalk biosynthetic machinery (S5B Fig). Notably, however, PbpX appeared enriched in the stalk compartments, but the significance of this observation remains unclear.
Finally, we analyzed the role of the two predicted LD-TPases LdtD and LdtX in stalk formation. Although these proteins make a significant contribution to PG crosslinking in the stalk compartment (Fig 3), their inactivation did not have any apparent phenotypic effect (Fig 5B). LD-TPase activity may thus not contribute to the establishment of the stalk structure per se but rather have an accessory function that serves to modify the biophysical properties of the PG layer. Localization studies indicate that LdtD and LdtX do not accumulate at the stalk base, suggesting that they may act independently of the polar stalk biosynthetic machinery (S5C Fig).
Apart from PG synthases, stalk formation must also involve autolytic enzymes that cleave the PG sacculus and, thus, enable the insertion of new cell wall material at the stalk base. However, to this point, the nature of the factors involved has remained unknown. To address this issue, we systematically screened mutants lacking one or multiple predicted PG hydrolases for defects in stalk growth under phosphate-limiting conditions. The enzymes tested included all LytM-like and NlpC/P60-like endopeptidases, AmiC-like and CHAP domain-containing amidases, soluble and membrane-bound lytic transglycosylases, and carboxypeptidases identified in the Caulobacter genome (S2 Table). In most cases, the lack of single factors and even the absence of whole enzyme families had no apparent effect on stalk length (S6 Fig). Four strains, however, displayed obvious morphological defects (Fig 6; see S7 Fig for the phenotypes in PYE medium). One of them was a mutant lacking the protein DipM, a catalytically inactive LytM-like endopeptidase homolog that was previously shown to be critical for proper PG remodeling during cell division [61–63]. The absence of DipM led to a severe reduction in stalk length, combined with the formation of branches within the stalk structure or the establishment of multiple stalks, often emanating from the same pole (Figs 6B, 6C and S8). Even shorter stalks were observed in the combined absence of the soluble lytic transglycosylases SdpA and SdpB (Fig 6B and 6C), two proteins previously found to be associated with the divisome complex [64]. Apart from its aberrant morphology, the ΔsdpAB mutant frequently showed membrane blebs that were associated with the residual stalk structures, suggesting a defect in membrane attachment or homeostasis (S8 Fig). Milder effects on stalk length were caused by inactivation of the divisome-associated carboxypeptidase CrbA (Billini et al, unpublished) or the LytM-like endopeptidase LdpA, a thus-far uncharacterized protein encoded in an operon with the polarly localized scaffolding protein bactofilin A (BacA) [50, 65] (Figs 6B, 6C and S8). Importantly, despite their stalk elongation defects, none of the four deletion strains showed a significant reduction in cell length (Fig 6C), indicating that stalk and cell body growth are mechanistically distinct processes that proceed independently of each other.
To further investigate the functions of the five autolytic factors identified in the mutational screen, we generated fluorescently (mCherry-) tagged derivatives of these proteins and analyzed their localization patterns under conditions of phosphate starvation (Fig 7). Both the DipM and CrbA fusions accumulated at the stalk base and may, thus, be specifically associated with the polar stalk biosynthetic machinery. The SdpA and SdpB fusions, by contrast, were distributed throughout the cell envelope, suggesting that the two proteins may either act independently of the polar complex or associate with it in a very transient manner. Unlike the other proteins analyzed (S9 Fig), LdpA-mCherry was quantitatively cleaved at the junction between the two fusion partners, preventing further analysis.
In order to determine how the absence of the different autolytic factors influenced the pattern of PG biosynthesis, mutants lacking these proteins were grown in phosphate-limiting conditions and subjected to HADA staining (Fig 8). Consistent with their relatively mild stalk elongation defect, ΔldpA cells still displayed a pattern similar to that of the wild-type strain. In the ΔdipM and ΔcrbA strains, by contrast, the polar signals were much fainter and new cell wall material was often incorporated at non-polar sites. An even more pronounced effect was observed in the ΔsdpAB mutant, which virtually lacked polar foci and instead showed patchy or even HADA fluorescence throughout the cells. Thus, the severity of the stalk elongation defect scales with the loss in polar PG biosynthesis.
To obtain more detailed insight into the effects of the different mutations on the structure of the PG layer, we isolated whole-cell sacculi from wild-type and mutant cells after prolonged (24 h) phosphate starvation and subjected them to muropeptide analysis (S3 Table). For the wild-type strain, whole-cell sacculi gave similar results as PG from isolated from a cell body fraction (compare Fig 3 and S1 Table), indicating that the characteristic features of stalk PG are largely obscured by the excess of cell body PG in the whole-cell preparations. Interestingly, there were hardly any differences between the muropeptide profiles obtained under phosphate-limiting (S3 Table) and phosphate-replete [66] conditions. The average composition of cell body PG thus appears to be independent of the phosphate supply. Among the mutant strains, ΔldpA cells showed essentially the same average PG composition as the wild-type strain. The same was true for the ΔdipM mutant, with exception of a significant increase in the proportion of non-crosslinked tetrapeptides (S3 Table), which could indicate an elevated level of endopeptidase and/or carboxypeptidase activity. The muropeptide profiles of the remaining strains, by contrast, showed marked global changes. In line with the notion that CrbA acts as a carboxypeptidase, removing the terminal D-Ala residue of pentapeptide side chains (Billini et al, unpublished), the ΔcrbA mutant displayed a considerable decrease in the total content of tetrapeptides that was accompanied by a proportional increase in the content of pentapeptide-containing muropeptide species (S3 Table). In ΔsdpAB cells, on the other hand, the average glycan chain length increased from 7 to 9.4 disaccharide units, consistent with the loss of lytic transglycosylase activity. Surprisingly, the mutant cells additionally showed a severe reduction in the degree of crosslinkage. At the same time, their total content of pentapeptide side chains was reduced, whereas the proportion of tripeptide side chains was considerably elevated (S3 Table). These results suggest that the lack of SdpAB leads to reduced transpeptidation or, more likely, elevated endopeptidase activity.
Collectively, our results show that several components of the autolytic machinery are critical for proper PG remodeling during stalk formation, with some of them localizing to the stalk base under phosphate-limiting conditions (see Fig 4 for a summary). Notably, most of the proteins, including DipM, SdpA, SdpB and CrbA, are associated with the cell division apparatus under standard growth conditions [61–64] (M. Billini, unpublished), suggesting parallels in the mechanisms that reshape the PG layer during cell constriction and stalk growth.
Polymer-forming scaffolding proteins are critical for the regulation of many growth processes in bacteria [4, 67], suggesting that this group of proteins may also play a critical role in stalk formation. Previous work has indeed implicated the bactofilin homolog BacA in stalk biogenesis [50]. Re-analysis of a ΔbacA mutant revealed a significant reduction in both stalk and cell body length during phosphate starvation (Figs 9 and S10). Notably, deletion of the endopeptidase gene ldpA, which lies in a putative operon with bacA, had a very similar effect on stalk length, whereas it barely affected the cell body (Fig 6). These results suggest that LdpA and BacA may specifically cooperate in stalk formation, whereas BacA is additionally involved in a distinct pathway involved in cell body elongation.
As another scaffolding protein, MreB was shown to be required for stalk formation in media containing moderate to high levels of phosphate [48, 68]. To clarify the contribution of this protein to stalk biosynthesis under phosphate starvation, we employed strains producing MreB or the adapter protein RodZ under the control of an inducible promoter. When starved for phosphate in the absence of inducer, both mutants showed a drastic reduction in stalk length or occasionally even failed to form stalks at all (Figs 9 and S10). The effects on the cell bodies, by contrast, differed depending on the protein depleted. Cells lacking RodZ showed a length distribution indistinguishable from that of the wild-type strain. Depletion of MreB, by contrast, markedly decreased the median cell length. Similar effects were observed for wild-type cells treated with the MreB inhibitor A22 [69, 70] (Fig 9). These findings indicate that, under phosphate-limiting conditions, RodZ appears to be specifically required for stalk biosynthesis, whereas MreB additionally contributes to cell body elongation, again supporting the idea that these two processes are driven by distinct mechanisms.
Apart from MreB, the MreCD complex has been identified as a factor critical to lateral growth in many rod-shaped bacteria [4]. MreC is thought to serve as a scaffold that interacts with various PG biosynthetic enzymes, including the monofunctional TPase PBP2 [71, 72] and the monofunctional transglycosylase RodA [73]. In E. coli, it is part of the elongasome complex [74], whereas it was shown to establish an elongasome-independent structure in Caulobacter cells [68, 75]. To test for a role of this protein in stalk formation, we analyzed the morphology of a conditional mreC mutant grown under phosphate-limiting conditions (Fig 9). In the absence of inducer, the cells started to elongate but eventually became amorphous and lyzed. In most cases, stalks were either absent or barely recognizable, indicating that the MreCD complex may be essential for both cell wall integrity and stalk biosynthesis during phosphate starvation. Given the role of MreC in the activation of PBP2 and RodA in E. coli, the differences in the phenotypes obtained after inactivation of these proteins are unexpected. Either Caulobacter MreC has functionally diverged from its E. coli homolog or its depletion is more complete than the depletion of MreB or the inhibition of PBP2 with mecillinam. Collectively, our results show that the bactofilin homolog BacA and the elongasome components MreB, RodZ and MreC are required for proper stalk biosynthesis in Caulobacter cells.
To clarify whether the role of the different scaffolding proteins in stalk formation involves their recruitment to the stalked pole, we analyzed the localization patterns of fluorescently tagged derivatives in cells subjected to phosphate starvation (Fig 10). Both the MreB and RodZ fusion formed a distinct focus at the stalk base and, in rare cases, also a second focus at the pole opposite the stalk. Together with the polar localization of PBP2 (Fig 5A), these findings indicate that key components of the elongasome complex relocate to the site of stalk biosynthesis in phosphate-limiting conditions. There, they colocalize with BacA, which retains its polar position irrespective of changes in the phosphate supply (Fig 10). The MreC fusion, by contrast, formed a broad band at midcell, whereas it was largely excluded from the polar regions (Fig 10). In line with the global morphological defects caused by its depletion, MreC may have a general role in cell wall biosynthesis, but it does not appear to be stably associated with the polar stalk biosynthetic machinery.
To determine the role of the different scaffolds in polar PG biosynthesis, cells lacking these factors were subjected to HADA staining after phosphate deprivation (Fig 11). Interestingly, despite its severe stalk elongation defect (Fig 9) the ΔbacA mutant still displayed intense polar foci, indicating that BacA is an accessory factor that is not critical for the global reorganization of PG biosynthesis induced under phosphate-limiting conditions. Consistent with this idea, muropeptide analysis showed that deletion of bacA did not have any appreciable effects on global PG composition (S4 Table). Depletion of MreB or RodZ, by contrast, strongly decreased the intensity of the polar HADA signals, and frequently led to the insertion of cell wall material at pole-distal sites. In both cases, these defects were accompanied by significant changes in the whole-cell muropeptide profiles. Similar to the ΔsdpAB mutant (compare S3 Table), the degree of crosslinkage was significantly reduced, mostly due to a decrease in the proportion of highly crosslinked (trimeric and tetrameric) muropeptide species. Moreover, there was a striking increase in the proportion of muropeptides with tripeptide side chains, indicative of high levels of LD-TPase activity. Thus, cell wall stress caused by reduced levels of PBP2-mediated DD-transpeptidation may trigger a fail-safe mechanism that stabilizes the PG meshwork through the formation of abundant 3–3 crosslinks.
Collectively, these results demonstrate that MreB and its transmembrane adapter RodZ play a central role in the establishment of the polar PG biosynthetic zone that gives rise to the stalk structure (see Fig 4 for a summary).
Our data demonstrate that several components of the PG biosynthetic machinery localize to the stalked pole in phosphate-starved cells, suggesting that they assemble into a complex mediating the synthesis of stalk PG. To obtain more insight into the factors mediating the recruitment of these proteins, we reanalyzed the localization patterns of DipM-mCherry, CrbA-mCherry, Venus-MreB, CFP-RodZ, and BacA-CFP in all deletion strains that showed defects in stalk elongation (ΔdipM, ΔsdpAB, ΔcrbA, ΔldpA, and ΔbacA). However, in all cases, the positioning of the fusion proteins remained unaffected, indicating that neither lytic factors nor the bactofilin cytoskeleton are required for complex assembly. Given the prevalence of elongasome components among the polarly localized proteins, we then tested the role of MreB in the recruitment process. Treatment of cells with the MreB inhibitor A22 not only led to the delocalization of the known MreB interactor RodZ but also abolished the polar foci of DipM and CrbA (Fig 12). Thus, MreB appears to be a key organizer of the stalk biosynthetic complex. Notably, A22 had no effect on the polar localization of BacA, indicating that the bactofilin scaffold acts independently of MreB.
To analyze the dynamics of the polar MreB assembly, we constructed a sandwich fusion in which mCherry was inserted into a surface-exposed loop of the MreB protein, following the strategy previously used for its E. coli homolog [26] (Fig 13A and 13B). A strain carrying the respective allele (mreBsw) in place of the endogenous mreB gene showed normal growth rates (S11A Fig). However, in rich medium, the sandwich fusion failed to fully relocate to midcell during the early phases of cell division. Moreover, cells were slightly longer and more highly curved than the wild type and occasionally formed branches and/or filaments (Figs 13C and S11B). Under phosphate-limiting conditions, by contrast, the distribution of cell lengths was similar to that of the wild-type strain (S11B Fig). Strikingly, mreBsw cells failed to form stalks in both standard and phosphate-limited medium. Consistent with this observation, the fusion protein no longer condensed into polar foci during phosphate starvation but retained the patchy localization pattern typically observed in exponentially growing cells [76] (Fig 13C and 13D). This unusual behavior went along with changes in the global muropeptide profile that were qualitatively similar to those observed for MreB- and RodZ-depleted cells but considerably less pronounced (S4 Table). Consistent with the slightly aberrant morphology of the mutant cells, this finding suggests that the insertion of mCherry leads to a mild general defect in MreB function. Importantly, however, it appears to additionally block a specific set of interactions that are critical for the polar recruitment of MreB, thereby preventing stalk formation. To our knowledge this is the first report of a Caulobacter strain that is completely stalkless under all growth conditions.
Collectively, these results demonstrate that MreB has a key role in the assembly and function of the polar stalk biosynthetic complex in Caulobacter (see Fig 4 for a summary).
Bacterial cells come in a variety of different shapes, but in most cases the mechanisms generating this morphological diversity are poorly understood [77]. This study uses the Caulobacter stalk as a readily amenable model system to investigate the molecular principles underlying the development of species-specific morphological traits. Previous work has shown that stalk growth is driven by zonal PG incorporation at the old cell pole [38, 41, 42]. Initially, this process was thought to be mediated by FtsZ and mechanistically similar to pre-septal cell elongation [68, 78, 79]. However, localization studies revealed that FtsZ is not detectable at the stalked pole, neither during normal cell cycle progression [54] nor during phosphate starvation (Fig 1B), excluding the divisome as a relevant player in stalk formation. Other reports implicated MreB and RodZ in stalk growth, suggesting that the elongasome could have a dual role in both cell body and stalk elongation [48, 68]. Clarification of this issue is complicated by the fact that the inactivation of factors with a global role in PG biosynthesis leads to pleiotropic morphological defects. Exploiting the fact that phosphate starvation suppresses Caulobacter cell division while strongly promoting stalk elongation, we were able to disentangle cell body- and stalk-specific growth processes and specifically identify proteins involved in the synthesis of stalk PG. Our results indicate that stalk biogenesis is driven by a specialized biosynthetic complex whose composition and biosynthetic activities are clearly distinct from those of the generic cell elongation and division machineries.
Interestingly, the stalk biosynthetic complex is a hybrid composed of factors typically associated with the elongasome (MreB, RodZ, RodA, PBP2) or divisome (DipM, SdpA, SdpB, CrbA) (Fig 14). The recruitment of components from the cell elongation machinery may reflect the need to incorporate new cell wall material into an existing sacculus to drive the elongation of the stalk structure, a process that may be mechanistically similar to dispersed PG biosynthesis during lateral cell growth. Notably, HADA (Fig 2A) and D-Cysteine [41] labeling clearly indicate that, during stalk growth, newly synthesized PG is not primarily detected in the basal stalk segment but rather in the adjacent polar regions of the cell body. This observation suggests that stalk elongation does not occur simply by addition of new material to the existing stalk template. Instead, it appears to be mediated through expansion of the stalk-proximal polar cap and its simultaneous remodeling into a new stalk segment, a process reminiscent of the medial growth and constriction of the PG sacculus during Caulobacter cell division. The common requirement for extensive PG remodeling may explain why the cell division and stalk biosynthetic complexes show a considerable overlap in their autolytic machineries. Interestingly, the importance of some of these shared components varies substantially between the two complexes. For instance, combined inactivation of the lytic transglycosylases SdpA and SdpB has no obvious effect on cell division [64], whereas it largely abolishes stalk formation, indicating that the functional context of these proteins varies depending on the process they mediate. It remains to be clarified to what extent the different cell wall biosynthetic complexes compete for their shared components. Interestingly, during the Caulobacter cell cycle, stalk growth occurs predominantly within a short time window at the transition from dispersed to medial peptidoglycan biosynthesis (S1 Fig). It is therefore tempting to speculate that the elongasome and divisome have higher priority in the recruitment of shared factors, thereby restricting assembly of the stalk biosynthetic complex to phases in which they are not fully active. Overall, stalk formation clearly demonstrates how the reshuffling of preexisting machinery can serve as a straightforward means to generate novel morphological features in bacteria (see also [80]). The striking diversity of cell shapes observed in certain lineages, such as the alphaproteobacteria, may therefore not be based on major new additions to the repertoire of cell wall biosynthetic proteins but rather on subtle changes in protein activities and localization patterns.
A key finding of our work is the central role of MreB in the stalk biosynthetic complex. We show that this cytoskeletal protein condenses at the stalked pole during phosphate starvation and facilitates the polar recruitment of several other factors that are critical to stalk formation. Notably, our attempts to integrate mCherry into a surface-exposed loop of MreB led to the serendipitous identification of a Caulobacter strain that was completely devoid of stalks under both phosphate-limiting and -replete conditions. This is in stark contrast to other mutants described previously, which are stalk-less in rich medium but still elaborate stalks upon phosphate starvation [43–46], suggesting that they have a defect in the regulation of stalk formation rather than in the biosynthetic machinery mediating this process. Importantly, apart from their failure to form stalks, cells producing the MreB sandwich fusion showed only mild general cell shape defects. The region surrounding the insertion site of mCherry may thus contain determinants that are specifically required for MreB’s function in stalk formation but largely dispensable for elongasome-mediated longitudinal growth of the cell body. Previous work has shown that the positioning of MreB filaments is strongly influenced by their intrinsic curvature [81, 82], a parameter controlled by the concentration of the membrane adapter RodZ [83]. The high enrichment of the MreB-RodZ complex at the stalked pole may thus be sufficient to change the architecture of MreB filaments such as to facilitate their interaction with the more highly curved stalked pole. However, the cues promoting the relocation of MreB from the lateral regions of the cell to the stalked pole still remain unknown.
Although MreB clearly has a key role in stalk biogenesis, it is not the only scaffolding protein contributing to this process. Previous work has shown that the bactofilin BacA is required for proper stalk length [50], and our analyses revealed an additional role for this protein in cell body elongation during phosphate starvation (Fig 9). Notably, the bacA gene lies in a putative operon with ldpA, a gene encoding a putative LytM-like endopeptidase that also functions in stalk formation. This genetic context is conserved in a variety of other species, suggesting a functional link between the two gene products [84, 85]. Support for this notion comes from studies in the human pathogen Helicobacter pylori, which demonstrated that both genes in this conserved operon are required to establish the characteristic helical cell shape of this species [84]. Notably, apart from its putative interaction with LdpA, Caulobacter BacA was shown to recruit a class A PBP (PbpC) involved in stalk elongation and in the targeting of proteins to the stalk lumen [50, 86]. Importantly, the polar localization of BacA was independent of the presence of MreB. The bactofilin cytoskeleton thus appears to constitute a functionally independent morphogenetic module that has been coopted by Caulobacter to modulate stalk formation. This module appears to act downstream of the MreB-dependent stalk biosynthetic complex, as it was not able to establish a stalk structure in the absence of a functional MreB cytoskeleton.
The ultimate determinant mediating the polar recruitment of the stalk biosynthetic machinery in Caulobacter still remains unknown. In Asticcacaulis excentricus, a member of the Caulobacteraceae that is characterized by subpolar stalks, the site of stalk formation was shown to be defined by the polarity determinant SpmX [87]. However, despite its conservation, this protein is not required for proper stalk localization in Caulobacter cells [88]. Notably, deletion of SpmX or transfer of the cells to phosphate-limited media restores polar stalk growth in A. excentricus [87], suggesting that the pathway observed in Caulobacter is still present in A. excentricus but normally obscured by the the action of the newly coopted localization factor SpmX. It will be interesting to see whether the A. excentricus SpmX homolog organizes an alternative stalk biosynthetic complex or simply recruits the polar machinery to a pole-distal position.
Although the functionality and localization of the peptidoglycan biosynthetic machinery changes drastically upon transition of Caulobacter cells from phosphate-replete to phosphate-limiting media, the overall composition of their PG layer remains largely unaffected. This finding is unexpected because significant changes in both glycan chain lengths and the degree of cross-linking were observed in other species in response to changes in their growth conditions [89]. However, analyzing the muropeptide profiles of isolated stalk and cell body fractions, we identified clear differences between these two compartments that are likely obscured in whole-cell analyses due to the small contribution of stalks to the total cellular PG content. Most importantly, stalk PG showed a significantly higher degree of crosslinkage, which was mostly due to a higher frequency of 3–3 crosslinks, indicative of elevated LD-TPase activity. The precise reason for this difference remains unclear. It is conceivable that the LD-TPases LdtD and LdtX are part of the polar stalk biosynthetic complex and, thus, preferentially act on newly synthesized PG produced by this machinery. However, localization studies did not give any evidence for an enrichment of these proteins at the stalked pole. An alternative explanation may be provided by the observation that the turnover rate of PG is significantly lower in the stalk than in the cell body. Thus, LD-TPases may act uniformly throughout the entire cell envelope, but most of the 3–3 crosslinks formed in the cell body may be lost as a consequence of PG remodeling, whereas those in the stalk are retained over prolonged periods of time. Notably, peptides with 3–3 crosslinks are stiffer than those with 3–4 crosslinks and adopt a more extended conformation that is better suited to connect glycan strands in stressed PG [90]. Their increased frequency may therefore help to modulate the mechanical properties of the stalk and render it more resistant to bending or breakage under conditions of high laminar flow [37, 91]
Collectively, our study shows that, in Caulobacter, multiple cell-wall biosynthetic machineries act in concert to generate stalks of proper size and stability, thereby ensuring optimal performance of this cellular structure in the environmental context. It will be interesting to see how the nature and the regulation of these components have changed during evolution to bring about the large variety of morphologies found in other stalked members of the alphaproteobacterial lineage.
Caulobacter strains [92] were grown at 28°C in peptone-yeast-extract (PYE) medium [35], supplemented with antibiotics at the following concentration when appropriate (μg ml-1; liquid/solid medium): spectinomycin (25/50), streptomycin (-/5), gentamicin (0.5/5), kanamycin (5/25), chloramphenicol (1/1), mecillinam (15/-). Gene expression from the xylX promoter (Pxyl) or vanA promoter (Pvan), was induced by supplementation of the media with 0.3% D-xylose and 0.5 mM sodium vanillate, respectively, prior to analysis of the cells. To induce phosphate starvation, stationary cells were diluted 1:20 in M2G-P medium [50] and incubated at 28°C for the indicated times. In case of the conditional mreB, rodZ, and mreC mutants, cells were grown to exponential phase (OD600 ~ 0.5) in PYE medium supplemented with xylose, washed three times, and then resuspended to an OD600 of 0.05 in inducer-free medium. The cultures were then grown for 7 h to achieve protein depletion, diluted (1:20) in M2G-P medium, and cultivated for additional 24 h before analysis. The conditional amiC and dipM mutants were treated in a similar fashion, with 12 h of cultivation in PYE medium prior to transfer into M2G-P. The synchronization of Caulobacter was achieved by density gradient centrifugation using Percoll (Sigma-Aldrich) [93]. To determine the viable-cell count in cultures, various dilutions of the cell suspensions were spread on PYE plates, and the number of colony-forming units (CFU) was determined after three days of incubation at 28°C. E. coli strain TOP10 (Invitrogen) and its derivatives were cultivated at 37°C in LB broth (Karl Roth, Germany). Antibiotics were added at the following concentrations (μg/ml; liquid/solid medium): spectinomycin (50/100), gentamicin (15/20), kanamycin (30/50), chloramphenicol (20/30).
The bacterial strains, plasmids, and oligonucleotides used in this study are listed in S5–S8 Tables. E. coli TOP10 (Invitrogen) was used as host for cloning purposes. All plasmids were verified by DNA sequencing. Caulobacter was transformed by electroporation. Non-replicating plasmids were integrated into the Caulobacter chromosome by single-homologous recombination at the xylX (Pxyl) or vanA (Pvan) locus [94]. Gene replacement was achieved by double-homologous recombination using the counter-selectable sacB marker (M.R.K. Alley, unpublished) [54]. Proper chromosomal integration or gene replacement was verified by colony PCR.
Cells were grown to exponential phase in PYE medium, harvested by centrifugation, and resuspended in the same medium to an OD600 of 0.05. The suspensions were then transferred to 24‐well polystyrene microtiter plates (Becton Dickinson Labware), incubated at 32°C with double‐orbital shaking in an Epoch 2 microplate reader (BioTek, Germany), and analyzed photometrically (OD600) at 15 min intervals.
For light microscopic analysis, cells were transferred onto pads made of 1% agarose. Images were taken with an Axio Observer.Z1 (Zeiss, Germany) microscope equipped with a Plan Apochromat 100x/1.45 Oil DIC and a Plan Apochromat 100x/1.4 Oil Ph3 phase contrast objective, an ET-mCherry filter set (Chroma, USA), and a pco.edge sCMOS camera (PCO, Germany). Images were recorded with VisiView 3.3.0.6 (Visitron Systems, Germany) and processed with Metamorph 7.7.5 (Universal Imaging Group, USA) and Illustrator CS6 (Adobe Systems, USA). To generate demographs, fluorescence intensity profiles were measured with ImageJ 1.47v (http://imagej.nih.gov/ij). The data were then processed in R version 3.5.0 [95] using the Cell Profiles script (http://github.com/ta-cameron/Cell-Profiles) [96]. Box and violin plots for the statistical analysis of imaging data were generated in R version 3.5.0 using the ggplot2 [97] and Reshape2 [98] packages, respectively.
10 μl cell suspension were applied to an electron microscopy grid (Formvar/Carbon Film on 300 Mesh Copper; Plano GmbH, Germany) and incubated for 1 min at room temperature. Excess liquid was removed with Whatman filter paper. Subsequently, the cells were negatively stained for 5 sec with 5 μl of 1% uracyl acetate. After three washes with H2O, the grids were dried, stored in an appropriate grid holder, and analyzed in a 100 kV JEM-1400 Plus transmission electron microscope (JEOL, USA).
Western blot analysis was performed as described [54], using anti-CtrA [99], anti-FtsZ [63], anti-MipZ [54], anti-DnaA [100], or anti-SpmX [88] at dilutions of 1:10,000 (anti-CtrA, anti-FtsZ, anti-MipZ, and anti-DnaA), and 1:50,000 (anti-SpmX). Goat anti-rabbit immunoglobulin G conjugated to horseradish peroxidase (Perkin Elmer, USA) was used as secondary antibody. Immunocomplexes were detected using the Western Lightning Plus-ECL chemiluminescence reagent (Perkin Elmer, USA). Signals were recorded with a ChemiDoc MP imaging system (Bio-Rad) and analyzed using the Image Lab 5.0 software (Bio-Rad).
HADA-staining experiments were conducted as described [42]. Briefly, 50 μl of a culture were incubated for 2 min with 0.5 mM HADA. The cells were then fixed by addition of ice-cold ethanol to a concentration of 70% and incubated at 4°C for 20 min. Subsequently, they were washed three times with PBS and subjected to fluorescence microscopic analysis. For chase experiments, phosphate-starved Caulobacter cells were grown for 90 min in the presence of 0.5 mM HADA. The cells were washed three times with M2G-P medium, resuspended in fresh M2G-P medium, and further cultivated for the indicated time intervals. Cells were fixed and washed as described above prior to imaging.
Protein sequences containing the indicated domains were retrieved from the UniProt Knowledgebase [101]. Their overall domain composition was determined using the SMART server [102]. The prediction of protein localization and membrane topology was performed with Signal-BLAST [103] and TMHMM [104], respectively.
Cultures were grown in the indicated media and supplemented with 20 μg/ml rifampicin 3 h prior to analysis to block the re-initiation of chromosome replication. At the indicated time points, cells were diluted to an OD600 of 0.1–0.2, incubated for 25 min under vigorous shaking with the DNA-specific fluorescent dye Hoechst 33342 (10 μM; Thermo Fisher Scientific, Germany), and fixed by addition of ethanol to a final concentration of 70%. Subsequently, the suspensions were analyzed by flow cytometry in a customized Fortessa Flow Cytometer (BD Biosciences, Germany), using the UV 440/40 nm channel. Data were acquired with FACSdiva 8.0 (BD Biosciences) and processed with FlowJo v10 (FlowJo LLC, USA).
For whole-cell analyses, cultures were rapidly cooled to 4°C and harvested by centrifugation at 16,000 rpm for 30 min. The cells were resuspended in 6 ml of ice-cold H2O and added dropwise to 6 ml of a boiling solution of 8% sodium dodecylsulfate (SDS) that was stirred vigorously. After 30 min of boiling, the suspension was cooled to room temperature. Peptidoglycan was isolated from the cell lysates as described previously [105] and digested with the muramidase cellosyl (kindly provided by Hoechst, Frankfurt, Germany). The resulting muropeptides were reduced with sodium borohydride and separated by HPLC following an established protocol [105, 106]. The identity of eluted fragments was assigned based on the retention times of known muropeptides from Caulobacter [107].
To prepare stalk and cell body fractions, 100 ml cultures grown in M2G-P medium were rapidly cooled to 4°C and harvested by centrifugation at 16,000 rpm for 30 min. After resuspension in M2G-P medium, the cells were vigorously agitated for 2 min at maximum speed in a kitchen blender. The suspension was submitted to three rounds of centrifugation at 9,000 rpm and 4°C. The supernatants (stalk fraction) and the first pellet (cell body fraction) were collected separately and kept in ice. The stalk fraction was subjected to an additional centrifugation step at 10,000 rpm and 4°C to remove residual cell bodies and cell debris. Subsequently, stalks were collected by centrifugation at 20,000 rpm and 4°C for 30 min, resuspended in 3 ml ice-cold H2O, added dropwise to 3 ml of a boiling 8% SDS solution, and then further processed as described above to isolate stalk PG. The isolation of cell body PG was achieved as described for whole-cell samples.
|
10.1371/journal.pcbi.1003057 | Two Misfolding Routes for the Prion Protein around pH 4.5 | Using molecular dynamics simulations, we show that the prion protein (PrP) exhibits a dual behavior, with two possible transition routes, upon protonation of H187 around pH 4.5, which mimics specific conditions encountered in endosomes. Our results suggest a picture in which the protonated imidazole ring of H187 experiences an electrostatic repulsion with the nearby guanidinium group of R136, to which the system responds by pushing either H187 or R136 sidechains away from their native cavities. The regions to which H187 and R136 are linked, namely the C-terminal part of H2 and the loop connecting S1 to H1, respectively, are affected in a different manner depending on which pathway is taken. Specific in vivo or in vitro conditions, such as the presence of molecular chaperones or a particular experimental setup, may favor one transition pathway over the other, which can result in very different monomers. This has some possible connections with the observation of various fibril morphologies and the outcome of prion strains. In addition, the finding that the interaction of H187 with R136 is a weak point in mammalian PrP is supported by the absence of the residue pair in non-mammalian species that are known to be resistant to prion diseases.
| Transmissible spongiform encephalopathies, which include the “mad cow” disease and the Creutzfeldt-Jakob disease, are related to the abnormal folding of a host protein termed the prion protein (PrP). Many aspects of the underlying molecular mechanism still remain elusive. Among the hypotheses that have been put forward in the past few years, it has been suggested that PrP could be destabilized by the protonation of a specific residue, H187, when the protein passes through acidic cell organelles. We have modeled PrP at the atomistic level, with the neutral and protonated forms of H187. Our simulations show that the destabilization process can follow two alternative pathways that could lead to different final structures. This discovery may shed some light on one of the most puzzling aspect of prion diseases, the fact that they exhibit various strains encoded in the structure of misfolded PrP. In addition, the atomistic details provided by our model highlight a key interactions partner in the destabilization process, R136. The residue pair is not present in non-mammalian species that do not develop prion diseases.
| The misfolding of the prion protein (PrP), which is a key aspect of transmissible spongiform encephalopathies (TSE), has been the subject of intense research during the past decades. Nonetheless, little is known about the underlying molecular mechanism. One serious hurdle remains the determination of the structure of the resulting misfolded isoform () [1]. As a consequence, various models have been suggested with substantially different packing arrangements and monomer structures, and a consensus about the structure of is far from being reached [1]. A particular subject of controversy is about the actual region of PrP that undergoes a deep refolding during the PrP conversion. According to the so-called “spiral” [2] and “” [3], [4] models, extended are formed in the N-terminal region and at the beginning of the C-terminal domain up to H1 (H1 is kept intact in the former and is refolded in the latter model). However, it has been recently shown that the H2H3 core is also highly fibrillogenic by itself [5], [6]. Finally, it has also been suggested that could be entirely refolded in an in-register extended [7].
Many in vitro [5], [8]–[11] and computational [2], [12]–[16] studies have tackled this issue using acidic conditions. They have consistently shown that low pH destabilizes PrP and favors its misfolding. This represents biologically relevant conditions insofar as endosomal organelles, whose typical pH is about 5 but can be as low as 4.3 [17], have been highlighted as possible locations for growth [18]–[20]. Importantly, mammalian PrP contain one slightly buried residue, H187, that titrates right in the range of endosomal pH [11], [13]. Several lines of evidence indicate that its protonation [13], or more generally the addition of a positive charge at site 187 [11], destabilizes the protein fold.
Whereas many theoretical studies have been performed on the globular C-terminal domain (residues 121–231 using the numbering of the human sequence) of mouse PrP (mPrP, Fig. 1-A), it is worth noting that the cellular form of PrP () also contains a long unstructured N-terminal tail (residues 23–120) [21]–[29], a glycosylphosphatidyl-inositol (GPI) anchor [30]–[32] and can be mono or diglycosylated [27], [33]. Nevertheless, previous MD simulations have suggested that the structure and dynamics of the globular domain of is rather independent of the anchoring to the membrane and the glycosylation [34]. In addition, our previous study of the misfolding propensity of mPrP using extensive REMD simulations [16] has revealed that various monomers can be formed from the C-terminal domain alone, which is also consistent with the results of Ref. [5], [6].
Here, we have performed microsecond MD simulations of the structured C-terminal domain of mPrP at pH 4.5, which corresponds approximately to the lowest pH value observed in endosomes [17]. To this end, we assigned the protonation state of all titrable residues with the program PROPKA [35] (see also Materials and Methods section). The only buried residue for which the protonation state cannot be uniquely assigned is H187. The quantitative evaluation of its is challenging, because the protonation/deprotonation of a buried residue usually affects the protein structure drastically [36], [37]. Nevertheless, several semi-quantitative estimates of the of H187 have been obtained [13], [38] and they all indicate that mPrP coexists in both, neutral and protonated-H187 forms at pH 4.5. Thus we have performed two sets of acidic pH simulations, with H187 in either its neutral or protonated form. It is worth noting that other residues in mPrP also titrate at pH 4.5. However, they are all located at the protein surface, so that their electrostatic effect on the global structure of the protein is much less important than that of H187. Thus, we have considered only one protonation state for these residues (see Materials and Methods section).
Our micorsecond simulations show that the mechanism of mPrP destabilization upon protonation of H187 involves R136 as a key partner (Fig. 1-B,C). There is an electrostatic repulsion between the imidazole ring of the protonated H187 () and the guanidinium group of R136 (), to which the system responds by pushing away either or . Because R136 and H187 belong to two very different structural regions of the protein, namely the loop connecting S1 to H1 () and the C-terminal part of H2 (H2(Cter), Fig. 1-A), the effect on the structure is different depending on which of the two transition routes is taken. It is possible that specific in vivo or in vitro conditions may favor one route over the other, which could lead to completely different structures. Our findings thus seems to provide some rational to the various conclusions reached by different authors regarding the actual region of the protein that is refolded upon misfolding.
Fig. 2 shows the effect of protonating H187 on the backbone of mPrP. The structure is very stable and remains close to the NMR structure when H187 is neutral, whereas simulations with the protonated H187 exhibit important backbone fluctuations and reorganizations. As depicted in Fig. 2-C, these enhanced fluctuations are mainly located in two specific regions of the protein, namely H2(Cter), which hosts H187, and .
Fig. 2-E shows that the protonation of H187 induces a drastic change in the free energy surface. The projection of the free energy on the of H2(Cter) and shows a single minimum when H187 is neutral, which corresponds to the native structure of PrP, and a complicated multiple minima landscape when H187 is protonated. The new free energy basins are located Å away from the native basin, thus corresponding to substantial conformational changes.
The two example snapshots provided in Fig. 2-B,D show that this reorganization is accompanied by a significant modification of the secondary structure of the protein. We will provide a more detailed analysis of the secondary structure changes later in the following sections. For the time being, it is interesting to rationalize how the perturbation that is introduced at one side of the protein (the protonation of H187 located in H2(Cter)) is transmitted through the macromolecule and affects strongly the structure at the opposite side ().
In order to understand the mechanism by which the protonation of H187 induces the reorganization of the protein structure, it is necessary to have a closer look to the environment of H187 in PrP. It is particularly interesting to focus on nearby charged residues because they are expected to play a major role in the reorganization of the protein when H187 gets a positive charge upon protonation. In the NMR structure of mouse PrP, the closest charged residues are R136, R156, K194, E196 and D202 (Fig. S5-A). R136 is somewhat isolated in terms of proximity with charged residues other than H187 (when protonated), whereas K194, E196, R156, and D202 form a network of salt-bridge interactions. These four latter residues have been pointed out has possible key residues in the misfolding of PrP [13], [39]. As shown in Fig. S5-B, our simulation provide a consistent picture with that of Ref. [13], because the protonation of H187 leads to the disruption of the salt bridge between E196 and R156 and the transient formation of a new salt bridge between the protonated E196 and H187, while a tight salt bridge is maintained between R156 and D202 (K194 is highly solvated, independently of the protonation state of H187, and never interact strongly with E196). Nevertheless, the fact that R136 does not have any close alternative partner makes it more sensitive to the positive electric field created by the protonated H187, as we shall see in the next section.
The observation of structural rearrangements in , which is located far from H187, has motivated us to perform a thorough analysis of the mobility of each residue in this region. It turns out that R136 is a key partner of H187 in the destabilization of mammalian PrP upon protonation of H187. In the NMR structure of mPrP (Fig. 1), and are about 8 Å apart and loop (residues 154–158) is located in between. is stabilized by a series of dipole-charge interactions with four peptide bonds while is H-bonded to one carbonyl group and establishes van der Waals contacts with the ring of P158 (Fig. 1-B).
Because of the proximity of and in the native structure of mPrP, the protonation of the former should induce an electrostatic repulsion between the two groups. A discussion of the corresponding energetics is provided in Text S1. Fig. 3 shows the effect of the protonation of H187 on the position of (or ) and . When H187 is neutral, and are mostly located in their respective native cavities, whereas they cover a much wider portion of conformational space upon protonation of H187. We define four conformational states according to the position of (or ) and inside or outside their respective native cavities. To do so we consider the bivariate histogram of the distances between (or ) and from their respective cavities (Fig. 3-C). The conformational state in which both groups stay close to their original location will be termed , and we define states and according to the departure of or , respectively. Interestingly, the state is almost not populated. The picture that is the most consistent with these data is that PrP exhibits a dual response to the protonation of H187, by pushing away either or (but not both at the same time), thus decreasing the electrostatic repulsion between them. Because H187 and R136 are attached to H2(Cter) and , respectively (Fig. 1-A,B), the local reorganization of either or affects the global structure of these two regions (Fig. 2). We stress that, once H187 is protonated, the dynamics of the system proceeds smoothly through a series of locally thermalized states giving rise, in a reproducible way, to either the or state. Fig. S6 and S7 show that and remain in their native pockets during at least 100 ns before one of the two moves out.
A similar electrostatic repulsion can be expected for the H187R mutation, for which the positive charge of the introduced arginine has been suggested to destabilize the overall fold of human PrP [11], [40], [41]. An interesting aspect of this finding is that none of the non-mammalian PrP exhibit this specific spatial arrangement (Fig. S4). In other words, these non-mammalian proteins do not have this pH-sensitive “weak point” in there structure and this probably explains the fact that non-mammalian species do not exhibit TSEs.
Due to the buried character of H187 and the fact that its protonation induces a substantial modification of the protein structure, the quantitative evaluation of its (and the corresponding contributions of other residues) during the misfolding is challenging [36], [37]. Nevertheless, PROPKA calculations [35] provide physically sound estimates that can help to rationalize the underlying physics. Such calculations for representative snapshots of our simulations are provided in Fig. S8. The of H187 is systematically shifted up as soon as the protein starts to misfold, independently of the pathway ( vs ) that is taken. This is in agreement with the fact that the proximity with the positive charge of in the native structure of mPrP induces a down-shift of the of H187 (this is supported by the fact that our PROPKA calculations report R136 as a key residue in the electrostatic environment of H187, see the corresponding PROPKA output file for a representative structure in Dataset S1). As soon as moves out of its cavity ( state) the electrostatic repulsion between and decreases and the protonated form of H187 becomes much more stable ( shifted up). When the protein adopts a state, is much more solvated by water and the of H187 approaches the corresponding value in water ().
The positioning of (or ) and has a strong influence on the length of the S1,S2 , as shown in Table 1. Typically, both and states correspond to structures with a short native , while the state is characterized by the preference of an elongated . This is illustrated by the simulation depicted in Fig. 4. At the beginning of the simulation, the protein is in its native conformation. As depicted in the insets of Fig. 4, the native location of at is a key aspect of the protein fold because it forms a sort of “clip” that forces the backbone to remain packed against the rest of the protein (Fig. 1-A) in a specific conformation. The permanent departure of out of its cavity at ns induces an important release of backbone constraint and the system is consequently more prone to reorganize in this region. Then the system relaxes during about 400 ns, and and come close together. The number of hydrogen bonds between the two strands increases concomitantly and the elongates (Fig. 4-A,B).
As shown in Fig. 5, both and states are characterized by an unraveling of H2(Cter). However, the underlying mechanisms (and the corresponding transition pathways) differ substantially. The portion of the helix that undergoes an unraveling is represented by a dashed purple arrow in Fig. 1-A (see also the example snapshot depicted in Fig. 2-D).
The departure of ( conformation) out of its cavity obviously destabilizes H2 because the helix looses a key tertiary contact with loop (Fig. 1-A). The unraveling of H2(Cter) in the state has its roots in the polar interactions of with the nearby residues. A closer look to the shape of the cavity (Fig. 1) reveals that it is a narrow groove at the bottom of which lies the carbonyl group of T183. The contact analysis shown in Fig. 6 reveals that the neutral is H-bonded to R156 only, consistent with the NMR structure of mPrP [21], whereas new contacts are formed with the CO group of T183 when H187 is protonated. A key aspect of these extra contacts is that they involve not only the group of H187, but also the group. They reflect dipole-charge interactions between the extra positive charge of and the dipole moments of the 156–157 and 182–183 peptide bonds. In other words, the imidazole ring can take two conformations around the bond and still maintain a significant interaction with one of the two nearby backbone CO groups, which results in four stable conformations inside the pocket.
The formation of new contacts between and T183(CO) has two effects that explain the loss of helical character in H2(Cter). First of all, it weakens the tertiary contact between H2(Cter) and . Second, the native intra-helix H-bond between T183(CO) and H187(NH) is lost. The tighter the interaction between and T183(CO) the weaker the local stability of H2.
In this paper we have shown that the protonation of H187 in mPrP at pH 4.5, which corresponds approximately to the lowest pH observed in endosomes [17], leads to extensive conformational changes on the microsecond time scale. The picture that emerges from our simulations is that the protonation of H187 leads to an electrostatic repulsion between the positive charges of and , which results in conformational transitions in the regions to which H187 and R136 are linked, namely H2(Cter) and respectively.
Our findings hence highlight two possible routes for PrP misfolding with either the unraveling of H2(Cter) alone ( route) or the unraveling of H2(Cter) with simultaneous elongation of S1,S2 ( route). This dual behavior seems to reconcile the various observations and proposals that have been made regarding the actual PrP region that undergoes a deep refolding upon conversion to [2]–[5]. It is indeed possible that a particular computational or experimental setup favors one of the or substates at the beginning of the misfolding process. Such conformational shift could be assisted in vivo by molecular chaperones such as polyanionic molecules [1], [42]. This variability in misfolding pathways may also be connected to the fact that prion exhibits a variety of strains, because it is believed that changes in conformations of encodes for strain properties [30], [43], [44].
Finally, it is interesting to note that the pattern is not present in those non-mammalian species who are known to resist to TSEs. This is a possible explanation for the observed resistance to TSEs in these species.
All simulations were started from the NMR structure of mPrP published by Riek et al. (PDB code 1AG2). We aimed at modeling mPrP with a neutral or protonated H187 at pH 4.5.
The protonation state of titrable residues apart from H187 was first estimated from PROPKA [35] calculations. The protonation state of most of them can be determined without ambiguity. All buried or semi-buried residues other than H187 are all aspartic or glutamic acids whose side chains are hydrogen-bonded to other groups in the protein. This has the effect to shift up their above the typical values of that they adopt in water, i.e. significantly above the pH we want to model. Hence they are expected to be protonated. The solvent-exposed histidines are expected to exhibit a of so they can be considered protonated at a pH of 4.5. The remaining solvent-exposed aspartic or glutamic acids are more ambiguous because their is close to the pH we want to model. Nevertheless, their solvent-exposed character makes them much less important for the global fold of the protein. We chose their protonation state according to the estimated with PROPKA [35]. The relevance of this choice was verified a posteriori by observing that the fold of the protein is very well conserved over microsecond simulations with a neutral H187.
Two topologies (one with H187 neutral and one with H187 protonated) were built with the GROMACS 4.0.7 [45]–[47] suite of programs. For each of them, the protein was immersed in a rhombic dodecahedral water box. The size of the box was chosen so that the distance between the protein and the edge of the box was Å. The system was neutralized by adding 2 or 3 chloride counterions (depending on the protonation state of H187). The resulting system contained about 30000 atoms.
The AMBER99SB force field [48] was used to describe the protein and the TIP3P model [49] was employed for the water molecules. The force field was included in GROMACS thanks to the ports provided by Sorin and coworkers [50], [51]. The particle mesh Ewald method [52] together with a Fourier grid spacing of 1 Å and a cutoff of 12 Å was used to treat long-range electrostatic interactions. A cutoff of 12 Å was used for van der Waals interactions. The water box was first relaxed by means of NpT simulations with restraints applied to the positions of the heavy atoms of the protein. Then the system was optimized in a series of energy minimization runs in which the restraints on the protein were progressively removed. Finally, we run eight simulations with a time step of 2 fs. Three and five of them were conducted with a neutral or protonated H187, respectively.
Each simulation was initiated with a set of velocities taken at random from a Maxwell-Boltzmann distribution corresponding to a temperature of 10 K. Then the system was heated up to 300 K in 300 ps using two Berendsen thermostats [53] (one for the protein and one for the solvent) with a relaxation time of 0.1 ps each. The simulation was prolonged for 100 ps and the Berendsen barostat with a relaxation time of 2 ps was switched on during 100 ps. Finally, we switched to production phase using a Nose-Hoover [54], [55] thermostat and a Parrinello-Rahman barostat [56] with relaxation times of 0.5 and 10.0 ps, respectively. The total simulation lengths were 1.9, 1.3 and 1.6 for simulations with a neutral H187, and 1.9, 1.5, 1.6, 1.2 and 1.2 for simulations with a protonated H187. The plot of each simulation is provided in Figure S1.
All the representations were done with the program VMD [57]. Secondary structure assignments were done using the STRIDE algorithm [58].
|
10.1371/journal.ppat.1004574 | Neutral Sphingomyelinase in Physiological and Measles Virus Induced T Cell Suppression | T cell paralysis is a main feature of measles virus (MV) induced immunosuppression. MV contact mediated activation of sphingomyelinases was found to contribute to MV interference with T cell actin reorganization. The role of these enzymes in MV-induced inhibition of T cell activation remained equally undefined as their general role in regulating immune synapse (IS) activity which relies on spatiotemporal membrane patterning. Our study for the first time reveals that transient activation of the neutral sphingomyelinase 2 (NSM2) occurs in physiological co-stimulation of primary T cells where ceramide accumulation is confined to the lamellum (where also NSM2 can be detected) and excluded from IS areas of high actin turnover. Genetic ablation of the enzyme is associated with T cell hyper-responsiveness as revealed by actin dynamics, tyrosine phosphorylation, Ca2+-mobilization and expansion indicating that NSM2 acts to suppress overshooting T cell responses. In line with its suppressive activity, exaggerated, prolonged NSM2 activation as occurring in co-stimulated T cells following MV exposure was associated with aberrant compartmentalization of ceramides, loss of spreading responses, interference with accumulation of tyrosine phosphorylated protein species and expansion. Altogether, this study for the first time reveals a role of NSM2 in physiological T cell stimulation which is dampening and can be abused by a virus, which promotes enhanced and prolonged NSM2 activation to cause pathological T cell suppression.
| Though the ability of measles virus (MV) to impair T cell activation has long been known, it is mechanistically not well understood. We have shown earlier that MV can contact dependently trigger activation of sphingomyelinases which is known to affect compartmentalization of membrane lipids and proteins. Because these are particularly important in the activity of the immune synapse (IS), we investigated whether MV-induced sphingomyelinase activity would interfere at that level with T cell activation. Our study for the first time revealed that the neutral sphingomyelinase 2 (NSM2) is transiently activated in primary T cells by co-stimulation through CD3 and CD28, and that this does occur to dampen early T cell responses. The virus appears to exploit this inhibitory activity of the enzyme to suppress T cell activation by promoting an enhanced and prolonged NSM2 activation. These findings do not only assign a hitherto novel role of the NSM2 in regulating T cell responses, but also reveal a novel strategy for viral T cell suppression.
| Plasma membrane ceramides are released in response to activation of sphingomyelinases and condense into large platforms which alter biophysical properties of the cell membrane. In addition to other stimuli, ligation of certain surface molecules, also including death receptor family members and viral attachment receptors, efficiently activates neutral and/or acid sphingomyelinase (NSM or ASM, respectively) followed by ceramide release (reviewed in [1]–[3]). Ceramide enriched membrane microdomains act to regulate sorting of membrane proteins and their signalosomes, and this affects a variety of biological responses including lateral and vertical receptor segregation as particularly relevant for pathogen uptake, apoptosis, cell motility and proliferation [3]–[6].
Measles virus (MV) causes profound generalized immunosuppression and interference with T cell viability, expansion and function is one of its major hallmarks. A plethora of findings supports the interpretation that MV is acquired and transferred by CD150+ antigen-presenting cells to the secondary lymphatic tissues where it can be transmitted to and deplete CD150+ lymphocytes, especially memory T cells [7]–[9]. Though being infected to a very limited extent, peripheral blood cells of patients, however, are generally refractory to expansion driven by polyclonal and antigen-specific stimulation, implying they had been paralysed by mechanisms independently of direct infection. In line with this hypothesis, exposure of uninfected lymphocytes to UV-inactivated MV or the MV glycoprotein complex (gpc) was sufficient to induce their arrest in vitro and in vivo [10]–[13]. For this, the gpc interacts with an as yet unknown receptor on the surface of T cells (which is not identical to CD150 [14]) to abrogate relay of T cell receptor (TCR) signaling at the level of the phosphatidyl-inositol-phosphate-3-kinase (PI3K) and its downstream effectors, Akt kinase, Vav1, Rac1 and Cdc42 [15]–[17]. Because these are also major regulators of actin cytoskeletal dynamics, MV contact induced physical T cell paralysis is reflected by collapse of actin based protrusions and loss of polarity and motility on fibronectin [15]–. Thus, by targeting the PI3K, MV abrogates activation of downstream effectors essentially mediating S-phase progression, but also actin dynamics which is of key importance in organizing the functional architecture of the immune synapse (IS) where T cell signaling is initiated and sustained [19].
In order to interfere with TCR signaling, MV has to initiate signaling upon binding to T cells itself. This involved sequential activation of NSM2 (the NSM species abundant at the plasma membrane) and ASM, which almost entirely accounted for MV interference with actin cytoskeletal integrity and dynamics [18], though the role of sphingomyelinase activation on T cell activation remained unknown.
The role of ceramide release in regulating T cell activation by CD3/CD28 co-stimulation is unclear. Ligation of either CD3 or CD28 alone caused activation of NSM or ASM, respectively [20], [21]. CD3-mediated NSM activation was dispensable for TCR-induced overall tyrosine phosphorylation, however, required for IL-2 production and MAPK activation [20], while ASM-mediated ceramide release was important for CD28-dependent NF-κB activation [21]. In the latter study, exogenous supply of ceramides fully replaced ASM activity in co-stimulation. In contrast, other studies indicated an inhibitory activity of ceramide in co-stimulating CD3 signaling [22]–[24], and ceramide metabolites are of low abundance in CD3 associated domains immuno-isolated from T cells following activation [25]. This indicates that sphingomyelin breakdown, if occurring at all, has to be tightly controlled and compartmentalized at the level of the IS, and, if aberrantly induced, might translate into T cell inhibition.
With the present study we addressed activation of sphingomyelinases and subcellular distribution of ceramide accumulation upon CD3/CD28 co-ligation. Strikingly, co-stimulation entirely abrogated ASM activation while caused an early rise in NSM activity. This proved to be important for dampening thresholds of T cell activation because cell spreading as required for formation of IS interaction platforms and overall tyrosine phosphorylation were initiated earlier and were enhanced upon NSM knockdown as was expansion of T cells. Enhancement of sphingomyelinase activity upon pre-exposure to MV was associated with accumulation of ceramide enriched membrane domains at the synaptic interface. Importantly, MV induced NSM activation substantially accounted for loss of spreading responses and partially for MV interference with TCR-induced tyrosine-phosphorylation and expansion. Altogether, our findings indicate that the ability of NSM to dampen the threshold of T cell activation is exploited by MV for T cell suppression.
CD28 ligation causes ASM activation followed by ceramide release [21], which can be inhibitory to T cell stimulation [23], [24]. We therefore investigated whether CD3/CD28 co-stimulation would also promote ASM activation in primary T cells. Using a planar system (plate bound antibodies) for stimulation, ligation of CD28 alone (but not that of CD3) expectedly induced ASM activity which peaked after 15 min. This, was, however, completely abrogated upon CD3/CD28 co-stimulation where ASM activity did not significantly increase during the observation period (Fig. 1A). Activated ASM is the major source of extrafacial ceramide, and consequently, ceramide was displayed at the cell surface following ligation of CD28, but not of CD3 or CD3/CD28 co-ligation on T cells as detected by flow cytometry. This again supported the interpretation that CD28-mediated ASM activation is abolished upon co-stimulation (Fig. 1B, and S1A Fig.). Because NSM2 activation after CD3 ligation and cross-regulation of sphingomyelinases have been described earlier [18], [20], [26], we analyzed whether NSM2 would be involved in ASM inhibition by knocking down NSM2 in T cells by specific siRNA (further referred to as NSMKD T cells) (Fig. 1C). In contrast to what has been reported for other cell types [26], NSM knockdown did not augment basal ASM activity (S1B Fig.). As measured in co-stimulated untransfected T cells (Fig. 1A), the ASM activity remained at or below basal levels in control siRNA transfected cells (further refered to as CTRL T cells) (Fig. 1D, left panel). In NSMKD T cells, biphasic activation of the enzyme occurred to levels comparable to those measured upon CD28 ligation (Fig. 1A and 1D, left panel) indicating that NSM activity was required to extinguish that of ASM in co-stimulation. ASM activity was enhanced in NSMKD T cells, however, accumulation of extrafacial ceramides was not and rather remained at levels below that of CTRL T cells (Fig. 1D, right panel). In line with previous findings [20], early NSM activation in response to CD3 ligation did occur in untransfected T cells, and did not significantly differ from that induced upon co-stimulation, while a late peak of NSM activation was only seen after CD3 ligation (Fig. 1E). As revealed by direct comparison of activities of both enzymes, ASM activation does not occur in co-stimulated T cells while NSM was transiently activated clearly exceeding background levels, retained within 15 minutes and dropped thereafter (Fig. 1F). To the best of our knowledge, cross-regulation of ASM and NSM upon co-stimulation of receptors individually promoting their activation has not been previously observed.
Cross-regulation of ASM and NSM activity upon co-stimulation implies that activation of these enzymes and ceramide release have to be limited, tightly controlled and eventually compartmentalized during T cell activation. To assess the latter question directly, we co-detected ceramide with f-actin in T cells seeded onto co-stimulatory slides after 10 min. After this period, cells had spread on the planar support with typical IS formation where lamellum (corresponding to the pSMAC where centripetal transport of microclusters occurs), lamellipodium (corresponding to the dSMAC with particularly high actin dynamic reorganization) (as detected by f-actin, Fig. 2A) and the actin-free cSMAC (where TCR activity is terminated) can be visualized (Fig. 2A). Most interestingly, ceramide-enriched membrane domains were virtually absent from the lamellipodium, while they were readily detectable within the lamellum (Fig. 2A). Analyses addressing the subcellular distibution of NSM2, the enzyme most likely involved in generating ceramides at the IS (Fig. 1F), could not be performed in fixed cells using antibodies because commercially available NSM-antibodies revealed a broad reactivity in Western blot analyses rendering them unsuitable to obtain reliable staining patterns. We therefore analyzed NSM2 distribution following nucleofection of GFP-tagged neutral sphingomyelinase [27]. As for ceramide, NSM2-GFP appeared to localize mainly to the lamellum (Fig. 2B). However, the enzyme was not entirely absent from the lamellipodium and the cSMAC (Fig. 2B, 3D reconstruction). Also of note, although the protein expressed well in other cell types ([27], [28] and 293 cells, not shown), expression levels of NSM2-GFP in T cells were low Fig. 2B, second panel). Altogether, these data reveal that NSM, but not ASM activation does occur early during co-stimulation, however, ceramide release is spatially confined and excluded from areas of high actin activity.
Early NSM activation during co-stimulation suggested that the enzyme may have a role in T cell activation. To address this, we comparatively analyzed parameters important in early T cell activation in CTRL and NSMKD T cells (which did not detectably differ with regard to surface expression of CD3 or CD28, S1C Fig.). When seeded onto co-stimulatory slides, NSMKD T cells more efficiently adhered and spread than CTRL T cells already after 5 min (Fig. 3A). 10 min following activation (when comparable amounts of CTRL T cells had adhered) NSMKD T cells revealed an enhanced spreading response acquiring extended cell areas framed by actin-dense lamellipodial extensions. Moreover, they detectably polarized (usually representing late stages of activation) giving rise to extended protrusions 20 or 60 min following activation (Fig. 3B). These observations were confirmed by scanning electron microscopy analyses, where NSMKD T cells acquired a flattened, lamellar appearance tightly interacting with the planar support already 10 min following stimulation, which was seen in CTRL T cells only after 60 min. Remarkably, at that time NSMKD, but not CTRL T cells, had developed veil like protrusions indicative for acquisition of a more motile phenotype (Fig. 3C, arrows, and S2A Fig.). This indicates that T cell activation occurs more rapidly upon NSMKD and is associated with enhanced actin cytoskeletal activity. Moreover, accumulation of tyrosine phosphorylated protein species (p-tyr), p-ERK and p-Akt was accelarated and enhanced in NSMKD T cells as was initiation and magnitude of Ca2+-fluxing indicating that NSM depletion facilitated early T cell activation (Fig. 3D and, for overall p-tyr, Fig. 6A). Importantly, this also translated into proliferative responses of NSMKD T cells which expanded significantly more efficient in response to CD3/CD28 ligation than CTRL T cells. As seen for spreading responses, NSMKD also appeared to especially support early expansion of co-stimulated T cells (Fig. 3E, left panel, 48 h). As for human cells, proliferation of splenocytes isolated from Smpd3 deficient fro/fro mice driven by syngenic, superantigen-loaded bone marrow derived DCs was significantly enhanced as compared to that of Smpd3 sufficient littermates (Fig. 3E, right panel). In contrast to expansion, neither release of cytokines (IL-2, IL-4, IL-5, IL-10, IFN-γ or TNF-α) 4, 10, 24 or 72 h following α-CD3/CD28 stimulation or intracellular accumulation of IL-2, IL-10, IFN-γ or IL-17α) following a 4 h restimulation were detectably affected by NSM knockdown in human T cells (not shown). Altogether, these observations suggest that NSMKD facilitates initiation of T cell activation and therefore, NSM activity acts to dampen early T cell activation thresholds.
If NSM activity regulates the initiation threshold of physiological T cell activation, conditions additionally enhancing NSM activity could possibly further dampen T cell activation by promoting timely or spatially aberrant ceramide release. MV is known as an efficient inhibitor of T cell activation and its ability to cause sequential NSM/ASM activation in these cells has been established by us earlier [18]. In line with our previous findings, MV caused NSM activation in T cells (Fig. 4A). When compared to NSM activity induced upon co-stimulation alone, that induced upon additional MV exposure was elevated and persisted indicating that MV supports exaggerated and sustained NSM activation during T cell activation. Because studies involving bacterial sphingomyelinase (bSMAse) or short to middle chain ceramides suggested an inhibitory activity of ceramides in T cell stimulation [23], [24], we assessed whether augmented NSM activation would affect subcellular redistribution of ceramides. When exposed to bSMase or MV (both of which caused comparable ceramide release within 20 min, S3A Fig.), T cells barely adhered and spread on co-stimulatory slides (S2B Fig.). Because of the substantial differences in cell areas, it was not possible to evaluate whether the apparent central concentration of ceramide clusters in bSMase or MV pre-exposed cells resulted from condensation of the contact plane or mis-localization of ceramides (Fig. 4B). As an alternative approach, we analyzed ceramide accumulation at interfaces formed between T cells pre-exposed to MV and co-stimulatory beads coated with CD3/CD28-specific antibodies. In line with our observations made in the planar system (Fig. 2A and 4B), ceramides were largely excluded from the center of the interfaces formed with MOCK treated T cells, and this was unaffected by NSMKD (Fig. 4C). As revealed for adhesion to other supports, MV pre-exposure impaired adhesion of CTRL T cells and thereby the frequency of conjugates with beads (not shown). A substantial fraction of conjugates, that did form with MV-exposed CTRL T cells, however, did not efficiently exclude ceramide from the interface, which was, however, entirely corrected for in NSMKD T cells (Fig. 4C). Thus, MV-induced NSM activation is followed by mis-compartmentalization of ceramide within the IS. To address whether NSM is also excluded from the IS center in this system, and if so, wether this is altered in MV-exposed T cells, we studied distribution of the enzmye in bead assays involving NSM-GFP expressing T cells. In MOCK treated cells, NSM-GFP was detected in association with the plasma membrane, but also with intracellular, presumably Golgi, compartments as described previously for MCF-7 cells [28] (Fig. 4D). Interestingly, these efficiently polarized to the distal pole of the T cell while NSM-GFP was mainly excluded from the bead interface (Fig. 4D, upper panel). In MV-exposed cultures, T cells remained mainly round and failed to detectably polarize which also referred to the localization of the intracellular NSM-GFP enriched compartments (Fig. 4D, bottom panels and S2 Fig.). As seen for ceramides, NSM2 was not efficiently excluded from the interfaces involving MV-pre-exposed T cells indicating that MV signaling interferes with compartmentalization of both NSM2 and ceramides in the IS.
To evaluate whether defective spreading responses were associated with impairments of T cell activation, accumulation of tyrosine phosphorylated protein species (p-tyr) indicating relay of TCR signaling in response to α-CD3/CD28 stimulation were analyzed in NSMKD T cells pre-exposed to MV or MOCK. In CTRL T cell cultures, MV exposure appeared to compromize accumulation of certain, yet not all p-tyr protein species as compared to MOCK-treated cells (Fig. 6A, left, asterisks). NSM knockdown substantially increased accumulation of p-tyr protein species in MV exposed cells (NSMKD+MV), which, however, still remained below the levels seen in MOCK-treated NSMKD T cells which by far exceeded those seen in CTRL T cell cultures (Fig. 6A, right panel). These data support the overall importance of NSM activity in downmodulating initiation of T cell activation (Fig. 3), but also reveal that MV mediated NSM activation contributes to alterations seen with regard to p-tyr accumulation after TCR triggering. To study the impact of NSM activation on MV-induced inhibition of stimulated T cell expansion, NSMKD and CTRL T cells were exposed to MV (or the corresponding amounts of MOCK) and proliferation was analyzed 48 h following αãCD3/CD28 stimulation. MV dose dependently inhibited T cell expansion both in CTRL and NSMKD T cells, however, the latter expanded substantially more effective indicating that NSM ablation enables high proliferative activity in spite of MV inhibitory signaling (Fig. 6B). Therefore, NSM activation by MV contributes, however, not fully accounts for MV T cell inhibition. Altogether, these findings suggest that NSM activation during TCR triggering acts to dampen T cell activation, which, when inappropriately elicited, results in T cell suppression.
Using siRNA mediated genetic NSM ablation, our present study reveals the hitherto unknown activity of the enzyme to dampen the activation of co-stimulated T cells. This is because T cell adhesion, spreading, actin cytoskeletal dynamics, accumulation of tyrosine phosphorylated protein species and expansion are significantly enhanced upon NSM knockdown (Fig. 3, 5–7). At the subcellular level, accumulation of ceramide enriched membrane domains is compartmentalized within the lamellum, whereas ceramides appear to be excluded from the IS center and the lamellipodium (Fig. 2). Indicating that MV exploits and augments the dampening activity of NSM to interfere with T cell activation, MV exposure causes substantial accumulation of ceramide within the IS. This and MV-induced loss of adhesion and spreading responses, is entirely, while MV-interference with p-tyr and T cell expansion are partially rescued upon NSM ablation (Fig. 7). Unfortunately, it is impossible to evaluate the role of the enzyme in MV immunosuppression in vivo. Firstly, mice are not permissive for peripheral MV infection and therefore cannot be used to study MV T cell paralysis, and secondly, Smpd3 deficient animals (also referred to as fragilis ossium (fro/fro) mice) suffer from severe chondrodysplasia and dwarfism which, together with their low birth frequencies (90% of the embryos are lost prenatally) precludes infection experiments [30]–[32].
Ceramide enriched microdomains usually generated in response to stimulated activation of sphingomyelinases and subsequent sphingomyelin breakdown are sites of lateral protein segregation also including receptors and associated signalosomes [3], [33], [34]. Both phenomena efficiently regulate cellular responses to exogenous triggers in a variety of cells also including T lymphocytes where relay of extracellular cues translates into regulation of motility, adhesion and activation all of which are associated with massive reorganizations of the cytoskeleton and signalosome complexes [19], [35]. The role of sphingomyelinases and ceramide release early in TCR activation remained, however, controversal.
While confirming earlier finding that CD3 ligation alone causes NSM activation which peaked late after activation [20], we found that co-stimulation resulted in an early rise of the NSM activity which subsequently returned to background levels. Thus, NSM activity upon CD3/CD28 co-ligation is timely restricted and not prolonged within the first 30 min of stimulation (Fig. 1). As reported earlier [21], ligation of CD28 alone induced ASM and surface ceramide display. This was, however entirely ablated upon co-ligation (Fig. 1), indicating that ASM activity is dispensable for early T cell activation. In line with this, ASM activation or exogenous ceramide supply were not required for or even inhibitory to early TCR signaling and co-stimulation [22], [23]. Indicating that ASM activity interferes with T cell activation, abrogation of TCR-stimulated Ca2+ mobilization following CD95 or TNFR-ligation was attributed to ASM activation [18], [36], [37]. In contrast, ASM activation has also been reported to be beneficial for CD28-mediated NF-κB activation or discharge of cytokines from T cells [22], [38]. In co-stimulated cells, early ASM activation did not occur (Fig. 1), and we did not detect any significant impact on NSM ablation on accumulation levels of the cytokines measured intracellularly or in supernatants (not shown). Though NSMKD clearly enhanced kinetics of early T cell activation, activation of CTRL cells to comparative levels occurred (though with delay, Fig. 3,6) and therefore, NSM relating alterations of late activation functions such as cytokine production might not be highly pronounced. Moreover, to the best of our knowledge, except for its role in exosome production [39]–[41], a role of NSM in vesicular trafficking and or discharge of vesicular compartments has so far not been revealed.
NSM/ASM cross-regulation has been reported earlier. Asm was substantially elevated at RNA, protein and activity level in fibroblasts of Smpd3-deficient fro/fro mice in the absence of stimulation [26], which was, however, not the case for NSMKD T cells (S1B Fig.). In MV exposed T cells, NSM activity was required for ASM activity triggered by the virus via an unknown receptor [18]. It is beyond the scope of this study to unravel mechanisms underlying NSM/ASM crossregulation in our system, which, however, describes the first example of crossregulation by a receptor complex.
By revealing that NSM ablation clearly had an enhancing effect on steady state actin dynamics, early TCR signaling and expansion, we established the role of the enzyme in regulating kinetics and magnitude of T cell activation (Fig. 3, 6, 7). In contrast to enhanced proliferative responses seen in NSMKD or Smpd3-deficient splenocytes, re-introduction of Smpd3 corrected for a strongly impaired cell cycling of Smpd3-deficient fro/fro fibroblasts [26] indicating that either compartmentalized receptor-mediated activation or cell type specificity defines the impact of NSM on proliferation. In line with our observations made with regard to an inhibitory role of the enzyme in T cells, choleratoxin B mediated inhibition of human CD4 cell proliferation has been related to activation of NSM in lipid rafts [42]. Notably, however, reduced ceramide levels in fro/fro fibroblasts were associated with elevated PI3K activity and p-Akt levels indicating a role of NSM in general prevention of activated cellular steady state [43]. In line with this, NSM ablation in T cells elevated basal levels of p-tyr prior to αãCD3/CD28 stimulation (Fig. 6).
Molecular targets of NSM in physiological T cell activation remain unclear. As obvious from the p-tyr analysis, ablation of the enzyme enhanced both kinetics and magnitude of signaling responses which arguably might reflect its role in regulating compartmentalization of signaling microclusters rather than specific downstream targets (Fig. 6). Confinement of ceramides and the the majority of the NSM2 to the lamellum where actomyosin dynamics regulates microcluster transport might indicate a role in communication with cytoskeletal adaptors [44], [45]. In turn, ceramide exclusion from the lamellipodium would be compatible with the requirement of sphingomyelin rather than ceramide at areas of high actin turnover [18], [46], [47]. When overexpressed in MCF-7 cells, a NSM-GFP pool was found to shuttle from the Golgi to the plasma membrane [28]. Similarly, a fraction of NSM appeared to be redistributed from a perinuclear compartment upon T cell stimulation, and, most interestingly, late after stimulation, to be segregated from the IS towards the distal pole altogether indicating that trafficking of this enzyme during T cell activation is tightly controlled (Fig. 4D). Mechanisms underlying NSM trafficking are, however, unknown as yet.
Both ceramides and NSM are excluded from the IS center upon physiological TCR activation, and strikingly, ceramides and NSM2 were shifted to the central IS upon pre-exposure to MV prior to stimulation (Fig. 4C, D). Though this was entirely corrected for by NSM ablation, we cannot rule out that ceramide release there reflected aberrant activation of ASM as found to be activated in MV exposed cells- [18].
Rescue of MV-induced ceramide mis-localization, loss of adhesion and spreading responses by NSM ablation identifies the enzyme as of key importance in pathogen induced suppression of early T cell activation (Fig. 4,5). It certainly also contributes to, but not fully accounts for MV interference with p-tyr accumulation and T cell expansion (Fig. 6). Possibly, identification of specific p-tyr targets of MV (which are rescued upon NSM knockdown, examples marked in Fig. 6A) would be informative in delineating proteins activation of which is required in conferring partial resistence of T cells to MV inhibition.
Primary human cells were obtained from the Department of Transfusion Medicine, University of Würzburg, and analysed anonymously. All experiments involving human material were conducted according to the principles expressed in the Declaration of Helsinki and ethically approved by the Ethical Committee of the Medical Faculty of the University of Wuerzburg.
Primary human PBMCs were isolated from peripheral blood obtained from healthy donors by Ficoll gradient centrifugation. CD3+ T cells (purity ≥90%) were enriched from the PBMC fraction using nylon wool columns and maintained in RPMI 1640/10% FCS. The MV wild-type strain WTF was grown on human lymphoblastoid BJAB cells kept in RPMI 1640/10% FCS and titrated on marmoset lymphoblastoid B95a cells. For exposure experiments, MV (or MOCK preparation obtained from uninfected BJAB cells) was purified by sucrose gradient ultracentrifugation. T cells were exposed to MV or a MOCK preparation in the presence of a fusion inhibitory peptide (Z-D-Phe-L-Phe-Gly-OH; 200 mM in DMSO; Bachem) to prevent infection of T cells.
1×105 T cells when indicated pre-exposed to GW4869 (2 h, 1,3 µM), recombinant bacterial sphingomyelinase (30 mins, 12,5 mU/ml)(both: Sigma-Aldrich), MV or MOCK (each: 2 h on ice) were pre-incubated with CD3- (clone UCHT-1) and/or CD28-specific antibodies (clone CD28.2) (each 1 µg/ml)(both Beckton-Dickinson Biosciences Pharmingen) on ice, subsequently transferred onto 8-chamber slides for immunostaining (LabTekII, Nunc) or 96 well plates (for proliferation assays) precoated with 25 µg/ml α-mouse IgG (Dianova) (1 h at 37°C) and stimulated for the time intervals indicated at 37°C. For pseudo-IS formation, 2×105 T cells were stimulated for 30 min at 37°C in 100 µl RPMI 1640/0,5% BSA with α-CD3/CD28-coated beads (Dynabeads Human T-Activator CD3/CD28; Invitrogen) at a ratio of 4∶1, captured onto a poly-L-lysine-coated slides (LabTekII, Nunc) and fixed at RT for 15 min in 4% PFA/PBS. For proliferation assays, NSMKD or CTRL T cells exposed to MV or MOCK were stimulated for 48 or 72 h including last 24 h labeling period ([3H]-thymidine (Amersham)) and analyzed using a microplate scintillation counter. For mixed leukocyte reaction with murine cells, Smpd3-deficient splenocytes were isolated from Smpd3 knockout fro/fro mice (or Smpd3-sufficient littermates) and co-cultured with syngenic, superantigen-loaded bone marrow derived dendritic cells (ratio 10∶1) for 5 days.
Nucleofection of human T cells was performed according to the manufacturer's protocol (Amaxa) using pNSM2-GFP [48] (kindly provided by Y. Hannun). 24 h following nucleofection, transfection efficiencies were determined by flow cytometry (the percentage of GFP+ cells ranged between 30 and 45%). For silencing of NSM2, human T cells were nucleofected twice with a two days interval with 400 pmol siRNA targeting human SMPD3 (NSM2) [49] or, for control, a non-targeting siRNA (Sigma-Aldrich). Cell aliquots were harvested at day 5 for activity assays. On average, knockdown efficiencies were higher than 80% at enzyme activity level (as exemplified in Fig. 1C). When indicated, cells were exposed to dexamethasone (10−5 M) (Sigma-Aldrich) or ETOH (used as solvent) for 1 h. ASM or NSM activities were determined as previously described [20] with modifications. 3×106 T cells were disrupted by freeze/thawing (methanol/dry ice) in ASM or NSM lysis buffer (pH 5.2 for ASM and pH 7.4 for NSM). Nuclei were removed by centrifugation for 5 min at 1600 rpm. Post-nuclear homogenate was centrifuged for 1 h at 26 000 rpm in PBS with protease inhibitors for detection of NSM activity and in 100 mM Na-acetate, pH 5,2 for ASM. Cell membrane extracts in ASM or NSM specific lysis buffer were incubated with 1,35 mM HMU-PC (6-hexadecanoylamino-4-methylumbelliferyl-phosphorylcholine) (Moscerdam substrates) as an artificial sphingomyelinase substrate at 37°C for 17 h (final volume 30 µl). Fluorescence reading was performed using excitation at 404 nm and emission at 460 nm according to the manufacturers protocol.
For surface detection of ceramides, T cells were fixed with 1% PFA/PBS for 15 min on ice after the indicated time intervals, stained with primary (a-ceramide; clone MID15B4; Alexis Biochemicals; 1 h, 4°C) followed by a secondary (30 min, 4°C) antibody, and analyzed by flow cytometry (FACS Calibur; Becton Dickinson).
Cytokines (IL-2, IL-4, IL-5, IL-10, IFN-γ and TNF-α) were detected in supernatants 4, 10, 24 and 72 h following aãCD3/CD28 stimulation by cytometric bead assay (CBA) according to the manufacturers instructions (human soluble protein flex sets; BD Biosciences) using DIVA and FCAP Array Software. For intracellular cytokine detection, cells were restimulated 72 h following activation with 40 µM PMA/0,5 µM ionomycin for 4 h in the presence of 35 µM Brefeldin A (all: Sigma). Following a mouse Ig bocking step, cells were stained with α-CD4 (Biolegend), fixed and permeabilized prior to staining with antibodies specific for IL-2 (Pharmingen), IL-10 (eBioscience) IFN-γ and IL-17α (both: BioLegend).
For Ca2+-mobilization experiments, T cells (1×106) were washed once and loaded with 1 µM Fluo-4 as cell-permanent AM ester (Molecular Probes) in Hanks balanced salt solution (without CaCl2, MgSO4, and phenol red) containing 5% FCS and 25 mM HEPES (pH 7.5) according to manufacturers' protocol. α-CD3 and α-CD28 antibodies (5 µg/ml) were added in complete Hanks medium and Ca2+ flux was determined by flow cytometry.
T cell activation was stopped by adding 4% PFA (in PBS) for 15 min, permeabilized with 0.1% Triton-X100 for 5 min, blocked with 5% BSA and incubated with primary antibodies (α-ceramide, α-LAT; FL-233; Santa Cruz) diluted in 1% BSA overnight at 4°C. Cells were stained with appropriate Alexa488-conjugated secondary antibody (Invitrogen) for 45 min at RT. F-actin was detected with 488 or 555 Fluorochrom-conjugated Phalloidin (Cytoskeleton). Samples were mounted with Fluorochrome G (Southern Biotech). Confocal Laser Scanning Microscopy (CLSM) imaging was performed using a LSM 510 Meta (Zeiss, Germany), equipped with an inverted Axiovert 200 microscope and a 40x or 63x EC Plan-Apo oil objective (numerical aperture 1.3 or 1.4, respectively) and laser lines 488 and 543. Image acquisition was performed with Zeiss LSM software 3.2 SP2. When indicated, 0.15 µm thick z-stacks were acquired and 3–dimensional reconstructed using LSM software or processed by Huygens deconvolution software (SVI, Hilversum, The Netherlands). Quantification of cell area measurement was calculated using the imaging processing program ImageJ (http://rsb.info.nih.gov/ij/). For live analysis, 24 h following nucleofection of pNSM2-GFP T cells were exposed to MV or MOCK, stimulated with α-CD3/CD28-coated beads, subsequently transferred to 6-channel µ-slide VI (ibidi) and immediately imaged by confocal microscopy.
For SEM, 3×105 T cells were stimulated as described onto co-stimulatory coverslips (12 mm) in a 24-well plate for the time intervals indicated, fixed by addition of pre-warmed 6.25% glutaraldehyde in 50 mM phosphate buffer (pH 7.2) for 10 min at RT and subsequently at 4°C overnight. After a washing step, samples were dehydrated stepwise in acetone, critical point dried and sputtered with platin/paladium before SEM analysis (Jeol JSM 7500 E).
3×106 CTRL or NSMKD T cells were stimulated for the time intervals indicated by cell-bound α-CD3/CD28 antibodies (1 µg/ml each; Becton Dickinson) and plate-bound α-mouse IgG (25 µg/ml; Dianova) in 12 well cell culture plate, lysed in 200 µl Western blot sample buffer after freezing of samples by −80°C followed by boiling for 5 min. p-tyr were detected using Mouse monoclonal antibodies were used to detect tyrosine phosphorylated protein species (p-tyr) (clone 4G10, Millipore), GAPDH, Akt (Santa Cruz), p-Akt and pERK (both: Cell signaling). Quantification of signal intensities was performed using AIDA software (Raytest).
Overall, data shown were acquired in at least three independent experiments involving individual donors. For statistical analyses of data sets, two-tailed Student's t test (*p<0,05, **p <0,005 and ***p<0,0005; ns: non significant) was used for throughout the manuscript. Bars show standard deviations.
|
10.1371/journal.pgen.1004736 | Rad59-Facilitated Acquisition of Y′ Elements by Short Telomeres Delays the Onset of Senescence | Telomerase-negative yeasts survive via one of the two Rad52-dependent recombination pathways, which have distinct genetic requirements. Although the telomere pattern of type I and type II survivors is well characterized, the mechanistic details of short telomere rearrangement into highly evolved pattern observed in survivors are still missing. Here, we analyze immediate events taking place at the abruptly shortened VII-L and native telomeres. We show that short telomeres engage in pairing with internal Rap1-bound TG1–3-like tracts present between subtelomeric X and Y′ elements, which is followed by BIR-mediated non-reciprocal translocation of Y′ element and terminal TG1–3 repeats from the donor end onto the shortened telomere. We found that choice of the Y′ donor was not random, since both engineered telomere VII-L and native VI-R acquired Y′ elements from partially overlapping sets of specific chromosome ends. Although short telomere repair was associated with transient delay in cell divisions, Y′ translocation on native telomeres did not require Mec1-dependent checkpoint. Furthermore, the homeologous pairing between the terminal TG1–3 repeats at VII-L and internal repeats on other chromosome ends was largely independent of Rad51, but instead it was facilitated by Rad59 that stimulates Rad52 strand annealing activity. Therefore, Y′ translocation events taking place during presenescence are genetically separable from Rad51-dependent Y′ amplification process that occurs later during type I survivor formation. We show that Rad59-facilitated Y′ translocations on X-only telomeres delay the onset of senescence while preparing ground for type I survivor formation.
| In humans, telomerase is expressed in the germline and stem, but is repressed in somatic cells, which limits replicative lifespan of the latter. To unleash cell proliferation, telomerase is reactivated in most human cancers, but some cancer cells employ alternative lengthening of telomeres (ALT) based on homologous recombination (HR) to escape senescence. Recombination-based telomere maintenance similar to ALT was originally discovered in budding yeast deficient in telomerase activity. Two types of telomere arrangement that depend on two genetically distinct HR pathways (RAD51- and RAD59-dependent) were identified in post-senescent survivors, but the transition to telomere maintenance by HR is poorly understood. Here, we show that one of the earliest steps of short telomere rearrangement in telomerase-negative yeast is directly related to the “short telomere rescue pathway” proposed 20 years ago by Lundblad and Blackburn, which culminates in the acquisition of subtelomeric Y′ element by shortened telomere. We found that this telomere rearrangement depends on Rad52 strand annealing activity stimulated by Rad59, thus it is distinct from Rad51-dependent Y′ amplification process observed in type I survivors. We show that continuous repair of critically short telomeres in telomerase-negative cells delays the onset of senescence and prepares the ground for telomere maintenance by HR.
| Telomeres are nucleoprotein structures found at the physical ends of chromosomes. Their terminal location defines their two main functions: protection of the chromosome ends from illegitimate repair reactions and prevention of the loss of terminal DNA due to either degradation or incomplete replication [1]. In Saccharomyces cerevisiae, the first function is accomplished primarily by Rap1, which wraps tandem telomeric DNA repeats to inhibit NHEJ [2], and recruits Rif1 and Rif2 to restrain MRX-mediated 3′ end resection [3], [4], thus limiting recruitment of HR factors and checkpoint signaling [5]. The second function is mediated by Cdc13 bound to the single-stranded G-rich 3′ overhang at the extreme terminus of a telomere. Cdc13 forms alternative complexes with either Est1 or Stn1-Ten1 to coordinate telomerase-mediated synthesis of the G-rich strand with the synthesis of the complementary strand by DNA polymerase α [6].
As in mammals, telomeres in yeast cells with reduced telomerase activity progressively shorten with each cell division until they are recognized as DNA damage and recruit Mec1 kinase that initiates irreversible G2/M arrest [7]–[9]. At the level of cell population, telomere dysfunction is manifested as crisis, when majority of the cells irreversibly arrest in G2/M [7]. Most of the cells die, but at low frequency survivors emerge, which maintain their telomeres via recombination [10], [11], implying that homologous recombination (HR) can serve as a bypass pathway to sustain viability in the absence of telomerase.
The survivors are classified in two types based on their telomere arrangement and growth characteristics [12], [13]. The type I survivors have tandem arrays of subtelomeric Y′ elements separated by short tracts of TG1–3 repeats at most chromosome ends, and also short terminal TG1–3 repeats [10]. Their growth is interrupted by frequent periods of arrest and in the competitive conditions of liquid culture they are outcompeted by the more robust type II survivors. In type II survivors, terminal TG1–3 repeats are abnormally elongated and are very heterogeneous in length. It is believed that they are established by stochastic lengthening events that likely involve rolling circle replication [14]. RAD52 is required for generation of both types of survivors. RAD51, RAD54, RAD57 are specifically required to generate type I, whereas type II survivors depend on MRX complex, RAD59 and SGS1, encoding the only RecQ helicase in yeast [13], [15], [16]. In addition, POL32 encoding a non-essential subunit of DNA polymerase δ is required for generation of both survivor types, implying involvement of the processive repair DNA synthesis in the recombination-based telomere rearrangements [17], [18]. Recently, a genome-wide screen aimed to identify telomere-length-maintenance genes that regulate telomere structure in post-senescence survivors unveil new regulators of Type I and II recombination [19]. Notably, Type I recombination was shown to depend on the helicase Pif1 and on the chromatin remodelling complex INO80.
Although genetic requirements for the formation of two types of survivors and their telomere patterns have been well characterized, much less is known about actual recombination events that lead to reorganization of the original short telomere into the patterns observed in survivors. In budding yeast, telomere shortening does not cause end-to-end chromosome fusions, as does the removal of Rap1 from telomeres [2]; instead, gene conversion increases near short telomeres indicating de-repressed recombination [20]. There is a controversy, however, whether type II recombination preferentially takes place at long or short telomeres [14], [21], [22]. Little is known about the telomere length preference of type I pathway. Early studies looking at the propagation of linear plasmids in yeast uncovered that they can recombine with the yeast chromosome ends and acquire telomere-adjacent sequences called Y′ elements [23]. Y′ elements found at many chromosome ends fall into two size classes, 6.7 (Y′-L) and 5.2 (Y′-S) kb-long, that differ by a 1.5 kb insertion/deletion [24]. Another subtelomeric sequences called X elements are present at all chromosome ends immediately proximal to either Y's or terminal TG1-3 repeats when Y′ is absent. The junction between X and Y′ elements often, but not always, contain short tracts of TG1–3 repeats [24], [25]. Importantly, only the Y′ and not X elements can be transferred on linear plasmids, and this is mediated by recombination between the terminal TG1–3 repeats added onto the plasmid ends by telomerase and the internal TG1–3 tracts present between the X and Y′ elements [23]. Pioneering work of Lundblad and Blackburn showed that est1Δ survivors arose as the result of the acquisition of Y′ elements by X-only telomeres and amplification of these elements on many chromosome ends. They proposed a model of telomere rescue via a recombination event between the terminal TG1–3 repeats of one telomere and an internal TG1–3 tract in another [10].
We have previously demonstrated using single cell analysis that Rad52-containing foci are assembled at the telomeres in a length-dependent manner in presenescent cells many generations before the onset of senescence [26]. The recruitment of recombination factors to short telomeres is in accord with increased recombinogenic activity of short telomeres observed in both yeast [20] and mammals [27]. Of note, inactivation of HR, particularly via deletion of RAD52 and RAD51 (but to a much lesser extent of RAD59), causes early decline in proliferative capacity of telomerase-negative yeast indicating that telomere maintenance most likely becomes dependent on HR soon after telomerase inactivation [28], [29]. Surprisingly, the rate of telomere shortening (population average length) is unaffected in HR-deficient yeast. All these observations raise the question of telomere recombination dynamics in presenescent cells, the mechanism of Y′ acquisition by X-only telomeres and the role of recombination proteins in maintaining telomerase-negative strains alive during presenescence. Another unresolved issue is whether a single critically short telomere is sufficient to induce cell cycle arrest. Complete loss of a single telomere causes Rad9-dependent arrest even in telomerase-proficient cells [30]. This does not seem to be the case when a very short telomere is created in telomerase negative cells [8], [28], but the fate of this abruptly shortened telomere remains obscure. In this study, we aimed to characterize the primary recombination event that takes place at short telomeres in the absence of telomerase. To this end we put together a system to simultaneously shorten modified VII-L telomere and inactivate telomerase. Bulk liquid cultures turned out to be inappropriate to address the fate of the abruptly shortened VII-L telomere, so we adapted clonal analysis. We found that the subtelomereless VII-L end acquired Y′ element in clonal populations originated from transiently arrested cells. Cloning and sequencing of the Y′ translocation junctions from multiple clones revealed that Y′ acquisition was initiated by recombination between the short terminal TG1–3 repeats at VII-L and the Rap1-bound internal TG1–3-like tracts present between X and Y′ elements on other chromosome ends. Such recombination initiates Pol32-dependent BIR, which results in non-reciprocal translocation of the entire Y′ element and terminal TG1–3 tract from the chromosome-donor onto the shortened telomere. Surprisingly, Y′ translocation events were Rad51-independent, but were instead promoted by Rad59 that stimulates Rad52 strand annealing activity. We found that the same mechanism operates at short native X-only telomeres, but it is much more efficient since translocated Y's are readily detectable in bulk liquid cultures during presenescence. In addition, sequence composition of the translocation junctions is simpler at native telomere VI-R, indicating that Y′ translocation on a native end is relatively straightforward event. We further show that RAD59 deletion compromises the efficiency of Y′ translocation on native telomere XV-L, and results in both accelerated senescence and prolonged crisis. Our results extend the model of short telomere rescue proposed by Lundblad and Blackburn more than 20 years ago [10], and they reinforce the notion that it is the overall depletion of the TG1–3 repeats on multiple chromosome ends rather than abrupt shortening of a few telomeres that defines the onset of senescence.
To address the processing of a short telomere without Y′ elements in the absence of telomerase, we employed the site-specific recombination system to induce abrupt shortening of a single telomere [31]. In this system, Cre induction causes excision of the basal portion of the telomere VII-L (TelVII-L) flanked by loxP sites (Figure 1A). In the presence of telomerase, shortened telomere is extended until its length returns to equilibrium [32]. To examine how this telomere will be processed in the absence of telomerase, we combined the abrupt shortening of TelVII-L with an inducible deletion of the plasmid-borne EST2 [33].
As expected, inactivation of telomerase completely abolished elongation of the TelVII-L after it was shortened via Cre-loxP recombination. Instead, its length decreased further until the bottom of the telomere length distribution reached a defined limit beyond which no shortening was observed (Figure 1B). The lower tail of the TelVII-L length distribution in the control strain also reached the same limit albeit with a delay of ∼20 population doublings (PDs) consistent with its greater initial length. We estimated that the lower limit of the TelVII-L length distribution corresponds to ∼60 bp of TG1–3 repeats (Figure S1). Since the probe anneals to the unique sequence of the terminal PacI fragment (Figure 1A), even complete loss of TG1–3 repeats should not affect hybridization signal. Thus, we reasoned that shortening of the TG1–3 tract beyond 60 bp causes the elimination of cells with critically short telomeres from the exponentially growing culture propagated via serial dilutions.
To isolate the cells undergoing cell cycle arrest due to TG1–3 tract shortening beyond the 60 bp threshold, we conducted clonal analysis of the telomerase-negative cultures at ∼15 PD after Cre induction. To this end, single cells were micromanipulated on a grid of agar, and analyzed for their ability to form microcolonies. While many cells divided regularly, at least once every 2 hours, and formed microcolonies of more than 16 cells after 8 hours on agar, a fraction of cells never divided during this time or stopped dividing at the 2- or 4-cell stage. These arrested microcolonies were marked (Figure 2A). Unexpectedly, most of the cells, which initially failed to divide, formed colonies after four days at 30°C (Figure 2A). Therefore, the majority of cells was able to overcome cell cycle arrest and resumed divisions. The fraction of arrested cells was significantly greater in the strain with shortened TelVII-L compared to the control strain (Figure 2B and Figure S2) indicating that shortening of a single telomere aggravated the effect of telomerase inactivation on cell cycle progression.
The state of the TelVII-L in clonal populations was analyzed by Southern blotting (Figure 2C). The terminal VII-L fragments with typical smeary appearance were detected in the expected size range for the control (no arrest) clones. In contrast, most of the clones that had undergone transient arrest completely lost the VII-L signal in this range. Instead, much larger fragments hybridized with VII-L-specific probe (Figure 2C), suggesting that VII-L end has been rearranged. These larger fragments grouped in two size classes after digestion with either PacI or MfeI, and were remarkably consistent with Y′ element translocation (see schematic in Figure 2C). Grouping of the fragments in two size classes could be well explained by translocation of either long (Y′-L) or short (Y′-S) version of the Y′ elements, which differ by a 1.5 kb insertion/deletion [24]. Most of the clones showed the presence of both Y′ classes. This heterogeneity is likely created due to independent acquisition of Y′ elements by TelVII-L in different cells (2- or 4-cell stage arrest), or even by two sister TelVII-L chromatids within a G2-arrested cell. Thus, we were able to isolate subclones descended from single recombination events (group A in Figure S2) by sequential micromanipulation of the cells as they came out of the arrest. Those transiently arrested clones that did not show VII-L end rearrangement (Figure 2C and Figure S2) could have either repaired it by addition of the terminal TG1–3 repeats [22] or arrested due to shortening of one of the native telomeres.
To verify the hypothesis of Y′ translocation and the kinetics of this repair process, we designed primer pairs to amplify the putative junction region between the VII-L end and the Y′ element (Figure 3A, Table S1). We analyzed the presence of the junction PCR product at 0, 10, and 50 PDs after inactivation of telomerase in the strain bearing a critically short telomere (16 Rap1-bs) and in the control strain (0 Rap1-bs) (Refer to Figure 1B). We detected the junction-specific PCR product at VII-L (Figure 1A) as early as 10 PD in liquid culture after inactivation of telomerase and telomere shortening. The intensity and heterogeneity of the junction PCR product increased dramatically by 50 PD (Figure 3B). Moreover, the junction PCR product was more intense at 10 PD in the strain with short telomere compared to that in the control strain. When the same PCR approach to detect Y′ translocation was performed on the native VI-R end, the appearance of recombined telomere followed the same kinetics in the “0” and “16 Rap1-bs” strains (see further).
To assess the relationship between telomere length and Y′ acquisition in a more quantitative fashion, we analyzed the frequency of Y′ translocation on TelVII-L by Southern blotting in the random clones isolated from “0” and “16 Rap1-bs” strains. This analysis revealed significantly greater fraction of clones with Y′ element translocated on TelVII-L in the “16 Rap1-bs” (10/18) as compared to that in “0 Rap1-bs” (1/19) strain when the clones were isolated at 18 PD after Cre induction (Fisher's exact test P value 0.001; Figure S2). The same tendency was observed when the clones were isolated at 12 PD after Cre induction, but the frequency of Y′ translocation at this earlier time was too low for statistical evaluation (0/24 in “0 Rap1-bs” and 3/24 in “16 Rap1-bs” strain). In addition, the fraction of “16 Rap1-bs” clones with Y′ translocated on TelVII-L was significantly greater when the clones were isolated at 18 as compared to 12 PD after Cre induction (Fisher's exact test P value 0.006; Figure S2). The frequency of Y′ translocation among “0 Rap1-bs” clones was still low at these time points for statistical evaluation. We concluded that Y′ element is preferentially translocated on short telomeres; and the fraction of cells with translocated Y′ element grows with time after telomerase inactivation.
Cloning and sequencing of the junction PCR products revealed that they are in fact composed of the joint VII-L and Y′ element sequences, thereby confirming Y′ translocation on the VII-L end. Remarkably, all clones contained the TG1–3 repeats at the breakpoint between the VII-L and Y′ sequences, consistent with our assumption that Y′ translocation was initiated by VII-L terminal repeats, which recombined with internal TG1–3 tracts located between the subtelomeric X and Y′ elements of the chromosome donor. Sequence analysis of the Y′ segments of the junction clones revealed that the choice of the donor for Y′ translocation was not random: the chromosome arms VI-L and VII-R were used most frequently as the donors (Table 1).
To characterize recombination breakpoint we searched for the point of sequence divergence within TG1–3 repeats at the VII-L and Y′ junction. The exact pattern of irregular repeats newly synthesized by yeast telomerase differs between the molecules [34]. When telomere repeat sequences from clonal populations are aligned, only very distal 40–100 nt-long portions maintained by telomerase are divergent, whereas proximal regions are identical in sequence [35]. The divergent distal portion is quickly lost after telomerase inactivation. We found that TG1–3 repeats adjacent to the VII-L end shared 47±18 (SD) nt of identity before they diverged (Figures 4 and S3A), indicating that VII-L telomeres were at least that long at the time they engaged in recombination. Only in a few clones, the TG1–3 sequence past the divergence point continued without interruption by repeats found proximal to the acquired Y′ element (Figure S3B). In most of the clones, however, the repeats of the VII-L and Y′ origins were separated by divergent TG1–3 repeats (Figure 4). These intervening TG1–3 tracts varied in length and sometimes were quite long (e.g. 255 bp in the clone H10). We speculate that they might have resulted from abortive DNA repair synthesis due to rejection of homeologous heteroduplexes formed by TG1–3 repeats followed by template switch events [36]. Otherwise, the divergent regions at the junctions could be generated via recombination between two terminal TG1–3 repeat tracts, or even by residual telomerase activity during the first few PDs after Cre induction.
To get insight into the mechanism of short telomere repair, we examined its genetic requirements. To this end, we generated a set of isogenic mutants by deleting RAD52, RAD51, and RAD59 genes in a strain with the VII-L telomere modified for inducible shortening and TLC1 allele with a tetracycline-regulatable promoter (tetO2-TLC1). Upon addition of doxycycline (Dox), expression of TLC1 is tightly repressed and telomeres shorten progressively with each generation ([26] and Figure S4). We first examined the effect of gene deletions on the efficiency of Y′ translocation by semi-quantitative PCR assay. Telomerase was inactivated by addition of Dox and abrupt shortening of the VII-L was then induced following the scheme in Figure 5A. We analyzed junction PCR products in clonal populations of HR-proficient or mutant cells (Figure 5B,C and Figure S5). As expected, robust junction PCR products indicating Y′ translocation events were detected in the clonal populations of HR-proficient cells (Figure 5B). Weaker product was also detectable in the mixed population of cells grown in liquid culture (Bulk). RAD52 deletion nearly completely abolished Y′ translocation as judged by severe reduction of the amplified VII-L/Y′ junctions in either mixed or clonal populations (Figure 5B). Thus, Rad52 is essential for recombination between TG1–3 repeats that leads to Y′ translocation. Surprisingly, deletion of the RAD51 had little effect on the efficiency of Y′ translocation (Figure 5C, top panel), demonstrating that Rad51 may not be essential for recombination between TG1–3 repeats. In contrast, RAD59 deletion reduced both the amount and the heterogeneity of amplified VII-L/Y′ junctions (Figure 5C, midpanel) suggesting lower frequency of TG1–3 repeat recombination in the absence of Rad59. These results indicate that Rad51 filament assembly may not be required at the 3′ overhang of the short telomere, which pairing with an internal tract of the TG1–3 repeats could depend only on Rad52 strand annealing activity which is stimulated by Rad59.
While Y′ translocation is initiated by TG1–3 repeat recombination, its completion likely depends on break induced replication (BIR), a pathway used to repair a DSB when homology is limited to its one side [37]. Deletion of POL32, encoding a nonessential subunit of Polδ required for processive DNA synthesis during BIR [38], severely reduced the efficiency of Y′ translocation as evidenced by drastically reduced junction PCR product (Figure 5C, bottom panel). Notably, the PCR at the VII-L terminus, which served as a loading control, failed for two transiently arrested pol32Δ clones (clones c and d) implying disappearance of the primer(s) site from the VII-L terminus. Therefore, we verified the integrity of the VII-L terminus in pol32Δ clones by Southern blot and found that it was indeed rearranged in both clones in a similar manner, which is different from Y′ translocation (Figure S6). We failed to detect Y′ translocation in any of the pol32Δ clones that we have analyzed by Southern blot. The junction PCR product that is reduced but still detectable in pol32Δ clones could have been generated on single strand extension products that were terminated before reaching chromosome end. Thus, we concluded that Pol32-dependent BIR is essential for the completion of Y′ translocation.
To quantify the Y′ translocation efficiencies in the aforementioned deletion mutants, we employed real-time qPCR using two pairs of primers to amplify either the total amount of VII-L or the VII-L/Y′ junctions (Figure 5D). For this analysis we used DNA extracted from random clones isolated at ∼16 PD after Cre induction which were also analyzed by Southern blotting with VII-L-specific probe (Figure S7). We performed clonal analysis because it provides an unbiased snapshot of the repair frequencies at any given time, whereas bulk liquid cultures of telomerase-negative cells are strongly affected by selection for best-growing clones. We estimated that at the time of analysis (∼40 PD after Cre) on average 43.5% of VII-L ends in wild-type clones acquired Y′ element. The fraction of “repaired” VII-L was reduced to 18.2 and 6.6% in rad51Δ and rad59Δ, respectively, whereas it was at the background level in both rad52Δ and pol32Δ clones (Figure 5D). We inferred from these results that Rad52 predominantly cooperates with Rad59 to initiate Pol32-dependent BIR which leads to Y′ translocation. Nevertheless, Rad51 also appears to contribute in this process since it reduces the efficiency of Y′ translocations by at least two fold.
To show that Y′ translocation is not a unique phenomenon that takes place at the modified VII-L end, but is common for eroded X-only telomeres, we used the same PCR approach to detect Y′ translocation on the native VI-R end (Figure S8A), which normally has only X element in the subtelomere region (W303 genome sequencing project, contig 00111). PCR product encompassing the junction between the VI-R end and the Y′ element was already detectable at 10 PD and increased further by 50 PD after telomerase inactivation (Figure S8B). Thus, erosion of the native telomeres in the absence of telomerase also leads to Y′ element translocations. Sequence analysis of the cloned PCR fragments confirmed that the VI-R/Y′ junction regions were predominantly amplified. However, we also identified a few clones resulted from mispriming on the X- and Y′-containing chromosomes V-R and XIII-L, which explains the background amplification before inactivation of telomerase.
As expected, all clones contained TG1–3 repeat tracts at the junction between the VI-R end and Y′ element. Thus, similar to modified VII-L, native chromosome VI-R end also acquired Y′ element as a result of recombination between the eroded terminal and internal TG1–3 repeats. The extent of TG1–3 sequence identity at the VI-R side of the junction was limited to 60±45 (mean ±SD) nt (Figure 6), the minimum length of the native telomeres at the time they recombined. The divergent repeat sequence was shorter on average (or even absent) compared to that at the junction of modified VII-L (Figure 6). Notably, among the chromosome ends that served as Y′ donors we identified a greater variety of ends (Table 1), indicating that native telomere VI-R was less selective with respect to a Y′ donor. Of note, not all junctions between the X and Y′ elements contain perfect TG1–3 repeats, but they all tend to be G-rich and are all bound by Rap1 (Table1 and Figure S9).
Since the native VI-R/Y′ junctions showed simpler composition indicative of possibly greater efficiency of Y′ translocation, we attempted to detect them during the senescence time course by Southern blot directly in the bulk liquid cultures. To this end, we performed standard senescence assay in liquid with est2Δ spore clones. The samples were collected for DNA extraction daily at the time of culture dilution. XhoI-digested DNA was then Southern blotted and probed for the native X-only XV-L end using subtelomeric probe. We were surprised to find that two high-molecular-weight bands characteristic of Y′L and Y'S translocations were readily detectable early on during the outgrowth of the est2Δ cultures (Figure 7A, B; top panels). The same result was obtained with the native VI-R end (not shown). As expected, the timing of the onset of Y′ translocations correlated with the initial XV-L telomere length, which differed among the spore clones. The abundance of cells with Y′ translocated on XV-L in est2Δ cultures during presenescence was consistent with the notion of greater efficiency of the Y′ element translocation on the native X-only telomere than on the modified subtelomereless VII-L end.
We next asked whether the RAD59 deletion, which reduced the frequency of Y′ translocation on the modified VII-L, would also compromise (or delay the onset of) Y′ translocations on the native XV-L as cells progress into senescence. Using the approach described above, we detected a substantial delay in the onset of Y′ translocations in the est2 rad59Δ relative to est2Δ clones (Figure 7). In the two representative est2Δ rad59Δ clones that are shown in Figure 7 (bottom panels), Y′ translocations were not detectable until the cells with very short XV-L telomere nearly disappeared from the population. In contrast, the cells with unrecombined short XV-L telomere and the cells which have already undergone Y′ translocation coexisted in the cultures of Rad59-proficient est2Δ clones long before they reached growth nadir. This result clearly points to the role of Rad59 in promoting Y′ translocation on the short native telomere. Note that in clone rad59Δ-1, type II pattern is seen after 8 days without telomerase. We usually observe that in liquid cultures, 75% of the clones are type I while 25% display a mixed pattern of type I and II (see Figure S11B).
The finding of short telomere repair by Y′ translocation raised a question whether this process can delay the onset of replicative senescence. Since Y′ element translocation is facilitated by Rad59 at X-only telomeres, we assayed the effect of RAD59 deletion on the onset of proliferative decline in the absence of telomerase. To this end we performed standard senescence assays using multiple est2Δ and est2Δ rad59Δ clones. Clones lacking Rad59 lost proliferative capacity slightly earlier and exhibited fivefold lower cell densities at the nadir of growth (Figure 8A). This observation indicates that Rad59-facilitated repair of critically short telomeres contributes to sustain cell proliferation particularly when a population of telomerase-negative cells approaches growth nadir. Consistent with this notion, we found by ChIP-qPCR that Rad59 associates with telomeres during presenescence. Although Rad59 association varies considerably among the telomeres, it may peak abruptly (up to 25-fold increase) at certain X-only telomeres when culture approaches growth nadir (Figure S10A). On average, Rad59 association with terminal and internal (between X and Y′ elements) TG1–3 repeats increases many fold as telomerase-negative cells progress into senescence (Figure S10B), highlighting the role of Rad59 in the conversion of X-only into Y′ telomeres.
We next reasoned that Y′ element translocation could be the initial step of type I telomere pattern formation at X-only telomeres. This premise seemingly contradicts established genetic requirements for survivor formation since type I survivors are predominantly obtained in cells lacking Rad59 [12], [13], [29]. The latter is due to the fact that Rad59 deletion greatly impedes type II. However, this does not exclude a possibility that Rad59 also facilitates transition to type I, particularly at X-only telomeres, although it is not strictly required. To reveal such a role of Rad59, we had to prevent the dominant type II pathway. This was performed by deleting SAE2 whose deletion strongly inhibits type II formation (Figure S11). We therefore compared the efficiency of type I survivor formation in est2Δ sae2Δ and est2Δ sae2Δ rad59Δ mutants (Figure 8B). We found that the triple mutant clones spent longer time in crisis suggesting reduced efficiency of type I survivor formation.
The involvement of RAD59 in Y′ translocation raised a question whether the other genes of the type II survivor pathway may also contribute to Y′ acquisition. To this end, we compared the kinetics of Y′ acquisition by the native telomere XV-L between est2Δ and sgs1Δ est2Δ cells. We chose to delete SGS1, another gene required for type II survivors to arise [13], [15], [16], [19], because unlike the genes encoding the subunits of MRX complex it does not cause short telomeres, which would complicate comparative analysis of Y′ translocation kinetics since short telomeres acquire Y′ faster. We found that although est2Δ and sgs1Δ est2Δ clones started to senesce with XV-L telomere of comparable size, there was a substantial delay in Y′ translocation on XV-L in sgs1Δ est2Δ compared to est2Δ cells (Figure 9). Therefore, Sgs1 also appear to promote Y′ element acquisition. As expected, XV-L telomere converted to type I pattern by the time sgs1Δ est2Δ cells generated survivors. Remarkably, recombination between Y′ elements increases in sgs1Δ mutants [39], thus the inhibitory effect of SGS1 deletion on Y′ translocation further highlights the mechanistic difference between largely RAD51-independent Y′ acquisition and RAD51-dependent Y′ amplification, the two steps of type I survivor formation.
Since we found that short telomere repair is associated with transient arrest, we asked whether the checkpoint function is required for Y′ translocation. To address it, we deleted MEC1, which is required for G2/M arrest in telomerase-deficient cells [40] and also mediates type II recombination [41]. As expected, mec1Δ sml1Δ mutant clones exhibited flatter senescence profiles indicative of defective cell cycle arrest in response to telomere shortening and delayed senescence as reported previously [8], and yet Y′ translocation was not affected in any of the three clones analyzed (Figure 9). We concluded that checkpoint function is not required for Y′ acquisition by X-only telomeres. This is in contrast to deleting RAD59 and SGS1, which both substantially delay Y′ acquisition.
Many studies from several laboratories have characterized the genetic pathways contributing to survivor formation [42]. In this study we demonstrated that a single telomere that was experimentally shortened in the absence of telomerase can acquire during presenescence Y′ element along with the terminal TG1–3 repeats from other chromosome ends which terminal repeats are still sufficiently long. Sequencing of the Y′ translocation junctions clearly evidenced that Y′ translocation was initiated by recombination between the short terminal TG1–3 repeats of the TelVII-L and the internal TG1–3 tracts located between the X and Y′ elements on certain native chromosome ends thereby confirming the model of Lundblad and Blackburn [10]. In addition, our results suggest that there is a continuous repair of the shortest telomeres that delays erosion into the subtelomere and allow the cell to escape DNA damage checkpoint activation. We showed that Y′ translocation was entirely RAD52-dependent, and was further promoted by RAD59. In contrast, RAD51 deletion had rather modest effect on the efficiency of Y′ translocation (two fold reduction), and Y′ translocations on VII-L end were detectable in rad51Δ clones. Therefore, Rad51-ssDNA filament formation and strand invasion do not seem to be obligatory for TG1–3 repeat recombination; instead, it could be accomplished by Rad52 strand annealing activity that is stimulated by Rad59. This mechanism bears strong similarities with the “short tract homology” recombination that has been described in the other context [43]. It remains possible that Rad51 affects Y′ translocations indirectly via its involvement in the recombination among the Y′ elements themselves [44]. Another question that rises with respect to Rad51-independent recombination between TG1–3 repeats is how the G-strand overhang anneals to the internal TG1–3 repeats, which are normally double-stranded. Possibly, this single-strand annealing occurs during subtelomere replication when the single strand regions are exposed [45], particularly when the replication forks are posed at the internal TG1–3 sequences [46].
In this study, we showed that such recombination events also occurred at the native telomeres VI-R and XV-L indicating that these rearrangements are physiological in nature. We found that modified subtelomereless VII-L and native X-only telomeres behave surprisingly different with respect to the efficiency of Y′ acquisition. Low efficiency of Y′ translocation on the modified VII-L end might be explained by its poor clustering with the other telomeres due to absence of subtelomeric elements. Indeed, intranuclear localization of the truncated TelVII-L is insensitive to the absence of either Sir3 or Yku [47], whereas these proteins participate in targeting native telomeres to the nuclear periphery [48], [49]. Poor clustering and failure to localize to nuclear compartments which favour recombination could be responsible for low efficiency of Y′ translocation and may explain why single experimentally shortened telomere accelerates senescence [8], [28].
Our results suggest that the Rad59-facilitated recombination between the terminal and internal TG1–3 repeats could be responsible for initial spreading of certain Y′ elements on X-only telomeres. We demonstrated that disappearance of the bands corresponding to X-only telomeres from Southern blots as telomerase-negative cultures progress into crisis is indeed a consequence of Y′ element acquisition that is promoted by Rad59. One can, therefore, envisage that the clones that arise via Y′ translocations are the precursors of type I survivors. The homogenization of the subtelomeric sequences due to preferential translocation of a certain class of Y′ elements can further promote perhaps more efficient Rad51-dependent BIR initiated within Y′ elements resulting in their amplification that is observed in type I survivors [10]. Remarkably, the strains that harbour Y′ elements at all chromosome ends due to previous history as a telomerase-defective survivor form survivors more readily when rendered telomerase-negative again [50]. This observation suggests that spreading of the Y′ elements to all ends could be one factor that limits the rate of type I survivor formation. Consistent with this proposition, we found that Rad59 accelerates type I survivor formation when type II survivor pathway is inhibited.
The initial step of Y′ acquisition involves heteroduplex formation by TG1–3 repeats that contain mismatches and is likely recognized by mismatch-repair proteins [51], [52]. It is therefore possible that the homeologous heteroduplexes formed during Rad59-dependent single-strand annealing are often rejected, and this may lead to reiterative rounds of repair synthesis which could effectively provide a way of TG1–3 tract expansion in the absence of telomerase. Indeed, we have identified a few clones with substantially long TG1–3 tracts at the breakpoint of recombination, which could have resulted from reiterative repair synthesis. We could speculate that once very long internal tracts are generated, they can be excised as circles via intramolecular recombination events (perhaps in a few steps) [14], [53]. The ultimate escape mechanism would be achieved via conversion to type II survivors, which are thought to amplify their terminal repeats via rolling circle replication. This could explain the apparent contradiction between the fact that circular DNA molecules are efficiently generated only when telomeric repeat tracts are abnormally elongated [54] and that only short telomeres engage in type II recombination [14].
Interestingly, we found that Rap1 was strongly enriched at the internal TG1–3 tracts located between the X and Y′ elements of the donor telomeres (Figure S9 and Table 1, see also [55]). Rap1 at these internal TG1–3 tracts was not relocalized upon critical shortening of telomeres [55]. It is conceivable that these internal Rap1 binding sites are also bound by the Sir proteins [56]. Notably, there is no homology between the modified TelVII-L and the Y′ donor chromosome end outside of the TG1–3 repeat tracts. Nevertheless, the apparent efficiency of Rad51-independent single strand annealing that relies exclusively on short homeologous TG1–3 repeat sequences is rather high, which might be due to spatial clustering of the telomeres in the yeast nucleus [57]. Indeed, for double-strand break repair, it has been recently shown that proximity of the donor sequence promotes homologous recombination [58]. The choice of the Y′ donor does not appear random. Whether this preference reflects the proximity between the chromosome ends in the nucleus or is influenced by other factors is currently unknown. Of note, among all chromosome ends, the VI-L end harbours one of the longest (139 bp-long) internal TG1–3 tract between the X and Y′ elements; and VII-R end is located at approximately the same distance from the centromere as the experimental VII-L end on the opposite arm of the same chromosome, which is consistent with Rabl configuration [59]. It has been also reported that Rap1 had the intrinsic ability to locally distort telomeric double-stranded DNA provoking local conformational changes characteristic of single strand [60]. Therefore, subtelomeric Rap1 binding could favour homologous recombination by creating local structures that are amenable to annealing with single-stranded overhang. In addition, Rap1 contains four putative SIMs, and thus, it may potentially recruit SUMOylated Rad52 and Rad59 to TG1–3 tracts.
Although our study focuses on short telomere repair by Y′ translocation, the BIR events leading to only terminal TG1–3 tract extension are also possible in telomerase-negative yeast. These events are readily detectable by Southern blot in survivors [14], [22] since they often result in large increase of terminal TG1–3 tract length, but this is not the case in pre-senescent cells. Direct sequencing of the terminal repeats showed that they do occasionally get extended in pre-senescent cells, although there is a controversy of whether the short or long tracts are preferentially extended [21], [22]. In any case, recombination between terminal repeats cannot indefinitely sustain proliferation of telomerase-negative cells, whereas Y′ translocation leads toward type I telomere formation, which likely requires additional changes such as chromatin structure alteration [19].
In summary, we have characterized a repair mechanism that acts upon short X-only telomeres in budding yeast lacking telomerase during presenescence. This repair mechanism does not require Rad51 and depends on annealing between short homeologous sequences which is stimulated by Rad59. Unlike a single unrepairable DSB, a single critically short telomere is not immediately lethal for a cell in the absence of telomerase as long as there is a reserve of long telomeres on other chromosome ends. Remarkably, Rad51 independence of the short telomere rescue pathway points to yet another problem caused by telomerase inactivation which resolution depends on Rad51-dependent HR, since deletion of RAD51 is known to cause early loss of viability of telomerase-negative cells.
All yeast strains used in this study were from the W303 background (see genotypes in Table S2). Strains used to analyze telomere rearrangement in the absence of telomerase were derivatives of the YAB892 and YAB893, which have 0 and 16 Rap1-binding sites, respectively, flanked by loxP sites at the modified telomere VII-L [31]. To obtain Tet-Off TLC1 derivatives, these strains were crossed to the tetO2-TLC1 strain. The inducible EST2 deletion derivatives (double Cre-loxP) were generated by first transforming the cells with the pDS381 plasmid [33] carrying EST2 gene linked to the ADE2 marker and the loxP-flanked ARS/CEN region, and then replacing the endogenous EST2 locus with KANr cassette. All other gene disruptions were carried out by PCR-based methods resulting in the replacement of targeted loci with the TRP1 marker.
To induce genome-integrated GALp-CRE for simultaneous TelVII-L shortening and loss of the plasmid-borne EST2, overnight cultures growing in SC medium lacking lysine and adenine and containing 2% raffinose were diluted (1∶20) into complete YEP medium containing 2% galactose. After 24 hours of induction the cells were diluted into YPD to prevent genome damage by Cre at non-specific sites. To inhibit telomerase in the Tet-off TLC1 strains, the cultures grown in SC medium lacking lysine and containing 2% raffinose were supplemented with doxycycline (10 µg/µl), 12 hours before induction of TelVII-L shortening by switching cells to galactose for the next 24 hours as described above. The doxycycline concentration was maintained constant throughout the experiments in all media.
To determine the ability of individual telomerase-negative cells to form microcolonies, the double Cre-loxP cells were micromanipulated onto a grid of YPD agar at 36 h after induction of Cre with galactose in liquid culture. Alternatively, the Tet-off TLC1 derivatives were micromanipulated onto the YPD agar freshly supplemented with Dox (20 µg/µl) at 24 h after Cre induction and 48 h after repression of TLC1 with Dox. The microcolony formation was monitored by counting the number of cells in each grid position at 2, 4, and 8 h after micromanipulation. The plates were incubated for additional 4 days to allow formation of visible colonies. The colonies which exhibited growth delay at the time or soon after micromanipulation were chosen for VII-L end analysis. In addition, the colonies with deeply nibbled edges and typically smaller size were included regardless of the growth delay. The clones that exhibited robust expansion and formed large colonies with smooth edges were used as controls.
To determine the state of the VII-L end before and after induction of Cre expression, genomic DNA was digested simultaneously with MfeI and PacI. The resulting fragments were separated by 0.9% agarose gel electrophoresis, transferred on Hybond N+, and hybridized with 32P-labeled 252 bp-long probe that was generated by PCR (see Table S1 for primers sequences) using YAB892 genomic DNA as a template. The rearrangement of the VII-L end in clonal populations recovered from transient arrest was analyzed similarly, except for two aliquots of DNA were digested separately with either MfeI or PacI. To determine the length of native telomeres, XhoI-digested yeast DNA was subjected to 0.8% agarose gel electrophoresis and hybridized with a 32P-labeled (TG1–3)n probe. The probes were labeled by random priming using Klenow fragment exo-, and all hybridizations were performed in Church buffer at 58°C.
To amplify the regions encompassing putative Y′ translocation junctions at either modified VII-L or native VI-R ends, the chromosome end-specific primers were designed to anneal ∼500 bp away from terminal TG1–3 repeats, and the Y′ element-specific primer sites were chosen in the centromere-proximal region that is conserved among all Y′ elements (Table S1). The Y′ sequence alignments are available from Ed Louis at http://www2.le.ac.uk/colleges/medbiopsych/research/gact/resources/yeast-telomeres. Analytical PCR was performed using Phusion DNA Polymerase (Thermo Scientific) in 1xHF buffer and 1 ng/µl genomic DNA purified via phenol-chlorophorm extraction. Primers were used at 500 nM, and the annealing temperature was set at 3°C above the Tm of the least stable primer.
For cloning, PCR across the junction was performed using Taq DNA polymerase to produce fragments with 3′ A overhangs, and the purified product was inserted into the pCR2.1 vector via one-step TA cloning (Invitrogen). The inserts were sequenced from both ends using M13F-20 and M13-26REV primers by Sanger method (at Beckman Coulter). As a rule, the read going through the TG-rich strand failed at the junction, so the entire sequence was assembled from both reads.
Rad59 was 13xMyc epitope–tagged at the C-terminus using one-step PCR. ChIP was performed as previously described [61]. Briefly, chromatin was cross-linked with 1% formaldehyde and sonicated to an average 200- to 500-bp DNA fragment size. After clarifying centrifugation, soluble chromatin was incubated with mouse anti-Myc tag monoclonal antibodies (9E10) and immunocomplexes were bound to magnetic Dynabeads Protein G (Novex). Following successive washes in standard solutions, Rad59-Myc bound chromatin was eluted from beads and incubated at 68°C to reverse crosslinks. DNA purified from the immunoprecipitates and inputs was quantified by real-time qPCR using chromosome end-specific primers listed in Table S1. The enrichment of the telomere-specific sequences bound by Rad59 was normalized to input and an unaffected GAL2 locus.
Rap1 ChIP experiments (duplicates) were performed with anti-Rap1 antibodies kindly provided by David Shore (University of Geneva). Rap1 ChIP and input DNA samples were hybridized to Nimblegen S. cerevisiae high density tiling arrays that were designed by us in collaboration with Frédéric Devaux (Ecole Normale de Paris) and Nimblegen to cover the entire genome. They contain 50 nt-long oligonucleotides separated by ∼15 nt-long gaps. The chip covers both strands of S. cerevisiae genome. Hybridizations and data analysis were performed by Nimblegen (Roche NimbleGen). Rap1 peaks were visualized with the signalMap software (Nimblegen) or with the Integrative Genomics Viewer (Broad Institute). The genome-wide Rap1 binding profiles were consistent with the recent published studies [55].
|
10.1371/journal.ppat.1006732 | Influence of an immunodominant herpes simplex virus type 1 CD8+ T cell epitope on the target hierarchy and function of subdominant CD8+ T cells | Herpes simplex virus type 1 (HSV-1) latency in sensory ganglia such as trigeminal ganglia (TG) is associated with a persistent immune infiltrate that includes effector memory CD8+ T cells that can influence HSV-1 reactivation. In C57BL/6 mice, HSV-1 induces a highly skewed CD8+ T cell repertoire, in which half of CD8+ T cells (gB-CD8s) recognize a single epitope on glycoprotein B (gB498-505), while the remainder (non-gB-CD8s) recognize, in varying proportions, 19 subdominant epitopes on 12 viral proteins. The gB-CD8s remain functional in TG throughout latency, while non-gB-CD8s exhibit varying degrees of functional compromise. To understand how dominance hierarchies relate to CD8+ T cell function during latency, we characterized the TG-associated CD8+ T cells following corneal infection with a recombinant HSV-1 lacking the immunodominant gB498-505 epitope (S1L). S1L induced a numerically equivalent CD8+ T cell infiltrate in the TG that was HSV-specific, but lacked specificity for gB498-505. Instead, there was a general increase of non-gB-CD8s with specific subdominant epitopes arising to codominance. In a latent S1L infection, non-gB-CD8s in the TG showed a hierarchy targeting different epitopes at latency compared to at acute times, and these cells retained an increased functionality at latency. In a latent S1L infection, these non-gB-CD8s also display an equivalent ability to block HSV reactivation in ex vivo ganglionic cultures compared to TG infected with wild type HSV-1. These data indicate that loss of the immunodominant gB498-505 epitope alters the dominance hierarchy and reduces functional compromise of CD8+ T cells specific for subdominant HSV-1 epitopes during viral latency.
| Most HSV-1 disease, including potentially blinding herpes stromal keratitis, results from sporadic reactivation of latent HSV-1 within sensory ganglia. Latently infected ganglia of humans and mice are associated with a persistent immune infiltrate of CD4+ and CD8+ T cells, with ganglionic CD8+ T cells capable of blocking HSV-1 reactivation from ex vivo cultures of latently infected ganglia. Here we show that in the absence of CD8+ T cells that recognize a single highly immunodominant epitope, the CD8+ T cells specific for the remaining 19 subdominant viral epitopes are not only numerically enhanced, but show more function within latently infected ganglia. We propose this work could lead to strategies that broaden and expand the functional CD8+ T cell repertoire within latently infected sensory ganglia, which may reduce the incidence of HSV-1 reactivation and recurrent disease.
| Primary herpes simplex virus type 1 (HSV-1) infection at peripheral mucosal sites leads to infection of innervating axonal termini, retrograde virus transport to nuclei of sensory and sympathetic neurons, and the establishment of a persistent latent state that is then maintained for the life of the host[1–3]. During latency, numerous factors, such as viral and host encoded miRNAs [4–6]and host epigenetic regulation [7–9], contribute to a repression of most lytic viral genes. During latency, abundant transcription is limited to a family of non-coding RNAs, the latency-associated RNA transcripts (LATs), which have been proposed to have multiple activities that promote latency and survival of the infected neurons [10, 11]. Sporadic or induced full HSV reactivation in humans can result in virus delivery to the periphery and development of recurrent disease. Recurrence in the eye is particularly problematic, since it may initiate a recurring immune-mediated herpes stromal keratitis (HSK) that causes progressive corneal scarring and opacity. Indeed, HSK is the most frequent infectious cause of blindness in the developed world[12].
Many lines of evidence now strongly suggest that lytic gene expression is not fully repressed during latency, but is rather in a dynamic state where sporadic lytic viral RNA and protein expression can occur in the neuron without virus production. It has been proposed that such sporadic HSV gene expression is largely outside of the typical α, β, γ cascade seen in productive infections [4, 8, 13–16]. A key decision is whether such sporadic events revert to a repressive state or subsequently progress to virus production. Evidence suggests that such chronic and sporadic viral gene expression in the latently infected ganglia is immune recognized, particularly by a persistent resident ganglionic CD8+ T cell population [17–19]. Indeed, the mouse model of HSV-1 latency has been under particular scrutiny, with the initial viral occupancy of the ganglia accompanied by a large infiltration of immune cells, including both CD4+ and CD8+ T cells. This immune infiltrate peaks near the onset of latency and then rapidly contracts, leaving a persistent low-level infiltrate that is maintained for the life of the host. Persisting ganglionic immune infiltrates associated with HSV-1 latency have also been seen in other model species and in humans [20–24]. The ganglionic CD8+ T cells in mice show markers of an activated effector memory phenotype, which is capable of reducing HSV-1 reactivation events in ex vivo cultures of latently infected ganglia[19, 25]. These observations promote a suspected role of adaptive cellular immunity in regulating the HSV-1 latent/lytic decisions in vivo. As such, gaining insights that allow improvement of the size, antigenic diversity, or the functional state of the ganglionic immune infiltrate may help increase protection from HSV-1 reactivation and subsequent disease [26].
The C57BL/6 (B6) mouse ocular HSV infection model has been particularly useful to explore cellular CD8+ T cell directed immunity, because the entire HSV-1 specific CD8+ T cell target repertoire has been described [27]. CD8+ T cells recognize short peptides of processed proteins (epitopes) that are presented bound to major histocompatibility complex class I (MHC-I) at the cell surface. In B6 mice, the CD8+ T cell repertoire developed to HSV-1 is highly skewed; a single immunodominant epitope on the essential viral glycoprotein B (gB) accounts for approximately half of all HSV-1 specific CD8+ T cells (gB-CD8s). These gB-CD8s are directed to amino acids 498–505 of gB (SSIEFARL, henceforth referred to as gB498-505). The other HSV-1 specific CD8+ T cell populations (collectively termed non-gB-CD8s) recognize 19 additional subdominant viral epitopes on only 12 viral proteins [27]. The approximate 1:1 ratio of HSV gB-CD8s to non-gB-CD8s is maintained systemically, in the corresponding lymph nodes, the spleen, and in the TG during acute peak infiltrate and in the contracted population during latency [28, 29]. However, while gB-CD8s during HSV-1 latency show an activated and highly functional phenotype that responds to antigen stimulation, non-gB-CD8 populations show a partial loss of function, such that a significant fraction do not respond to antigen stimulation. Loss of function in the non-gB-CD8 population resembles functional exhaustion in that it is associated with increased expression of the inhibitory receptor programed death-1 (PD-1) and is regulated by IL-10 [30,31]. Thus, the TCR repertoire of functional CD8+ T cells narrows by a process akin to functional exhaustion of subdominant CD8+ T cells within the latently infected TG.
The CD8+ T cell dominance hierarchy seen in the C57Bl/6 model and the strong dominance of the HSV-1 gB498-505 epitope is remarkable. It is not clear why the gB498-505 epitope is so immunodominant; the position of an epitope within the dominance hierarchy can be influenced by many factors [32]. Theoretical host disadvantages to strong immunodominance include a less diverse TCR repertoire, and the potential for viral escape from CD8+ T cell control through mutations in the immunodominant epitope [32]. The combination of a strongly immunodominant gB498-505 epitope and a completely defined TCR repertoire now makes HSV-1 an excellent model to investigate the effect of removing an immunodominant epitope on the resulting CD8+ T cell repertoire and changes associated with the latent state. Here, we fully characterize this response in acute and latent HSV-1 infections.
A series of gB mutants in the 498–505 amino acid region was generated and evaluated for recognition by gB-CD8s. The eight gB mutations (Table 1) included: mutations of the predicted MHC-I anchoring residues in the peptide (L8A, S1L, S1G, L8A/S1G, and L8A/R7K); mutations in the predicted T cell receptor binding region (F5L and S1G/I3A): and a mutation that changed the HSV-1 gB498-505 region to that of VZV, S1G/l3N/F5L/E4S (SIFE). The L8A mutation was previously reported (also referred to as L505A) in HSV-1 to abrogate gB-CD8 development upon flank skin infection of B6 mice [33]. We found that most of the eight mutant gB genes expressed a protein from plasmids that were detected by gB-specific antibodies of the size expected for gB (Fig 1). When each mutant gB protein was expressed in transfected B6WT3 fibroblasts, only wild-type (WT) gB could stimulate the production of intracellular IFNγ in an expanded population of gB-CD8s (Fig 1). While only a modest fraction of gB-CD8s showed activation, it was clear that all 8 mutations of the gB498-505 region abrogated gB-CD8 recognition.
HSV-1 recombinants containing each mutation were developed (Fig 2) by rescue of the growth of a gB-null-EGFP virus on gB-complementing Vero cells and all yielded virus that was able to form plaques on non-complementing cells. Recombinant HSV-1 with the following mutations in gB formed small plaques, indicating they were growth impaired: S1G, F5L, S1G/I3A, S1G/L8A, and S1G/I3N/F5L/E4S (Table 1). This was confirmed following ocular infection of B6 mice, where low viral replication was detected at 4 dpi within the TG of mice (Fig 3A). These impaired viruses were not studied further. However, HSV-1 S1L and L8A viruses could replicate to levels not significantly different from WT virus at 4 dpi in the TG (Fig 3A). They also establish equivalent latent genomic loads in the TG at 8 dpi (Fig 3B), and replicate to the levels of parental WT virus in both multi-step (infected at MOI of 0.01) and single step (infected at MOI of 10) growth curves in cultured Vero cells (Fig 3C and 3D). Since viral load and fitness in mice might influence CD8+ T cell hierarchy [34], we focused on these two HSV-1 mutants initially and then conducted detailed studies on HSV-1 S1L.
Following ocular infection of B6 mice, the peak CD8+ T cell infiltration in the TG occurs at 8 dpi and subsequently contracts to a low but persistent level that remains for the life of the host. The total CD8+ T cell infiltrates into TG of mice that received corneal infections with HSV-1 S1L and L8A were not statistically different from those induced by WT HSV-1 infection (Fig 4A). However, while approximately half of the CD8+ T cells infiltrating TG infected with WT HSV-1 [35] stained positively with gB498-505 H2-KB tetramers, CD8+ T cells infiltrating TG infected with HSV-1 S1L and L8A showed extremely low numbers of gB498-505 tetramer positive cells (Fig 4B). The gB498-505 tetramer positive cells were also virtually undetectable in spleens and the local draining lymph nodes (DLN) of S1L infected mice (Fig 4C and 4D). This suggests that there is an HSV-specific CD8+ T cell response in the TG that compensates for the loss of the immunodominant gB-CD8 population, as seen previously[33].
Failure of gB-CD8s to recognize the S1L mutation was further assessed by testing the ability of exogenous gB-CD8s to block HSV-1 S1L reactivation in ex vivo ganglionic explant cultures. TG of B6 mice harboring equivalent WT HSV-1 or S1L virus were excised at latency (34 dpi), dispersed with collagenase, depleted of endogenous CD8+ T cells and then plated in 1/5 TG cell equivalents per culture well, either alone or with 2x104 gB-CD8s from a previously described T cell clone, all under conditions that maintain T cell viability [25]. In the absence of CD8+ T cells, approximately 50–60% of TG cultures showed HSV-1 reactivation and virus release into the media, with reactivation frequency that was not statistically different in TG infected with WT and S1L virus (Fig 5A). These data establish that HSV-1 S1L possesses robust ability to establish latency and reactivate in ex vivo cultures. However, in the presence of gB-CD8s, the reactivation of WT HSV was nearly abrogated, as expected from similar previous studies[19, 21]. In contrast, reactivation of S1L was not affected by addition of gB-CD8s, as a similar proportion of cultures reactivated with or without gB-CD8s (Fig 5A). However, the endogenous CD8+ T cells in TG that were latently infected with S1L HSV-1 were still able to inhibit S1L reactivation from latency as effectively as endogenous CD8+ T cells in TG latently infected with WT HSV-1 (Fig 5B). Taken together, these results indicate that S1L reactivation events do not appear to be recognized by gB-CD8s, but the endogenous S1L-induced CD8+ T cell response is equally effective as that of WT at reducing reactivation events.
Our previous work demonstrated that the vast majority of CD8+ T cells in WT HSV-1 acutely infected TG at 8 dpi are HSV-1 specific [27]. We considered several possible explanations for the compensated CD8+ T cell TG infiltrate observed here (Fig 4) in the absence of gB-CD8s. We assessed if a CD8+ T cell response developed to the mutated forms of the gB498-505 epitope. B6WT3 fibroblasts were pulsed with the S1L, L8A, or WT gB498-505 peptides and co-cultured with CD8+ T cells from TGs of S1L, L8A, or WT infected B6 mice taken at 8 dpi. The ability of CD8+ T cells to recognize these peptides was determined by staining CD8+ T cells for intracellular IFNγ production (Fig 6). Ganglionic CD8+ T cells induced by a WT HSV-1 infection responded robustly to stimulation with WT peptide-pulsed fibroblasts, but failed to respond to B6WT3 cells pulsed with S1L and L8A peptides. This suggested there is little or no cross-recognition of the mutant peptides by gB-CD8s. Furthermore, the compensated ganglionic CD8+ T cell populations induced by HSV-1 S1L or L8A infection failed to produce IFNγ when stimulated with fibroblasts pulsed with any of the 3 peptides (Fig 6). This strongly suggested that the compensation of the ganglionic CD8+ T cell response was not due to the development of CD8+ T cells directed to the mutated gB peptides.
We next addressed the possibility that the compensated CD8+ T cell responses to S1L contained an expansion of CD8s directed to other HSV-1 epitopes. CD8+ T cells obtained from TG of mice infected with either WT or S1L HSV-1 were assessed both for total ganglionic CD8+ T cell infiltrates (Fig 7A) and for their ability to be stimulated by infected fibroblasts at different times post-infection (Fig 7B and 7C). The total ganglionic CD8+ T cell infiltrate for both WT and S1L showed a similar characteristic peak at 8 dpi and a subsequent contraction by 30 dpi (Fig 7A). However, at 16 dpi S1L infected TGs had a significantly smaller CD8+ T cell infiltrate, suggesting a more rapid CD8+ T cell contraction occurred.
To assess the response to non-gB498-505 HSV-1 epitopes, TG from WT or S1L infected mice at different times post infection were dispersed, stimulated for 6 hrs with B6WT3 fibroblasts infected with HSV-1 gB-null-EGFP virus, and intracellular IFNγ in CD3+CD8+ T cell populations was measured by flow cytometry (Fig 7B). We found a significantly higher frequency of IFNγ+ CD8+ T cells in S1L-infected TG compared to WT infected TG following stimulation with UV-irradiated B6WT3 cells infected with HSV-1 lacking gB. At the acute stage (8 dpi), the frequency of S1L stimulated CD8s was approximately twice that of the WT-stimulated CD8+ T cells. Intriguingly, in a WT infection, the fraction of ganglionic non-gB-CD8s responding to antigen drops over time. These data are consistent with our observation that in WT infected TG, CD8+ T cells specific for subdominant HSV-1 epitopes become functionally compromised by 30 dpi [30, 31]. However, we found that while non-gB498-505 CD8+ T cells in an S1L infection also dropped as latency was established, over 4x more non-gB-CD8s responded to infected target cells by producing IFNγ in this assay. This is more than can be explained by a simple doubling of stimulated cells that would be expected to fill the gB-CD8 compartment. This suggests that these non-gB-CD8s still target HSV-1 epitopes and are more functional in TG infected with HSV-1 S1L.
To separately assess CD8+ T cell responses to the gB immunodominant epitope in a similar assay, we stimulated CD8+ T cells obtained from S1L and WT latently infected TG for 6 hrs with B6WT3 cells infected with a recombinant pseudorabies virus (PRV) that expresses the HSV gB residues 494–509 containing the HSV immunodominant peptide (SSIEFARL) under control of the CMV promoter (PRV-gB) and measured intracellular IFNγ (Fig 7C). CD8+ T cells in S1L infected TG responded minimally to stimulation with the PRV-gB494-509 virus infected cells whereas CD8+ T cells in WT HSV-1 infected TG showed a robust response that maintained high levels throughout establishment of latency. These data agree with our previous finding that CD8+ T cells specific for the immunodominant gB498-505 epitope remain highly functional in TG latently infected with WT HSV-1 [30, 31]. They also demonstrate that CD8+ T cells in S1L-infected TG do not cross react to any presented PRV MHC-I epitopes in this assay. These results indicate that the compensatory response to S1L in the ganglia appears to reflect an increased number of CD8+ T cells directed to HSV epitopes other than gB498-505, and that these CD8+ T appear more functional in S1L infected TG compared to those in TG infected with WT HSV-1.
It was previously shown that activated, exogenously introduced non-HSV-specific OT-I CD8+ T cells could enter the TG during acute infections, but were not retained in the TG over time, presumably due to lack of cognate antigen recognition within the tissue [35]. The availability of an HSV-1 lacking the immunodominant peptide allowed us to assess the necessity of HSV-1 antigen presence in retaining ganglionic CD8+ T cell populations in vivo. We performed simultaneous corneal infections with the S1L virus lacking the immunodominant gB498-505 epitope in conjunction with ocular or flank infections with WT virus (Fig 8A). These infection models were designed to induce a systemic CD8+ T cell response to the gB498-505 epitope and observe the retention of that response under conditions where the epitope was or was not expressed in the TG. As expected, mice receiving corneal infections of either WT or S1L virus developed acute systemic and TG CD8+ T cell responses that contained or lacked the gB498-505 specific CD8+ T cell populations, respectively (Fig 8B and 8C). When a gB498-505 specific CD8+ T cell response was primed in S1L ocular infected mice by coinfection with WT HSV-1, either in the other eye or by flank infection, we noted that the infiltrate of the S1L infected TG contained a CD8+ T cell response to gB498-505. Indeed, the CD8+ T cell infiltration of the S1L and WT infected TG were quite similar at 8 dpi, with equivalent levels of both gB-CD8s and non-gB-CD8s. This data fits with the previously reported observation that acute ganglionic infection draws most or all activated CD8+ T cell populations into the ganglia [35]. However, a different pattern emerged by 30 dpi at HSV latency. In WT latently infected TG, irrespective of whether or not they received simultaneous corneal infection with S1L virus, an approximate 50:50 proportion of gB-CD8 to non-gB-CD8 T cells was retained during latency (Fig 8D). However, in S1L latently infected mice that were co-infected with WT virus ocularly or by flank infection, the gB-CD8 populations were greatly reduced by 30 dpi in the S1L infected ganglia. These results strongly support the conclusion that while most activated CD8+ T cell populations are able to infiltrate the ganglia at acute stages of infection, the maintenance of ganglia-resident HSV-specific CD8+ T cell populations requires antigen expression within the TG.
Virtually the entire CD8 antigenic repertoire to WT HSV-1 in B6 mice was recently defined [27]. This gave us the opportunity to examine the specific nature of the compensation to HSV-1 subdominant epitopes. We utilized the known HSV-1 subdominant epitope library to determine the size of each subdominant-epitope population that responds to peptide within the TG. HSV-1 induced CD8+ T cells infiltrating the ganglia at 8 dpi were evaluated for their ability to produce cytokines following stimulation with B6WT3 cells pulsed with each known epitope peptide. For WT, just over half of the CD8+ T cells in the TG of infected mice at 8 dpi responded to the immunodominant gB498-505 epitope, while the remaining CD8+ T cells responded at much lower levels to the 19 subdominant epitopes reported previously, with the addition of one minor epitope that is discussed below [27]. In contrast, CD8+ T cells infiltrating S1L infected TG showed no reactivity to the gB498-505 epitope, but showed a statistically significant increase in responses to most of the tested subdominant epitopes (Fig 9A). The response to a few subdominant epitopes remained statistically unaffected in the HSV S1L induced CD8+ T cell population, though an overall trend of increased levels was evident. Epitopes that featured prominently in the altered dominance hierarchy induced by HSV-1 S1L included RR1 (ribonucleotide reductase large subunit 1, UL39), with T cell epitope frequencies being the most abundant (RR1982-989) and fourth most abundant (RR1822-829) in the altered hierarchy. Interestingly, gB560-567 rose to position two in the hierarchy. Addition of the fractions responding to each peptide indicated that most of the CD8+ T cells were accounted for, with the fraction responding to each epitope adding to a total of 125.6 +/-33.8 percent(Fig 9B). This strongly suggests that the compensation was not a result of the development of significant CD8+ T cell populations to new epitopes, although we cannot exclude the possibility that minor populations directed to previously unreported epitopes were present in HSV-1 S1L infected mice. We also evaluated a subgroup of five of the specific epitopes for recognition by the HSV-1 L8A induced ganglionic CD8+ T cell response, and the results suggested this virus induced a similar compensatory response to that of S1L (Fig 9C). Finally, it was reported by Stock, et al [33]that the systemic response to the only known epitope at that time (RR1822-829) did not expand systemically; In contrast, we found that RR1982-989-specific CD8+ T cells, defined by tetramer, were expanded at both acute and late times (Fig 9D). These results indicate that the compensatory CD8+ T cell response to HSV-1 lacking the immunodominant gB epitope at 8 dpi is due to increased CD8+ T cell populations directed to the other tested subdominant epitopes.
Intriguingly, a quite different pattern emerged at 30 dpi, when HSV-1 is considered to be latent. We show the total average numbers of non-gB CD8+ T cells per latently infected ganglion that produce IFNγ+ in response to peptide presentation (Fig 10A). In a WT infection, the ganglionic CD8+ T cells to gB498-505 still responded to stimulation with peptide pulsed cells, and accounted for nearly half of the total CD8+ T cells detected in the ganglia (Fig 10B). However, the combined non-gB-CD8 populations in WT HSV-1 infected TG showed a much poorer response, with a total of only 27.3% of the subdominant CD8s producing IFNγ when stimulated with each of the subdominant peptides on B6WT3 cells. This data fits well with earlier reports indicating that a significant fraction of the non-gB-CD8s in the WT latently infected ganglia are not able to respond efficiently to antigen[30]. In contrast, the landscape of epitopes to which S1L induced CD8+ T cells can respond was changed dramatically, with 3–4 epitopes rising to prominence or co-dominance (Fig 10A). The most prominent were to UL28629-637(DNA packaging terminase), ICP8168-176, and the two large subunit ribonucleotide reductase epitopes; RR1982-989 and RR1822-829. This contrasts to the sub-dominant response hierarchy against WT HSV-1, which changed little over time. Secondly, when the CD8+ T cells from S1L latently infected ganglia that responded to peptide were totaled, 89.3 ± 16.5% of the CD8+ T cells were functionally able to produce IFNγ upon stimulation (Fig 10B). This does not simply reflect enrichment due to the lack of a gB498-505 response since the proportion of functional non-gB498-505 cells in this assay were significantly higher (Fig 10C). This is also in line with the data presented in Fig 7B showing that loss of the immunodominant epitope results in a functional increase in non-gB-CD8s at latency.
Comparing the total fraction of a CD8+ T cell population present by tetramer, to that identified by IFNγ production after peptide stimulation should tell us what percentage of that specific population is functional. We could not obtain working tetramers for most CD8 populations described here, but we were able to further characterize the functionality of RR1982-989 and RR1822-829 CD8+ T cell populations. This was done both by tetramer staining, and by determining multifunctionality (ability to produce multiple cytokines) following a peptide stimulation (Fig 10D). In a wild-type infected TG at latency, about 95 CD8+ T cells stained positive for these tetramers, with about 64% of these cells producing IFNγ, and 43% exhibiting multifunctionality (IFNγ+, TNFα+, and CD107a+) after stimulation. In contrast, an average of about 204 CD8+ T cells were tetramer positive in an S1L infected TG, with 84% capable of producing IFNγ, and 54% being multifunctional. This increase in functionality was statistically significant. Taken together, these results suggest that the presence of a strongly immunodominant CD8+ T cell population in latently infected ganglia adversely affects the functionality of those reactive to subdominant epitopes.
Given the general propensity of CD8+ T cells to recognize only a limited and highly restricted array of the potential epitope repertoire of a pathogen and develop a hierarchical response to them, it is important to understand how the presence or absence of one epitope, particularly an immunodominant epitope, influences the functionality and hierarchy of the CD8+ T cell response to others for a specific pathogen. Recent studies have revealed that HSV-1 specific CD8+ T cell responses developing to prior HSV infections in humans are highly limited, with only select proteins arising to dominance or co-dominance[36]. The unusually strong immunodominance to gB498-505 in the B6 mouse model of HSV-1 infection and the known full CD8+ T cell hierarchy makes it a highly suitable and manipulatable model to address the issue of how loss of one epitope affects the remaining response. Previously, the nature of CD8 compensation for loss of immunodominance could not be more precisely defined for HSV, because the HSV-specific CD8+ T cell repertoire in B6 has only recently been described [27]. That study demonstrated that the entire HSV-1 CD8+ T cell repertoire generated in the spleen was represented and not significantly modified in the TG, and that the vast majority of CD8+ T cells in the acutely infected TG at 8 dpi were HSV-specific [35].
Prior to this work, only one study had addressed the question of immunodominance loss in HSV-1, wherein B6 mouse skin infection with the HSV-1 L8A point mutation in the gB498-505 epitope was shown to induce a normal-sized expansion of HSV-specific CD8+ T cells in the spleen that lacked any specificity for the immunodominant gB498-505 epitope [33]. However, the authors concluded that the compensation reflected an additional response to previously unrecognized cryptic epitopes, based largely on the observation that CD8+ T cells to the one known subdominant HSV-1 epitope defined at that time, RR1822-829, were not increased in frequency of recognition by the compensated response. This finding is similar to a study by Holtappels et. al with murine cytomegalovirus (MCMV), wherein they demonstrated that deleting two MCMV immunodominant CD8+ T cell epitopes resulted in an altered immune response that was still protective, but contained small changes to known subdominant CD8+ T cells to known epitopes, with one specific epitope identified as rising in the absence of these deletions [37]. Similarly, Kotturi et. al deleted several immunodominant epitopes within Lymphocytic Choriomeningitis virus (LCMV), and show the altered immune response contained only limited increases in the subdominant CD8+ T cell hierarchy [38]. In our current study, we have found a somewhat different result in HSV-1 latently infected ganglia. Here we define the complete nature of the compensatory response in the TG and show a broad increase in the numbers and function of subdominant CD8 populations. Furthermore, assessment of the splenic response for one epitope indicates that at least part of the subdominant compensation occurs systemically in our model (Fig 9).
Of the several recombinant viruses developed to lack the gB498-505 epitope, two retained WT levels of pathogenicity in the mouse model. We chose S1L for most of the detailed studies, since L8A trended toward a marginally reduced ganglionic infiltrate in the TG (Fig 4) although differences were not significant. Reduced virus replication in the eye or TG may affect the level of antigen available for CD8+ T cell priming, recruitment and/or retention, as has been observed in the LCMV model [39]. S1L not only established latency with the same genomic loads as WT HSV-1, but also reactivated to the same efficiency in the absence of T cells. Both S1L and L8A viruses induced total ganglionic CD8+ T cell responses numerically identical to WT, despite the fact that the ~50% gB498-505 specific response was absent. Importantly, we show that neither S1L nor L8A virus induced a CD8+ T cell response to the native or mutated gB498-505 epitope, as demonstrated by priming and peptide stimulation assays (Fig 6). The lack of response to the S1L modified gB498-505 epitope was also established by demonstrating that (i) CD8+ T cells from mice infected with WT HSV failed to respond to the S1L modified gB peptide: (ii) CD8+ T cells in TG harboring the S1L virus showed negligible binding to tetramers containing the native gB498-505 epitope; and (iii) gB498-505-specific CD8+ T cells that effectively prevented reactivation of the parental wild-type HSV-1 strain had no effect in reducing reactivation of S1L from latently infected TG in ex vivo cultures. The epitope-specific nature of the reactivation blockade by gB-CD8s (Fig 5) also provides strong evidence that MHC-I presentation and cognate CD8+ T cell recognition are absolutely required to stop viral exit from latency in this model. This is consistent with our previous finding that HSV-1 specific CD8+ T cells from C57BL/6 mice can block HSV-1 reactivation in latently infected TG of C57BL/6, but not BALB/c mice [19], and contrary to a recent study suggesting that CD8+ T cells do not control HSV-1 reactivation events[40].
The nature of the compensated CD8+ T cell response in the acute infected TG to an immunodominant epitope-mutated HSV-1 reported here appears to be predominantly or entirely due to CD8s specific to other HSV-1 epitopes (Fig 7B). The compensatory increase of subdominant CD8+ T cell responses to epitope deleted viruses has been seen in other viruses, for example, following infection with an influenza virus that lacked an immunodominant epitope [41]. However, CD8+ T cell compensation was not observed following infection with epitope deleted LCMV, where the CD8+ T cell response was numerically reduced and delayed [42,43]. As noted by Stock et al [33] it is not clear if this difference in the ability to compensate for the loss of an immunodominant epitope reflects the difference between a local infection, like influenza and HSV-1, and a systemic infection like LCMV. Our studies using B6WT3 cells pulsed with each of the 19 known subdominant CD8+ T cell epitopes to stimulate IFNγ expression in CD8+ T cells indicate that most subdominant CD8 populations had an increased size to compensate for the loss of an immunodominant epitope. The added total CD8+ T cell response accounts for all (125.6% ± 33.8%) CD8+ T cells in S1L acutely infected TG, not only confirming that they are still HSV-1 specific, but suggesting negligible induction of CD8+ T cells specific for new cryptic HSV-1 epitopes. One possible exception to this is UL25338-345, which was a low frequency epitope identified, but not reported, in our previous study[27]. This CD8+ T cell population was identified as a small response (~0.2% of the total) in WT HSV-1 infections in our initial screen, but not in the subsequent repeat, suggesting a possible low frequency response in WT HSV-1. In an S1L infection, however, this epitope represented ~4% of the total acute TG infiltrate. We interpret this finding as an expansion of an existing low-frequency epitope-specific CD8+ T cell response. While we cannot exclude the possibility that S1L induced additional CD8+ T cell populations arising to new or cryptic epitopes, these would necessarily be small populations that would be difficult to identify via peptide library screening.
The change in the acute dominance hierarchy in the absence of the immunodominant gB498-505 epitope was not predictable. Compared to WT, the frequency of acute CD8+ T cells in the TG reactive to some epitopes such as UL39982-989 increased 4-fold, while the response to others such as UL29876-883 were only modestly increased. The response to others such as US44-12 showed little change compared to WT. However, 14 of the 19 subdominant CD8+ T cell populations in the S1L infected ganglia showed a significant increase (Fig 9). The reason for this differential rise of epitope-specific CD8+ T cell responses in the absence of the immunodominant epitope is not clear. Priming of the initial CD8+ T cell hierarchy as well as expansion of memory subsets are very dependent on the context of naïve T-cell-APC interactions, and T cells directly compete for antigen on APCs very early in this process[44–46]. CD8+ T cell population hierarchies also depend on naïve precursor frequency as well as peptide processing and MHC binding strength[38]. Selection in our model is not completely attributable to MHC binding capacity, since epitopes that changed little in the dominance hierarchy such as UL29876-883 and UL4083-90 were previously reported [27] to have higher MHC binding capacity (IC50 = 6.2 nM and 5.4 nM, respectively). Epitopes such as gB560 that rose substantially in the S1L dominance hierarchy had a much lower MHC binding capacity (IC50 = 9206 nM). The naïve precursor frequencies for the full HSV-specific response in our model has not yet been fully explored, but is outside the scope of this study. The kinetics of viral protein production are also likely to influence CD8+ T cell hierarchy, since a majority (~80%) of the epitopes recognized by CD8+ T cells in B6 mice are encoded by viral early (β) and leaky late (γ1) genes. This would be consistent with our previous finding that expressing gB as a true late (γ2) gene greatly reduced gB498-505 immunodominance [47]. Other studies using MCMV, Vaccinia, and Epstein-Barr virus epitopes also suggest that the context and kinetics of viral antigen expression contribute heavily to CD8 immunodominance, with the trend of stronger and earlier expression and presentation being important for the development of immunodominant CD8 populations[48–51]. However, we note differential rise to dominance of epitopes on the same viral protein in TG infected with S1L, suggesting that the level of expression of a particular viral gene product cannot by itself explain changes in immunodominance.
In the LCMV model it has been observed that when the virus is cleared acutely, the lasting memory CD8+ T cell hierarchy is maintained at similar proportions, but when a chronic strain is used, the hierarchy changes over time[52]. It is thus intriguing that the hierarchy at latency in the S1L infected TG is quite different from the hierarchy at acute infection. With our dual eye infection or flank infection models, we show that gB-CD8s are not retained in TG latently infected with the S1L virus that lacks this epitope (Fig 8), agreeing with previously published data demonstrating that antigen is necessary to effectively maintain exogenously added memory CD8+ T cells in nonlymphoid tissues[18,53]. Thus, differential expression of viral proteins or peptides during contraction/latency appears to be one among a complex set of mechanisms that define the CD8+ TCR repertoire in latently infected TG. This indicates a requirement for ongoing antigen detection to maintain memory CD8+ T cells in TG during latency, and could imply ongoing recognition by non-gB-CD8s in the TG of S1L latently infected mice during latency. Various levels of antigen recognition during the contraction phase may also explain the faster CD8+ T cell contraction observed in the S1L infected TG at 16 dpi (Fig 7A). Although the subset of possible ganglionic memory CD8+ T cells is shaped early in the lymphoid tissue, it is likely that competition for antigen recognition is a major factor responsible for determining which CD8+ T cell populations are retained efficiently and which subsequently arise to prominence in latently infected TG.
We present several lines of evidence that suggest the subdominant CD8+ T cells in the TG are more functional than those in the WT latently infected TGs. Specifically, in stimulations with gB-null HSV-1 infected cells, the fraction of responding CD8s in TG latently infected with S1L was four-fold higher than the fraction of responding cells in TG infected with WT virus (Fig 7B). This was considerably more than that expected from a simple doubling of numbers due to the compensated response for loss of the gB-CD8s. Secondly, the totaling of all the CD8+ T cells responding to peptide stimulation in S1L latently infected TG was approximately twice that of the fraction of subdominant CD8s responding in a WT latent infection (Fig 10B and 10C). Indeed, the response seemed to account for most of the CD8+ T cells in those S1L latently infected TG. Finally, measuring of two major subdominant populations directed to RR1 by parallel tetramer staining and analysis in response to peptide stimulation indicate that proportionally more of these RR1-specific cells show multifunctionality than do those same T cells from a WT latent infection (Fig 10D). Taken together, these data suggest that eliminating the strongly immunodominant gB498-505 epitope increases the function of remaining CD8+ T cells specific for some non-gB498-505 HSV-1 epitopes.
We previously demonstrated that i) subdominant CD8+ T cells have a higher frequency and express higher levels of PD-1 and show less apoptosis when PD-1 ligation is blocked within latently infected TG [31]; and ii) in vivo IL-10R blockade significantly increased the number, but not the frequency of subdominant CD8+ T cells in TG latently infected with WT HSV-1, and increased their ability to block HSV-1 reactivation from latency in ex vivo TG cultures[30]. These characteristics are consistent with an exhausted phenotype. Here and in previous studies [30,31] we show that TG that are latently infected with WT virus contain functionally exhausted subdominant CD8s, but near fully functional immunodominant gB-CD8s. Thus, in TG latently infected with WT HSV-1 the frequency of cytokine producing cells following peptide stimulation was similar to the frequency of the corresponding tetramer positive cells in the immunodominant population, but significantly lower in subdominant populations. Using the same analysis we now show significantly reduced functional exhaustion of the subdominant CD8+ T cells in TG that lack the gB498-505 immunodominance. We further demonstrate a significantly increased number of multifunctional subdominant CD8+ T cells in TG latently infected with HSV-1 lacking the immunodominant epitope when compared to those infected with WT HSV-1.
It is unclear why i) subdominant CD8+ T cells become partially exhausted in latently infected TG while immunodominant gB498-505-specific CD8+ T cells retain nearly full functionality; and ii) subdominant CD8+ T cells show less functional exhaustion in latently infected TG that lack the immunodominant population. Given the recent evidence demonstrating the unregulated nature of antigen expression in periods of “animation” during latency[15, 17], effective CD8+ T cell surveillance of a variety of viral targets may provide better monitoring of neurons in the early stages of reactivation. There is likely an interplay between the amount of leaky viral antigen presentation during latency and the number of specific CD8+ T cells that can possibly “see” that antigen. However, this interplay may also explain the phenomena observed here- wherein there exist an apparently very low frequency of latently infected neurons that express viral proteins[16]. The absence of the gB498-505 recognition and subsequent responses by immunodominant T cells may itself influence sporadic antigen expression during latency in the TG, which may change how other T cells see such expression in general. This may in turn affect the function of the subdominant T cells. Furthermore, in TG infected with WT HSV-1, the frequency of gB498-505-specific CD8+ T cells is high, reducing the likelihood of any one cell receiving the multiple exposures to cognate antigen required for functional exhaustion over time. In contrast, the frequency of subdominant CD8+ T cell populations is low, increasing the likelihood of such repeat encounters with cognate antigen. In TG harboring latent S1L virus, the subdominant populations are greatly expanded, which should alter the dynamics and reduce the likelihood of repeat antigenic exposure. As such, the mechanisms underlying the functional differences may be very complex and multifactorial.
In summary, we have precisely defined the nature of a compensatory subdominant T cell response infiltrating the TG, a site of HSV-1 latency in our model, and show that it is due to expansion of the non-gB-CD8 populations. However, during latency, this population alters to a state where there are multiple CD8+ T cell populations that rise to co-dominance that seems to be in part due to ongoing antigen expression within the TG during contraction. Furthermore, these CD8s at latency demonstrate an increased functionality compared to their WT-induced counterparts. This results in a broader repertoire of functional HSV-specific CD8+ T cells, which should increase the set of antigenic targets that can be used by CD8+ T cells to prevent HSV-1 reactivation from latency.
Vero cells (ATCC, Manassas, Virginia), SV40-transformed B6 embryo fibroblast cell line B6WT3 (MHC-I compatible with C57BL/6 mice; [54]), and gB-Vero (Vero cells stably transfected with a plasmid expressing gB from the native gB promoter; obtained from Dr William Goins, University of Pittsburgh) were grown in Dulbecco's modified Eagle's medium (DMEM) supplemented with 10% fetal bovine serum (FBS), penicillin-G (100 units/ml), streptomycin (100 mg/ml) and fungizone (250 mg/ml). All HSV-1 are based on the KOS strain originally obtained from a master stock obtained from P Schaffer. This strain was recently sequenced (GenBank JQ780693.1) [55]
Pseudorabies virus (PRV) used in this study were generated using a Becker-strain bacteria artificial chromosome (BAC) (kind gift from Lynn Enquist, Princeton University) [56] PRV expressing HSV-1 gB494-509 in a thymidine kinase (TK) knockout virus. This was developed in the E. Coli strain GS1783 by placing a kanamycin-resistance cassette containing monomeric red fluorescent protein in frame with TK at site 59,595 (Genbank JF797219.1) followed by a CMV promoter driving expression of 4 repeats of HSV-1 gB494-509. This cassette was amplified using the following primers: Forward, 5'-GCGGCAACCTGGTGGTGGCCTCGCT GGACCCGGACGAGCACATG GCCTCCTCCGAGGACGTCAT-3'; Reverse, 5’- TCAGGTAGCGCGACGTGTTGACC AGCATGGCGTAGACGTTCCTCG CGAGGGATCGGCTAGAGTC -3’ and inserted into the PRV BAC using recombineering [57]. The kanamycin-resistance cassette was resolved using two-step red-mediated recombination, with a resultant disruption of the TK gene. PRV was propagated to high titer in PK15 (porcine kidney 15, ATCC CCL-33) epithelial cells, purified as done for HSV, and gB-peptide expression confirmed by stimulation of gB-specific T cells, as described below.
DNAs were amplified by polymerase chain reactions (PCR) using a proofreading polymerase [Expand (Roche) or Primestar (Takara)] with “Hot start” conditions and in reactions containing 4% DMSO, as detailed previously [47]. A plasmid was derived in which genomic recombination into HSV-1 would replace the approximately N terminal half of the gB coding sequence with eGFP, using a pUC19-based plasmid (gC Rescue construct, Fig 1) detailed previously [47] containing HSV-1 DNA from 54,475bp to 56,464bp (sequences given in reference to HSV-1 KOS Genbank JQ780693.1) containing the gB promoter and partial gB coding region. The construct had a unique EcoRI flanking at 54,475 and unique HindIII site flanked 56,464, and a unique AvrII site at position 58,476bp, immediately upstream of the gB ORF start and designed not to alter the coding of the upstream UL28 protein. The AvrII-EcoRI fragment of this plasmid was replaced with a PCR fragment encoding gB residues 507 to the stop codon, generated using primers 5’-GCGCCTAGGCTCGGATCCCAGTTTACGTACAAC-3’ and 5’-GAGCGGAATTCATTTACAACAAACCCCCCATCA-3’ (restriction sites underlined). This construct (termed pgBp-gBend) was then further modified by inserting a PCR-amplified fragment containing eGFP using the primers: 5’ CCCTAGGCTACCTGACGGCGGGCACGACGG 3’ and 5’ CGTAGGATCCTTACTTGTACAGCTCGTC 3’. The BamHI-AvrII fragment was inserted into BamHI-AvrII digested pgBp-gBend, resulting in the flanking of eGFP by the gB promoter sequences, and sequence encoding gB residues 507–904 (Fig 1). When this linearized plasmid was then cotransfected with infectious HSV-1 DNA on gB-Vero cells, recombinant viruses showing eGFP positivity were selected and grown on complementing gB-Vero cells. Infectious HSV-1 KOS DNA was obtained using methods outlined previously [47]. Recombinant viruses (HSV-1 gB-null-EGFP) were then verified for correct DNA insertion by DNA sequencing of the junctions and by Southern blot analyses.
Plasmid gBp-gBend was digested with AvrII-SnaBI to remove the EGFP cassette, and replaced with PCR amplified sequence encoding gB residues 1–509 with altered residues in the gB SSIEFARL motif (Table 1). The primers shown in Table 1 were used in conjunction with the primer 5’ GCCCTAGGCTACCTGACGGGGGGCACGACGGGCCCCCGTAG 3’. Each PCR was then cloned as an AvrII-SnaBI fragment and sequence verified. Linearized plasmids were then individually cotransfected with HSV-1 gB-null-EGFP infectious DNA on Vero cells, and HSV-1 recombinants replicating on Vero cells, lacking GFP expression, and containing the mutations in the gB498-505 region were plaque-purified. We also derived expression plasmids for each epitope-mutated gB protein by placing the AvrII-EcoRI fragment containing the entire altered gB protein coding region into the vector pEGFP-C3, digested with NheI and EcoRI to place gB under the control of the human cytomegalovirus (hCMV IE) Immediate early promoter and SV40 polyadenylation signals.
Six to eight week old female C57BL/6 mice (Purchased from Jackson Labs) or gB-T1 mice [58] were used in these studies. HSV-1 infection by the ocular route was detailed previously [47], Briefly, corneas of anesthetized mice were scratched using a 30g needle with care not to enter the basement membrane. HSV-1 (1 x 105 PFU of purified HSV-1 in 3 μl RPMI media) was then applied and massaged into the eye, animals were then allowed to recover from anesthesia and returned to housing. Infections of the mouse flank were at a small (5mm x 5mm) region of skin denuded of hair. 1 x 106 PFU of purified HSV-1 was applied to abraded skin. Tissues for subsequent flow cytometry analyses were obtained from anesthetized mice injected i.p. with 0.3 ml of 1000 U/ml heparin (Sigma-Aldrich, St. Louis, MO) and then euthanized by exsanguination. Lymph nodes, TG and/or spleen were removed and subsequently digested in RPMI containing 10% fetal bovine serum and 400 U/ml collagenase type I (Sigma-Aldich) for 45–60 minutes at 37°C. Tissues were mechanically dissociated and triturated into single-cell suspensions and then filtered through a 40 μm nylon cell strainer (BD Biosciences, Bedford, MA). Spleens were treated with red blood cell lysis buffer (BD Pharm Lyse) for three minutes prior to analyses.
For ex vivo reactivation, TG samples were first rendered into single cell suspensions as just detailed, but without filtering. In cases where endogenous CD8+ T cells were depleted, we used antibody/complement mediated lysis using Low-Tox M rabbit complement (Cedarlane) and anti CD8, as previously described[20]. The efficiency of depletion was assessed by flow cytometry and was considered effective if >95% of the CD8+ T cells were depleted. Mock depleted suspensions used an IgG isotype control under the same conditions. Single-cell TG suspensions were then plated at one-fifth TG equivalents per well in 48-well culture plates in 400 μl of DMEM containing 10% FBS, 10 mM HEPES buffer, 10 U/ml recombinant murine IL-2 (R&D Systems), and 50 μM 2-mercaptoethanol. Where indicated, cultures were supplemented with exogenous expanded gB-CD8s made as described previously at 2 x 104 CD8+ T cells/well [24]. TG cultures were monitored for reactivation by testing 50 ul culture supernatant fluid for live virus by standard viral plaque assays as previously described [20]. The 50ul was replaced each time with fresh media. Supernatants were tested every two days for a total of ten days in culture. Data are represented as the percent of wells that were positive for viral reactivation.
CD8+ T cells specific for gB498-505 were expanded from TG preparations taken 8 days post infection (dpi) from mice infected with HSV-1, as detailed previously [24], or with collagen-dissociated suspensions of TGs. Cultures were maintained for up to 10 days, followed by MACS bead purification of the CD8+ T cells. Resulting populations were >95% CD3+, CD8+, and positive for gB498-505/H-2KB tetramer. For preparing target B6WT3 fibroblasts, the (WT) and mutant gB proteins were expressed in B6WT3 (1 x 105 cells) transfections with 5 μg of plasmids expressing each gB protein or the epitope mutant protein under the CMV-IE promoter. At 24 hours post-transfection, co-cultures for stimulations were established, with 5 x 104 expanded gB-CD8+ or CD8+ T cells obtained from the TG of HSV-1 infected gB-T1 mice added per 1 x 105 target B6WT3 cells. Co-cultures were maintained for 6 h in the presence of Brefeldin A, and then subsequently stained for surface CD45, CD8 and intracellular IFNγ and/or TNFα as markers of activation.
T cell phenotypic characterization by flow cytometry was performed essentially as detailed previously [26]. Single cell suspensions of TGs and spleens were stained with antibodies to CD45, CD3, and CD8α, and with tetramers for 1 hr at room temperature prior to fixing for 20 minutes with Cytofix/Cytoperm (BD Biosciences, Bedford, MA). Washed cells were then analyzed by flow cytometry. CD8+ T cell recognition of HSV-1 target antigens was determined by pulsing cultured B6WT3 fibroblasts with the respective peptide [26] at a concentration of 1 μg/ml for 30 min at 37°C/5% CO2. Alternatively, B6WT3 fibroblasts were infected for 6–12 hours at an MOI of 5 with recombinant HSV-1 or PRV, and then virus was UV-inactivated prior to stimulation. The dispersed TG or spleen cells were added to peptide-pulsed or infected fibroblasts in the presence of Brefeldin A and anti-CD107a (BD clone 1D4B) for 6 hr at 37°C/5% CO2. After stimulation, cells were stained for surface expression of anti-CD45 and CD8α, permeabilized and fixed using Cytofix/Cytoperm and then subjected to intracellular stain for IFNγ and TNFα. The peptides used in this work were detailed previously [26]. For each peptide, both TGs from a minimal total of five mice per peptide were separately analyzed for reactivity.
Phycoerythrin (PE)-conjugated or BV421-conjugated H-2Kb tetramers complexed with the gB498-505, RR1982-989, or RR1822-829 peptide were provided by the National Institute of Allergy and Infectious Diseases Tetramer Core Facility (Emory University Vaccine Center, Atlanta, GA). Efluor-450 conjugated anti-CD3 (clone 17A2) was purchased from eBioscience. Pacific-Blue-conjugated anti-CD8α (clone 53–6.7), APC-conjugated anti- IFNγ (clone XMG1.2), PerCP-conjugated anti-CD45 (clone 30-F11), PE-Cy7-conjugated anti-TNFα (clone MP6-XT22), APC-conjugated anti-granzyme B (clone GB11), and BD Cytofix/Cytoperm Fixation/Permeabilization Solution Kit were purchased from BD Pharmingen (San Diego, CA). The appropriate isotype control antibodies were purchased from the same company used for the reactive antibody and used as controls for intracellular staining. All flow cytometry samples were collected on BD FACSAria cytometer and analyzed by FACSDiva and/or FlowJo software.
HSV-1 genome copy number in infected TG was determined by quantitative real-time PCR as previously described using primers that recognize the sequences of the gH gene [59].
All statistical analyses were performed using GraphPad Prism software package. The specific statistical applications are shown in the legends to each figure.
All animal experiments were conducted in accordance with protocol # 15076444, approved by the University of Pittsburgh Institutional Animal Care and Use Committee. This protocol meets the standards for humane animal care and use as set by the Animal Welfare Act and the NIH Guide for the Care and Use of Laboratory Animals.
|
10.1371/journal.pntd.0000253 | A Single-Step Sequencing Method for the Identification of Mycobacterium tuberculosis Complex Species | The Mycobacterium tuberculosis complex (MTC) comprises closely related species responsible for strictly human and zoonotic tuberculosis. Accurate species determination is useful for the identification of outbreaks and epidemiological links. Mycobacterium africanum and Mycobacterium canettii are typically restricted to Africa and M. bovis is a re-emerging pathogen. Identification of these species is difficult and expensive.
The Exact Tandem Repeat D (ETR-D; alias Mycobacterial Interspersed Repetitive Unit 4) was sequenced in MTC species type strains and 110 clinical isolates, in parallel to reference polyphasic identification based on phenotype profiling and sequencing of pncA, oxyR, hsp65, gyrB genes and the major polymorphism tandem repeat. Inclusion of M. tuberculosis isolates in the expanding, antibiotic-resistant Beijing clone was determined by Rv0927c gene sequencing. The ETR-D (780-bp) sequence unambiguously identified MTC species type strain except M. pinnipedii and M. microti thanks to six single nucleotide polymorphisms, variable numbers (1–7 copies) of the tandem repeat and two deletions/insertions. The ETR-D sequencing agreed with phenotypic identification in 107/110 clinical isolates and with reference polyphasic molecular identification in all isolates, comprising 98 M. tuberculosis, 5 M. bovis BCG type, 5 M. canettii, and 2 M. africanum. For M. tuberculosis isolates, the ETR-D sequence was not significantly associated with the Beijing clone.
ETR-D sequencing allowed accurate, single-step identification of the MTC at the species level. It circumvented the current expensive, time-consuming polyphasic approach. It could be used to depict epidemiology of zoonotic and human tuberculosis, especially in African countries where several MTC species are emerging.
| The Mycobacterium tuberculosis complex (MTC) comprises several closely related species responsible for strictly human and zoonotic tuberculosis. Some of the species are restricted to Africa and were responsible for the high prevalence of tuberculosis. However, their identification at species level is difficult and expansive. Accurate species identification of all members is warranted in order to distinguish between strict human and zoonotic tuberculosis, to trace source exposure during epidemiological studies, and for the appropriate treatment of patients. In this paper, the Exact Tandem Repeat D (ETR-D) intergenic region was investigated in order to distinguish MTC species. The ETR-D sequencing unambiguously identified MTC species type strain except M. pinnipedii and M. microti, and the results agreed with phenotypic and molecular identification. This finding offers a new tool for the rapid and accurate identification of MTC species in a single sequencing reaction, replacing the current time-consuming polyphasic approach. Its use could assist public health interventions and aid in the control of zoonotic transmission in African countries, and could be of particular interest with the current emergence of multidrug-resistant and extended-resistance isolates.
| The Mycobacterium tuberculosis complex (MTC) comprises several closely related species responsible for strictly human and zoonotic tuberculosis (Figure 1). In addition to M. tuberculosis, which represents the leading cause of human tuberculosis worldwide and is now emerging as extensively drug-resistant tuberculosis strains [1], other MTC species have been found in patients, typically in African countries (Figure 2). Mycobacterium bovis is a re-emerging, zoonotic agent of bovine tuberculosis [2] whose prevalence probably depends on variations in direct exposure to cattle and consumption of unpasteurised dairy products [3]. The prevalence of Mycobacterium africanum type I (West Africa) and type II (East Africa) [4] has decreased in several African countries over the last decades [5],[6]. Mycobacterium canettii, a rare MTC species, has been isolated recently in patients exposed in Africa [7]. Mycobacterium microti, a vole and small rodent pathogen [8] that is closely related to the so-called Dassie-bacillus and infects small mammals in South Africa and the Middle East [9],[10], has been isolated in humans [11]. Mycobacterium caprae is a rare cause of tuberculosis in cattle [12],[13] and zoonotic tuberculosis in humans [14] while Mycobacterium pinnipedii has been isolated from seal lions and fur seals [15]. A recent description of the re-emergence of M. bovis in cattle, along with the direct interhuman transmission of this zoonotic organism [16] in a six-case cluster that included one death in United Kingdom [17], illustrates the potential of emerging and re-emerging zoonotic tuberculosis due to MTC species other than M. tuberculosis and the necessity for accurate species identification.
Accurate species identification of all MTC members is warranted in order to distinguish between strict human and zoonotic tuberculosis and to trace source exposure during epidemiological studies. Indeed, phenotypic methods of identification relying on colony morphology, oxygen preference, niacin accumulation, nitrate reductase activity, growth kinetics and resistance to thiophene-2-carboxylic acid hydrazide (TCH) and PZA [18] are hampered by slow growth of MTC members and subjective interpretation of colony morphology and cross-resistance to drugs [19]. They do not always allow unambiguous species identification in every case. Recent studies of MTC species responsible for animal and human tuberculosis in tropical countries have relied on molecular methods including mycobacterial interspersed repetitive-unit-variable-number tandem-repeat (MIRU-VNTR) typing, IS6110-RFLP and spoligotyping [20]–[22]. Molecular differentiation of MTC members has been complicated by low sequence variability at the nucleotide level, illustrated by a 85–100% DNA/DNA relatedness and a 99–100% 16S rDNA sequence similarity [23],[24]. Nucleic acid-based assays such as acridinium ester-labelled DNA probes (AccuProbe; Gene Probe Inc, San Diego, CA) have proven to be reliable tools for assigning an isolate to the MTC [25],[26], but they do not allow for identification at the species level. Molecular identification based on deleted regions (RD), RD1, RD9 and RD10 [27], are limited by the necessity of interpreting negative results in the case of the absence of a specific deletion. The detection of single nucleotide polymorphisms (SNP) in the pncA gene [28], the oxyR locus [29], the mtp40 gene [30], and the restriction fragment length polymorphism of the hupB gene [31] differentiated M. tuberculosis from M. bovis but not from other MTC species. The major polymorphism of tandem repeat (MPTR) sequencing differentiated M. tuberculosis (Sequevar long), M. bovis and M. microti (Sequevar Med-G), M. bovis BCG (Sequevar Med-C) and M. africanum (Sequevar short), but other MTC species were not studied [32]. The gyrB gene proved to be an effective target [33],[34], as an identification scheme has been proposed based on Pyrosequencing analysis of four single nucleotide polymorphism (SNPs) in gyrB [35], and a DNA strip based on gyrB is commercially available (HAIN Genotype MTBC DNAstrip test, Hain Lifescience, Nehren, Germany) [36]. Both approaches, however, fail to differentiate M. tuberculosis from M. africanum type II and M. canettii; and M. africanum type I from M. pinnipedii. IS6110-RFLP, VNTR typing and Spoligotyping [22],[37] emerged as reference methods to study the diversity of MTC species in resource-limited countries, despite the fact that these methods may not recognize rarely encountered species and may not appreciate the entire genetic diversity of strains, as they are not based upon the sequencing of molecular targets [38].
When investigating intergenic spacers in the genotyping of M. tuberculosis, we found that one spacer, previously identified as the Exact Tandem Repeat D (ETR-D) [39] and aliased Mycobacterial Interspersed Repeat Unit 04 (MIRU04) [40], exhibited a variable sequence among M. tuberculosis isolates. Analysis of this spacer had been previously shown to distinguish between M. bovis and the M. bovis BCG type [41]. We therefore further investigated whether sequencing the ETR-D could identify all of the MTC at the species level. In this study, we demonstrate that ETR-D sequencing offers a new tool for the rapid and accurate identification of MTC species in a single reaction.
M. tuberculosis CIP103471, M. bovis CIP105050, M. africanum CIP105147T (type I), M. bovis BCG vaccine strain type 105060, M. microti CIP104256T, M. canettii CIP140060001T, M. pinnipedii ATCC BAA-688, and M. caprae CIP105776T reference strains were purchased from the Collection Institut Pasteur (CIP, Paris, France) and American Type Culture Collection (ATCC, Rockville, USA). The following non-tuberculosis mycobacteria were tested in order to assess the specificity of ETR-D spacer sequencing: Mycobacterium avium IWGMT49 T, Mycobacterium intracellulare CIP104243 T, Mycobacterium chimaera CIP107892 T, Mycobacterium colombiense CIP108962 T, Mycobacterium haemophilum CIP105049 T, Mycobacterium ulcerans CIP105425 T, Mycobacterium xenopi CIP104035 T, Mycobacterium abscessus CIP104536 T, Mycobacterium chelonae CIP104535 T, Mycobacterium fortuitum ATCC49404 and Mycobacterium mucogenicum CIP 105223 T. Quality of DNA was controlled by parallel partial rpoB PCR amplification as previously described [42]. One hundred and ten MTC clinical isolates (Table 1) recovered from Microbiology Laboratory in Marseille (n = 102), from Institut Pasteur in Djibouti (n = 3) and from Institut de Pharmacologie et Biologie Structurale, Toulouse (n = 5) were also analyzed. All isolates were identified as members of the MTC by phenotypic characterization and a gene probe assay according to the manufacturer (AccuProbe; Gene Probe Inc, San Diego, Calif). This study was approved by the ethics committee of the Institut Féfératif de Recherche 48, Marseilles, France.
Phenotypic characterisation included colony morphology, a urease test controlled after 3 and 18 hour incubation, and oxygen consumption measured after inoculation of a 0.2 ml actively growing mycobacterial suspension into 40 ml of Middlebrook 7H10 into the Bactec 9000MB system (Becton and Dickinson, Le Pont de la Claix, France) after a 3-week incubation. Drug susceptibility tests for thiophene-2-carboxylic acid hydrazide (TCH) and PZA were performed as previously described [43].
The identification of reference strains and clinical isolates identified as M. bovis BCG type, M. canettii and M. africanum by ETR-D sequencing (see below) was confirmed by parallel reference molecular tests. Every isolate coated on beads was inactivated as previously described [44] and the DNA was extracted using a Qiagen kit (Qiagen, Courtaboeuf, France). DNA was used as a template for PCR amplification of pncA, oxyR, hsp65, gyrB genes and sequence analysis of MPTR was performed as previously described [28], [29], [32]–[34],[45] In addition, we sequenced the Rv0927c-pstS3 intergenic region in all clinical isolates identified as M. tuberculosis in order to identify the Beijing genotype [46]. Amplified products were visualized by agarose gel electrophoresis and direct sequencing was performed as described above. Sequences were edited using the Auto assembler program (Applied Biosystems, Courtaboeuf, France) and aligned using CLUSTAL W (http://pbil.ibcp.fr). Original sequences were deposited into GenBank (http://www.ncbi.nlm.nih.gov/sites/entrez/).
Amplification and sequencing of the ETR-D spacer located between the putative histidine kinase Senx3 upstream and the sensory transduction protein Regx3 downstream were done using direct primers: 5′-GTTGATCGAGGCCTATCACG-3′ and 5′-GAATAGGGCTTGGTCACGTA-3′. The PCR mixture contained 33 µl H2O, 5 µl 10× buffer (Qiagen), 2 µl 25× MgCl2, 5 µl 10× dDNTP, 1 µl forward primer, 1 µl reverse primer, 0.25 µl hotstart Taq (Qiagen) and 2 µl target DNA. Appropriate negative controls consisting of PCR mix without target DNA were also included. PCRs were performed using the following program: 15 min enzyme activation at 95°C, followed by 34 cycles consisting of 95°C for 30 s, 58°C for 30 s, 72°C for 1 min, followed by a 5 min elongation step at 72°C. After agarose gel electrophoresis, PCR products were purified and subjected to sequencing in both directions by using the BigDye Terminator 1.1 Cycle Sequencing kit (Applied Biosystems). Sequencing electrophoresis was performed on a 3130 genetic analyzer (Applied Biosystems). The sequences were edited using the Auto assembler program (Applied Biosystems) and aligned using CLUSTAL W (http://pbil.ibcp.fr). Original ETR-D sequences were deposited into Genbank (http://www.ncbi.nlm.nih.gov/sites/entrez/).
As for reference strains, M. tuberculosis exhibited eugonic growth that was inhibited by the presence of PZA but not by TCH and showed aerophilic growth on Middlebrook agar positive urease at 18 hours. M. bovis and M. bovis BCG type strains exhibited microaerophilic dysgonic growth, did not grow in the presence of TCH, but were resistant to PZA and exhibited a positive urease activity at 3 hours for M. bovis BCG type and at 18 hours for M. bovis. M. africanum type I differed from M. bovis by its susceptibility to PZA. M. canettii exhibited eugonic growth in the presence of PZA and TCH, and showed a positive urease activity at 3 hours and aerophilic growth on Middlebrook. M. microti, M. capare and M. pinnipedii exhibited eugonic growth that was inhibited by TCH and PZA, and a positive urease at 18 hours. As for clinical isolates, 101/110 of isolates were phenotypically identified as M. tuberculosis, 5 as M. canettii, 3 as M. bovis BCG type and one as M. africanum.
In all PCR experiments, negative controls remained negative. All reference strains and clinical isolates yielded an amplicon of the expected size when amplified for pncA, oxyR, hsp65, gyrB genes, Rv0927c-pstS3 intergenic region and MPTR. By comparison with M. tuberculosis, the 410-bp oxyR gene sequence exhibited a previously known A285G polymorphism in M. bovis and M. bovis BCG type [29] and a newly identified T136G polymorphism in M. canettii. The 561-bp pncA gene sequence exhibited a previously known G253C polymorphism in M. bovis and M. bovis BCG type [28] and a G222A polymorphism in M. canettii [47]. The 441-bp hsp65 gene exhibited a previously known T235C polymorphism in M. canettii [45] and a newly identified G376C polymorphism in M. africanum type I. The 1.020-bp gyrB gene sequence exhibited an identical sequence in M. tuberculosis, M. canettii and M. caprae, a previously known A756G polymorphism in M. bovis and M. bovis BCG type, a T675C polymorphism in M. microti, and an identical sequence was identified in common with M. africanum type I and M. pinnipedii [34]. Sequence analysis of MPTR (300-bp) exhibited a unique sequence for M. tuberculosis (Sequevar Long), M. africanum type I strain (Sequevar Short), M. bovis and M. microti (Sequevar MED-G) and M. bovis BCG type (Sequevar MED-C) [32]; MPTR sequencing newly identified C99T, G164C and A267G polymorphisms in M. canettii; the M. caprae strain exhibited Sequevar Long in common with M. pinnipedii and M. tuberculosis reference strains. Original sequences found in this study were deposited in GenBank under the following accession numbers (GenBank: EF656461, EF 656463, EF 656464).
Sequence analysis of clinical isolates using the five previous targets yielded four different profiles. One profile comprised 98 isolates identified as M. tuberculosis, including three isolates identified as W-Beijing strains using a G127A polymorphism in Rv0927c-pstS3 intergenic region, a second profile comprised five isolates identified as M. bovis BCG type; a third profile included five isolates identified as M. canettii and a fourth profile included two isolates identified as M. africanum type I.
For all ETR-D experiments, negative controls remained negative. All the non-tuberculosis mycobacteria yielded a negative ETR-D PCR amplification whereas they yielded an amplicon of the expected size through rpoB PCR amplification. The size of PCR products obtained with MTC reference strains varied from 497-bp for M. canettii, 545-bp for M. bovis, 564-bp for M. caprae, 598-bp for M. bovis BCG type, 651-bp for M. africanum type I, 805-bp for M. tuberculosis and 959-bp for M. microti and M. pinnipedii. Each of the eight reference strains exhibited a unique ETR-D sequence, exhibiting one to seven copies of a tandem repeat, six mutations and two deletions/insertions.
M. tuberculosis exhibited three different alleles combining two or five copies of a 77-bp repeat unit and one T/G SNP at the fifth base of the tandem repeat, in addition to one 53-bp repeat unit. The M. tuberculosis reference strain exhibited five copies of a 77-bp repeat unit followed by one 53-bp repeat unit copy; M. microti and M. pinnipedii exhibited seven copies of a 77-bp repeat unit and one 53-bp repeat unit, M. africanum type I exhibited three copies of a 77-bp repeat unit and one 53-bp repeat unit, in addition to one T/C polymorphism at position 75. M. bovis exhibited four copies of a 77-bp repeat unit and one 53-bp repeat unit in addition to an A/G SNP at position 773; M. bovis BCG type exhibited three copies of a 77-bp repeat unit in addition to an A/G SNP at position 773. M. caprae exhibited three copies of a 77-bp repeat unit, a 34-bp deletion and one 53-bp repeat unit. M. canettii exhibited one copy of a 77-bp repeat unit and three SNPs following the tandem repeat in addition to a 53-bp repeat unit. ETR-D sequences of MTC type strains were deposited in GenBank under the following accession numbers (GenBank: EU180228-EU180234). ETR-D sequencing identified 98/110 MTC clinical isolates as M. tuberculosis including 45 isolates presenting allele 1, 26 isolates presenting allele 2, and 27 isolates presenting allele 3; the ETR-D allele did not correlate with Beijing genotype (P = 0.2). 5/110 isolates were identified as M. bovis BCG type, 5/110 isolates as M. canettii, and 2/110 isolates as M. africanum type I. All unique ETR-D sequences were deposited into our freely available database at http://ifr48.timone.univ-mrs.fr/MST_Mtuberculosis/mst.
ETR-D identification was in agreement with phenotypic identification in 107/110 (97.27%) of clinical isolates. Three isolates phenotypically identified as M. tuberculosis were identified by ETR-D sequencing and reference molecular methods as M. bovis BCG type in two cases and M. africanum type I in one case. Reference molecular identification agreed with ETR-D identification in 100% (110/110) of clinical isolates.
Previous methods for MTC species identification either combined the amplification of several genomic regions in order to identify all species [27],[48] or analyzed one gene polymorphism to distinguish between only two species. ETR-D spacer sequencing herein developed proved to be specific for the MTC and allowed the differentiation of the 7/8 MTC species in a single reaction. Indeed all the non-tuberculosis mycobacteria yielded a negative ETR-D PCR amplification as previously described [49].
The fact that M. africanum type II was not included in the present study may not modify this conclusion. In fact, the taxonomic status of M. africanum type II has been disputed [50], but it is now regarded as a phenotypic variant of M. tuberculosis (genotype Uganda) [51],[52]. ETR-D sequencing agreed in all cases with reference molecular identification. In this study, new mutations were identified because some genes were sequenced for the first time in some MTC species including the oxyR gene and MPTR in M. canettii and the hsp65 gene in M. africanum type I (Table 3). ETR-D sequencing revealed that 3/110 clinical isolates identified as M. tuberculosis by phenotypic tests comprised two M. bovis BCG type isolate and one M. africanum type I isolate. The 497-959-bp size of ETR-D allows one-step sequencing using a modern capillary sequencer and software and may be easily sequenced using Pyrosequencing and additional internal primers. Cost was decreased in comparison with the current polyphasic approach and any microbiologist could compare the ETR-D sequence with those that we deposited in the versatile, freely accessible databank at http://ifr48.timone.univ-mrs.fr/MST_MTuberculosis/mst. This identification technique, based on PCR amplification, could be directly applied to clinical specimens exhibiting acid-fast bacilli.
ETR-D sequence identification relied not only on the variation in the number of tandem repeats illustrated by various PCR product sizes, as previously described [39] for M. tuberculosis, M. africanum, M. bovis group [41], but also on specific SNPs, which are stable events [53] accounting for 55.5% of genetic events observed in this study and on insertion/deletion events (accounting for 22.2% of genetic events). However, the ETR-D sequence was not correlated with the Beijing genotype as defined by Rv0927c-pstS3 intergenic region sequencing. This indicates that, although 3 ETR-D genotypes were found among M. tuberculosis isolates in this study, ETR-D sequencing alone cannot be used for genotyping. It is not surprising that the same, limited genomic region does not have the potential to identify at the species and strain levels. ETR-D sequencing provides, for the first time, a unique sequencing test capable of distinguishing all MTC species in a single step.
Accurate identification of MTC isolates at the species level is of particular interest in Africa where species other than M. tuberculosis were characterized in human tuberculosis and M. bovis remains a huge problem for cattle [21] (Figure 2). Their identification may direct specific epidemiological investigation. In Africa, the prevalence of M. bovis in human tuberculosis was correlated with the prevalence in the local cattle population [54]. Consumption of unpasteurised milk and of poorly heat-treated meat, and close contact with infected animals represent the main sources of infection for humans [3]. However, human to human transmission of M. bovis was recently reported in a 6-case cluster including one death due to M. bovis meningitis in United-Kingdom [17]. In addition, the emergence of MDR M. bovis has been documented, raising infection control in health care settings [55],[56]. M. bovis BCG type derived from the closely related virulent M. bovis after 230 serial passages had led to a considerably increased rate of disseminated BCG disease in HIV-infected infants reported in South Africa [57], although diagnoses were based on a few biochemical tests including the urease test and RD1deletion [58]. ETR-D sequencing allows unambiguous distinguishing of BCG type strains from M. bovis strains using a minute quantity of starting material. M. africanum identification indicated a tuberculosis microepidemic in a defined area when repeated isolation was observed [59]. Sporadic isolation of M. africanum strains has been reported in Europe and the United States, including outbreaks of multidrug-resistant (MDR) strains [60],[61]. In recent studies, variations in the reported prevalence of M. africanum among various African countries may also reflect difficulties in accurate identification of this species (Figure 2). M. microti, M. pinnipedii, M. caprae and M. canettii remain difficult to identify because of the extremely slow growth of these organisms, the difficulties with their identification under traditional bacteriological methods [62] and the fact that these recently described species have not been incorporated into current molecular identification schemes.
ETR-D spacer sequencing offers a new tool for the rapid and accurate identification of all MTC species in a single sequencing reaction without the need for expensive, time-consuming and potentially harmful polyphasic approaches. Its use could assist public health interventions and aid in the control of zoonotic transmission in African countries. Accurate identification of MTC isolates from Africa and tropical Asia would be of particular interest from the perspective of the current emergence of multidrug resistant and extended resistance isolates in these countries [63]. |
10.1371/journal.pntd.0001872 | Burden of Visceral Leishmaniasis in Villages of Eastern Gedaref State, Sudan: An Exhaustive Cross-Sectional Survey | Since December 2009, Médecins Sans Frontières has diagnosed and treated patients with visceral leishmaniasis (VL) in Tabarak Allah Hospital, eastern Gedaref State, one of the main endemic foci of VL in Sudan. A survey was conducted to estimate the VL incidence in villages around Tabarak Allah.
Between the 5th of May and the 17th of June 2011, we conducted an exhaustive door-to-door survey in 45 villages of Al-Gureisha locality. Deaths were investigated by verbal autopsies. All individuals with (i) fever of at least two weeks, (ii) VL diagnosed and treated in the previous year, and (iii) clinical suspicion of post-kala-azar dermal leishmaniasis (PKDL) were referred to medical teams for case ascertainment. A new case of VL was a clinical suspect with a positive rk39 rapid test or direct agglutination test (DAT).
In the 45 villages screened, 17,702 households were interviewed, for a population of 94,369 inhabitants. The crude mortality rate over the mean recall period of 409 days was 0.13/10'000 people per day. VL was a possible or probable cause for 19% of all deaths. The VL-specific mortality rate was estimated at 0.9/1000 per year.
The medical teams examined 551 individuals referred for a history of fever of at least two weeks. Out of these, 16 were diagnosed with primary VL. The overall incidence of VL over the past year was 7.0/1000 persons per year, or 7.9/1000 per year when deaths possibly or probably due to VL were included. Overall, 12.5% (11,943/95,609) of the population reported a past VL treatment episode.
VL represents a significant health burden in eastern Gedaref State. Active VL case detection had a very low yield in this specific setting with adequate access to care and may not be the priority intervention to enhance control in similar contexts.
| Visceral leishmaniasis (VL) is a life-threatening parasitic disease, transmitted by a sandfly. A survey was conducted to estimate the VL incidence in 45 villages located in the eastern part of Gedaref State, the main endemic focus of VL in Sudan. Between the 5th of May and the 17th of June 2011, we interviewed 17,702 households for a population of 94,369.
Sixteen individuals were diagnosed with primary VL through active case-detection, and 725 reported VL treatment over the past year. The overall incidence rate of VL over the past year was 7.0/1000 persons per year. The crude mortality rate over the mean recall period of 409 days was 0.13/10'000 persons per day. VL was a possible or probable cause for 19% of all deaths. Taking also into account the VL-specific mortality of 0.9/1000 per year, the incidence was estimated at 7.9/1000 per year. Overall, 12.5% of the population reported having been treated for VL in the past.
VL is a major public health issue in Gedaref. Active VL case detection had a very low yield in a context of adequate access to care. Such strategy seems redundant if patients already have access to care.
| Visceral leishmaniasis (VL), also called kala-azar, is a parasitic disease caused by members of the Leishmania donovani complex (L. donovani and L. infantum) and transmitted by the female phlebotomine sand flies of the genera Phlebotomus (Old World) and Lutzomyia (New World). It mainly affects areas in South Asia (India, Bangladesh, and Nepal) and Eastern Africa, where Sudan is the most affected country, followed by Ethiopia, Kenya, Somalia and Uganda. In Sudan, the causative agent is L. donovani, transmitted by the Ph. orientalis.
Gedaref State is the main endemic area of VL in Sudan. Passive detection figures from 1996 to 1999 have shown a mean yearly incidence between 6.6 and 8.4 VL cases per 1000 persons, with a large variation between villages (from 0 to 60 cases per 1000 persons per year) [1], [2]. Villages with high incidence are clustered along two rivers (Atbarah and Rahad), in areas of low altitude and high rainfall. Leishmanin skin testing, a marker of past exposure to the disease, has been shown to be positive in 21.6% of the population of the Atbarah area [3].
Many individuals infected by L. donovani have subclinical infection, while others develop clinical VL, a devastating illness that is usually fatal when left untreated. In Sudan, clinical signs develop gradually 2 weeks to 1 year after infection (in most cases after 2 to 4 months). Typical features are persistent fever, splenomegaly, weight loss and lymphadenopathies [4]. Post-kala-azar dermal leishmaniasis (PKDL) is a skin rash appearing after VL treatment, affecting up to 50% of treated cases in Sudan [5]. PKDL usually appears within 6 months after apparent cure and can last for months or years. Leishmania parasites can be found in smears of the skin lesions, and PKDL lesions are suspected to be an important parasite reservoir for human-to-human transmission. In Sudan, most lesions heal spontaneously. If not, PKDL treatment is challenging [6].
Some vector-control strategies such as indoor residual spraying have been shown to reduce the density of sand fly vectors in the Indian subcontinent [7], where the vector Ph. argentipes exhibits a behaviour different from the Sudanese vector. In Africa, bednet use has been shown to potentially reduce VL incidence [8]. Also, strategies to promote early detection and treatment have been shown to reduce case-fatality rates of VL in Brazil, where zoonotic VL is caused by L. infantum (syn. L. chagasi) [9]. Furthermore, early detection and treatment of anthroponotic VL patients is also believed to lower transmission through the reduction of the human reservoir [10]. This is supported by one pilot study that achieved good results using a combination of active detection, treatment and indoor residual spraying after a local outbreak of VL in one village located in the Bihar State of India [11]. However, although recommended, this strategy has never been formally evaluated in L. donovani endemic areas and is rarely implemented.
Since December 2009, Médecins Sans Frontières (MSF) has been diagnosing and treating patients presenting at Tabarak Allah Hospital, located in Al-Gureisha locality of the the Atbarah focus. MSF intended to conduct a cluster-randomized trial to evaluate the impact of an active VL and PKDL case detection strategy on the incidence of clinical VL. For the appropriate planning of this trial, a baseline survey was conducted in eligible villages around Tabarak Allah Hospital. The main objective of this survey was to estimate the incidence rate of VL over a one-year period at the village level. Additionally, we also aimed at retrospectively estimating the crude and VL-specific mortality rates, the proportion of VL cases missed by the passive case detection system in place, the proportion of the population treated for VL in the past, and the proportion of PKDL among patients previously treated for VL.
Ethical clearance was granted from the Sudanese National Ministry of Health's Research Ethics Review Committee. Written authorization to conduct the study was obtained from the Gedaref Ministry of Health and each head of village. Each head of household provided oral informed consent to the collection of demographical data, history of VL treatment, skin rash after treatment, and presence of fever of at least two weeks among household members. A referral form was given for each individual presenting with fever of more than two weeks, with suspicion of PKDL or having been treated for VL in the last year. The information included in these forms was not identifying and individuals were free to reach or not the medical team for clinical investigation. An additional oral consent was obtained from clinical suspects before testing for VL. The choice of oral consent was made because of the low literacy rate in the study area and the unlikelihood to easily find an impartial literate witness for each household. The Sudanese National Ministry of Health's Research Ethics Review Committee expressly approved the method of oral consent without use of a witness or written record of oral consent.
Between the 5th of May and the 17th of June 2011, we conducted an exhaustive door to door survey in the 45 villages of Al-Gureisha locality, covering a population of about 85,000 inhabitants. The survey villages were grouped into four geographical areas. Each area was surveyed by four field teams and one medical team.
Demographic information (age, sex, household composition on the day of survey and one year prior, number of births, deaths and movements within the past year) was collected by the field teams in each household. A household was defined as all people living together under the responsibility of one head of household and eating regularly together. For each household member, the history of VL treatment and possible subsequent PKDL was also recorded. The number and the causes of any death occurring in the past year were investigated in order to identify deaths possibly attributable to VL. Verbal autopsies were conducted for all reported deaths except for neonatal, delivery-related, and accidental deaths, as these were unlikely to be related to VL. Maternal deaths not directly related to delivery were investigated, as VL during pregnancy is known to be associated with increased treatment toxicity and mortality [12], [13].
Individuals with fever of at least two weeks duration, individuals diagnosed and treated for VL during the past year, and clinical suspects of either PKDL or VL relapse (independently of the time elapsed since treatment) were referred to the medical teams for clinical examination and case ascertainment. New clinical VL suspects (defined as fever for at least two weeks with at least one of the following: splenomegaly, lymphadenopathies or history of weight loss) were tested with an rK39 antigen-based rapid test (DiaMed IT-Leish) [14] and, if negative, with the direct agglutination test (DAT) [15], [16] for VL confirmation. A new VL case was defined as a clinical suspect who was confirmed either by the rK39 or the DAT. New VL cases, suspected VL relapses, and moderate and severe PKDL cases were referred to Tabarak Allah Hospital. Because of the self-healing nature of PKDL in Sudan and the potential toxicity of the recommended SSG treatment, mild PKDL cases were not offered SSG treatment [6] and therefore were not referred to Tabarak Allah Hospital.
To estimate the incidence rates at the village level, the population was exhaustively screened. We calculated a sample size of 266 deaths to estimate a proportion of deaths due to VL of 30% with a 5% precision (alpha 0.05). Based on an expected total number of deaths around 1500 (corresponding to an annual mortality rate of 0.5/10'000 persons per day), we planned to investigate the cause of every fifth death through verbal autopsy, using a systematic sampling procedure. All deaths were recorded consecutively on a tally sheet, with the death to be investigated pre-highlighted. As the data collected during the first three weeks of the survey showed a number of deaths much lower than expected, we later conducted verbal autopsies for every reported death. The analysis of the causes of death was weighted accordingly.
All verbal autopsies were reviewed independently by two clinicians experienced in VL and fluent in Arabic. In case of disagreement, the files were reviewed by a third expert clinician, with the help of a translator, and his verdict was final. Death was considered possibly due to VL if the respondent mentioned fever of at least two weeks duration and either one of the following: enlarged lymph nodes, a visible mass in the left upper part of the abdomen (spleen side), or weight loss, during the final illness of the deceased. Death was considered as probably due to VL if it occurred during treatment for VL (clearly mentioned by the relatives of the deceased) in a treatment facility offering reliable VL diagnosis (i.e. rk39 rapid test, DAT or microscopic examination of lymph node aspirate with quality control in place). If a death was reported to have occurred in another treatment facility during VL treatment, it was considered as possibly due to VL.
The event chosen to define the start of the recall period (covering the past year) was the presidential elections in Sudan, which occurred on the 10th and 11th of April 2010. The average recall period (referred hereafter as the “the past year”) was therefore 409 days. The end of the sesame harvest (end of October 2010) was used to define a 6-month recall period. VL incidence rate over the period was calculated by summing the new VL cases detected during the survey, the VL cases and the deaths possibly/probably due to VL reported over the recall period, divided by the mid-year population.
All documents were translated in Arabic and back-translated into English, and were subjected to pilot testing with subsequent update before the start of the survey. Data were entered in the EpiData software (EpiData, Odense, Denmark) by four data entry clerks. Data were analysed using the Stata 11 software (Stata Corporation, College Station, Texas, USA). Description of geographical information was performed using the QuantumGIS software, version 1.7.0. The coordinates of the Atbarah River were obtained by manually drawing along the river in Google Earth.
A total of forty-five villages were screened, corresponding to 17,965 households, 17,702 (98.5%) of which gave verbal consent to participate. The mid-year population was 94,369 inhabitants. The median household size at the time of the survey was 5 persons (interquartile range (IQR) 3 to 7 persons). The male/female sex ratio was 1.08. The median age was 15 years (IQR 7 to 30 years). The median population size by village was 1241 inhabitants (IQR 692 to 3113) at the time of the survey.
Overall, 12.5% (11,943/95,609) of the population reported having been treated for VL in the past, varying between 1.8% and 34.7% across villages. The medical teams assessed 725 individuals reporting VL treatment in the past year. Out of them, 125 (24%) mentioned a rash occurring within a median of 2 months after VL treatment (IQR 1 to 4 months) and lasting for a median of 3 months (IQR 1 to 10 months). Overall, PKDL was diagnosed in 260 cases (123 treated within the past year, 137 treated more than one year ago), corresponding to 0.3% of the survey population. The prevalence of PKDL cases ranged from 0 to 1.5% across villages. Most of the observed PKDL rashes were mild (81.5%) and none required treatment. In addition, the medical teams referred 40 patients for suspected VL relapse. Microscopic examination of lymph node aspirate was negative in 38 individuals and positive in 2 patients therefore diagnosed with VL relapse and treated.
The medical teams examined 551 subjects not previously treated for VL (Figure 1). Of these, 239 qualified as new clinical VL suspects, while the remaining 312 did not meet the case definition. Sixteen patients were ultimately diagnosed with primary VL (12 by rK39 rapid test and 4 by DAT), 85% of whom had actually sought care previously. Compared to the 725 cases treated in the past year, the active case detection therefore allowed to diagnose 16 (2%) additional of new cases. The age and sex distribution of the 741 VL cases newly diagnosed or treated in the past year is shown in Table 1. Males represented 54% of the cases, and 59.5% of the cases were aged from 5 to 14 years. The overall incidence rate of VL cases over the mean recall period of 409 days was 7.0/1000 persons per year. VL incidence rates by village varied between 0 and 23.0/1000 persons per year (Figure 2).
Five hundred and six deaths were reported, resulting in a crude mortality rate (CMR) of 0.13/10,000 persons per day. At the village level, the CMR varied between 0.02 and 0.30/10,000 per day, with a median of 0.14/10,000 per day. Accidental deaths represented 19.6% of all deaths, while neonatal and delivery-related deaths represented 14.8% and 1.8%, respectively. The remaining deaths were investigated by 171 verbal autopsies, corresponding to 299 deaths (32 sampled in the initial period of the survey given a weight of five, plus 139 for the remaining time of the survey). Taking into account the weighting, VL was a possible or probable cause of death in respectively 3.7% and 26.1% of verbal autopsies, or 2.4 and 16.6% when extrapolated to the total number of deaths. Other main causes of death were acute febrile illnesses (17.1%) and death related to chronic non-communicable diseases, mainly cardiovascular disease and diabetes (9.6%). Among the deaths probably/possibly due to VL (weighted n = 89), 45% occurred at home, 89% had received a medical treatment, and 53% had a clear history of VL treatment. The VL-specific mortality rate was 0.9/1000 persons per year. Taking into account these deaths possibly or probably due to VL, the overall incidence rate of VL cases over the recall period would reach 7.9/1000 per year.
In eastern Gedaref State, one out of 127 inhabitants was affected by VL over the past year. These incidence rates were lower than figures previously reported from the same region [1]. Still, a large proportion of the population (12.5%) has been affected by clinical VL in the past, reaching over one third in some villages. Also, one fifth of all deaths in the previous year may have been due to VL. However, there was no clear correlation between VL incidence and crude mortality rates at village level, which were overall lower than the reported average in Sudan [17], [18]. Still, VL represents an important burden in these communities, even between peaks of high incidence that occur approximately every six to ten years in Sudan [2], [18].
Interestingly, although gender is usually mentioned as a risk factor for VL [19], our data did not show a strong male predominance among cases compared to the general population. Project data from Tabarak Allah Hospital report 55% males among patients treated for VL, which is lower than figures reported from other treatment centres of Gedaref in the past [20], [21]. It is unclear whether this difference reflects differential access issues, changing epidemiology over time, or focal differences in transmission patterns. By contrast, age was clearly associated with VL: almost 60% of the cases were aged between five and 14 years, while this age group only represented 30% of the general population.
Up to one quarter of patients treated for VL within the past year reported some skin change consistent with PKDL, appearing within a median of two months after treatment and lasting for a median of three months. This is lower than reported in previous studies where up to 50% of treated VL patients developed PKDL [22], [23]. However, we only reported the proportion of PKDL among patients treated within the past year. Some patients were therefore likely to develop PKDL within the year after completion of the survey. Also, patients may not have reported mild and short-lasting PKDL. Although most PKDL cases are mild, they could still represent a reservoir of parasites, as L. donovani parasites can be detected in skin lesions [24]. None of the PKDL treatment currently available appears appropriate to treat mild cases, either because of toxicity (antimonials, conventional amphotericin B), teratogenicity (miltefosine) or cost (liposomal amphotericin B). As long as there is no definite evidence for the role of PKDL cases in the transmission chain, it is difficult to advocate for the development of better and simpler treatments for this condition.
Active VL case detection allowed us to detect an additional two percent of cases (n = 16/725). This appears as a very low yield for such a labour-intensive and costly operation. The survey was conducted at a time of the year when the number of new cases recorded at Tabarak Allah hospital is usually low. Thus, our results confirm that the incidence of clinical VL is low in May and June, and that this is not related to restricted access or use of health services. Active case detection may have detected more cases if it had been done from September to November, just after the rainy season, when incidence is believed to be higher and when many clinical cases have not yet sought medical care. However, the 2010–2011 Tabarak Allah hospital data neither show a clear seasonal trend, nor a large seasonal difference in delays for seeking care. Adequate access to care was confirmed by the short duration of symptoms reported by most VL cases on admission to Tabarak Allah Hospital (source: MSF project data). Also, most of the 16 new VL cases detected by the survey teams had actually previously sought care at health centres but were not adequately diagnosed with VL during that consultation. Our results show that when good-quality services are made accessible to a population that is well sensitized, active case detection might not be relevant. Based on our results, MSF decided not to proceed with the initially planned trial on active case detection, and not to recommend active case detection as a control strategy in the area.
A recent mathematical transmission model based on south Asian data suggested that VL treatment only had almost no effect on the overall intensity of transmission, which was mainly attributed to asymptomatically infected hosts [22], [23], [25]. These results cannot be extrapolated directly to Sudan where the VL epidemiology is very different, especially regarding the asymptomatic to symptomatic ratio that is much lower than in India [26]. Nevertheless it would be useful to adapt this model with Sudanese data, in order to guide efforts for VL control in the future.
The main limitation of our survey was the length of the recall period (over one year). Memory inaccuracies may have led to an overestimation of VL incidence, since VL cases that occurred more than one year prior the survey may have been reported as occurring within the past year. Some VL relapses treated during the past year may have been erroneously counted as new VL cases. Seasonal workers who left the area and developed VL later and elsewhere in Sudan were not included in the incidence results. These potential biases acted in opposite directions, which may have mitigated their impact on the estimated incidence of VL. Moreover, these biases being similar across villages, the relative differences in VL incidence still identify the villages most affected by VL in the study area.
The one out of five sampling of deaths submitted to verbal autopsy in the initial period of the survey may have caused some selection of the deaths investigated, which could have led to an overestimation of the proportion of deaths attributed to VL. However, the characteristics of deaths and the proportion attributable to VL were similar between the two periods, indicating no such phenomenon. We cannot exclude that some of the deaths attributed to VL may have been due to another disease causing similar symptoms, such as tuberculosis, advanced HIV infection, or cancer.
VL represents a significant health burden in the villages of eastern Gedaref State. The disease was among the major causes of death in the area. Active VL case detection through door-to door screening did not prove to be an efficient way to diagnose new VL cases likely due to current good access to VL care and relatively low prevalence of cases because the survey took place during a low-transmission period.
|
10.1371/journal.pcbi.1005453 | Effects of contact structure on the transient evolution of HIV virulence | Early in an epidemic, high densities of susceptible hosts select for relatively high parasite virulence; later in the epidemic, lower susceptible densities select for lower virulence. Thus over the course of a typical epidemic the average virulence of parasite strains increases initially, peaks partway through the epidemic, then declines again. However, precise quantitative outcomes, such as the peak virulence reached and its timing, may depend sensitively on epidemiological details. Fraser et al. proposed a model for the eco-evolutionary dynamics of HIV that incorporates the tradeoffs between transmission and virulence (mediated by set-point viral load, SPVL) and their heritability between hosts. Their model used implicit equations to capture the effects of partnership dynamics that are at the core of epidemics of sexually transmitted diseases. Our models combine HIV virulence tradeoffs with a range of contact models, explicitly modeling partnership formation and dissolution and allowing for individuals to transmit disease outside of partnerships. We assess summary statistics such as the peak virulence (corresponding to the maximum value of population mean log10 SPVL achieved throughout the epidemic) across models for a range of parameters applicable to the HIV epidemic in sub-Saharan Africa. Although virulence trajectories are broadly similar across models, the timing and magnitude of the virulence peak vary considerably. Previously developed implicit models predicted lower virulence and slower progression at the peak (a maximum of 3.5 log10 SPVL) compared both to more realistic models and to simple random-mixing models with no partnership structure at all (both with a maximum of ≈ 4.7 log10 SPVL). In this range of models, the simplest random-mixing structure best approximates the most realistic model; this surprising outcome occurs because the dominance of extra-pair contact in the realistic model swamps the effects of partnership structure.
| Pathogens such as HIV can evolve rapidly when the environment changes. One important aspect of a pathogen’s environment is the probability that an infectious contact (a sneeze for a respiratory disease, or an unprotected sex act for a sexually transmitted disease) encounters an uninfected person and thus has a chance to transmit the pathogen. As an epidemic grows the number of uninfected people shrinks, changing evolutionary pressures on the pathogen. While researchers have used models to explore pathogen evolution during epidemics, their models usually neglect important processes such as people entering and leaving sexual partnerships. We compared several evolutionary models for HIV that include partnership dynamics as well as sexual contact outside of stable partnerships. Models of intermediate complexity predicted lower virulence midway through the epidemic (a minimum of 15 years to progress to AIDS) than either more realistic models or simple models with no partnership structure (both with a minimum of 7.25 years to progress to AIDS), because random sexual contacts tended to wash out the effects of stable partnerships. Researchers trying to predict the evolution of pathogens must try to understand the implications of their modeling choices; models of intermediate complexity may not produce intermediate conclusions.
| The evolution of pathogen virulence (the harm done to the pathogen’s host) has both theoretical and, potentially, practical importance. Evolutionary theory suggests that pathogens with higher reproduction numbers (R 0)—the number of secondary infections caused by a single infected host over the course of its infectious period—will increase in prevalence relative to strains with lower reproductive ratios. Pathogens can increase their reproduction numbers either by increasing their transmission rate, the rate (per infected host) at which they infect new hosts, or by decreasing their clearance or disease-induced mortality rate, the rate at which hosts recover or die from disease. The trade-off theory [1] postulates that transmission and disease-induced mortality rates are both driven by the rate at which the pathogen exploits host resources for within-host reproduction, and that pathogen evolution will thus strike a balance between the pathogen’s rate of transmission to new hosts and its rate of killing its host (or of provoking the host’s immune system to eliminate it). Some biologists have criticized the tradeoff theory [2, 3], but others have successfully applied it to a variety of host-pathogen systems [4–7].
Fraser et al. have showed that HIV appears to satisfy the prerequisites of the tradeoff theory. The set-point viral load (SPVL: i.e., the characteristic virus load measured in blood during the intermediate stage of infection) is a measurable proxy for the rate of HIV within-host reproduction. Higher viral loads are correlated with faster progression to AIDS (higher virulence). Studies of discordant partnerships—stable sexual partnerships with one infected and one uninfected partner—have shown that SPVL is (1) positively correlated with transmission (people with higher SPVL transmit HIV to their uninfected partners sooner) and (2) heritable (when the uninfected partner does become infected, their SPVL is similar to their originally infected partner’s). Furthermore, the rate of increase in transmission has a decreasing slope as progression time decreases, fulfilling the requirements of the tradeoff theory [8]. Subsequent studies [9–11] used these data to parameterize mechanistic models of HIV virulence evolution, suggesting that HIV invading a novel population would initially evolve increased virulence, peaking after approximately 100-200 years and then declining slightly to a long-term stable virulence level.
The work of Shirreff et al. [9], and particularly the predicted transient peak in HIV virulence midway through the epidemic, highlights the importance of interactions between epidemiological and evolutionary factors [12, 13]. However, despite these studies’ attention to detail at the individual or physiological level, the population-level contact structures used in these models are relatively simple.
Many existing models of HIV eco-evolutionary dynamics use implicit models that incorporate the average effects of within-couple sexual contact—without representing the explicit dynamics of partnership formation and dissolution or accounting for extra-pair contact; agent-based formulations are more realistic, but can make it difficult to tease apart the reasons behind particular epidemic phenomena. Here we explore the effects of incorporating explicit contact structure in eco-evolutionary models.
Because our main goal is to explore how conclusions about virulence evolution depend on the way in which contact structure is modeled, we consider a series of models with increasing levels of complexity in the contact structure, but simplify some of the other epidemiological processes (such as the within-host life history of HIV). We evaluate our models across a wide range of parameters, using a Latin hypercube design; for each model run, we compute a set of metrics that summarize the evolutionary trajectory of SPVL over the course of the epidemic.
Our models explicitly track the evolution of the distribution of log10 SPVL (which we denote as α) rather than the rate of progression to AIDS itself (hereafter we use “virulence” to denote log10 SPVL). We use a single-stage model of HIV that assumes constant infectivity over the course of an exponentially distributed infectious period. This assumption contrasts with Shirreff et al.’s previous model which explicitly tracked three stages of HIV infection (primary, asymptomatic, and AIDS) and used a more realistic Weibull-distributed infectious period. We show below that our results are not overly sensitive to this simplification, although it could conceivably affect our conclusions about the evolution of virulence (e.g., Kretzschmar and Dietz [14] show that pair formation dynamics and multiple stages of infectivity have interactive effects on R 0). In Shirreff et al.’s model, the transmission rate and infection duration depend on virus load only during the asymptomatic stage. In order to adapt their parameterization to our single-stage model, we used their parameters to compute the SPVL-dependent transmission rate and duration during the asymptomatic stage and then derived the overall duration of the infectious period as the sum of the three stage durations and the average transmission rate as the duration-weighted average of the three stage-specific transmission rates. Thus the within-couple transmission rate, β (see “Contact Structure” below), for our models is given by:
β ( α ) = D P β P + D A ( α ) β A ( α ) + D D β D D P + D A ( α ) + D D , (1)
where the duration of infection (DP and DD) and rate of transmission (βP and βD) of the Primary and Disease stages of infection are independent of the host’s SPVL (Table 1 gives definitions, units, and values for all parameters). Following Shirreff et al., the duration of infection (DA) and rate of transmission (βA) for the Asymptomatic stage are Hill functions of the SPVL:
D A ( α ) = D max D 50 D k V ( α ) D k + D 50 D k , β A ( α ) = β max V ( α ) β k V ( α ) β k + β 50 β k , (2)
where V(α) = 10α.
In models that allow extra-pair contact, the uncoupled and extra-couple transmission rates (i.e., the rates of transmission among people outside of a stable partnership, or between people inside of a stable partnership and people other than their partner) are scaled by multiplying the within-couple transmission rate β by the contact ratios cu/cw and ce/cw (see S1 Appendix).
Over the course of infection, mutation occurs within the host. However, we follow Shirreff et al. in assuming that SPVL of the strain transmitted by an infected individual is determined by the SPVL at the time of infection and is not further affected by within-host mutation. Instead, the mutational effect is modeled as occurring in a single step at the time of transmission.
First, the distribution of log10 SPVL is discretized into a vector:
α i = α min + ( α max - α min ) i - 1 n - 1 i = 1 , 2 , 3 , … n . (3)
We have experimented with varying degrees of discretization in the strain distribution (i.e., values of n); in our model runs comparing results with Shirreff et al. [9] (Fig 1) we use n = 51 (i.e. a bin width of 0.1 log10 SPVL for α), but reducing n to 21 (bin width = 0.25 log10 SPVL) makes little difference; we use this coarser grid for all other simulations reported.
We then construct an n × n mutational matrix, M—which is multiplied with the transmission term—so that Mij is the probability that a newly infected individual will have log10 SPVL of αj given that their infected partner has log10 SPVL of αi. Finally, the probabilities are normalized so that each row sums to 1:
M i j = Φ( α j + d / 2 ; i ) - Φ ( α j - d / 2 ; i ) Φ ( α max + d / 2 ; i ) - Φ ( α min - d / 2 ; i ) , (4)
where Φ(x;i) is the Gaussian cumulative distribution function with mean αi and variance of σ M 2, and d = (αmax − αmin)/(n − 1). Unlike Shirreff et al., who allowed for variation in the expressed phenotype (SPVL) of each genotype, we use a one-to-one genotype-phenotype map. Thus there is a single value for within-couple transmission rate and for progression rate corresponding to each SPVL compartment in the model:
β i = β ( α i ) , λ i = 1 D P + D A ( α i ) + D D . (5)
We developed seven multi-strain evolutionary models covering a gamut including Champredon et al.’s relatively realistic [15] and Shirreff et al.’s relatively simple [9] contact structures, each of which is based on different assumptions regarding contact structure and partnership dynamics. Specifically, we focus on the effects of the assumptions of (1) instantaneous vs. non-instantaneous partnership formation; (2) zero vs. positive extra-pair sexual contact and transmission; and (3) homogeneous vs. heterogeneous levels of sexual activity on the evolution of mean log10 SPVL.
Our first four models (Fig 2) explicitly consider partnership dynamics [15]. The first (Fig 2d) assumes non-instantaneous partnership formation (i.e. individuals spend some time uncoupled, outside of partnerships) and consists of five states that are classified by infection status and partnership status; single (uncoupled) susceptible individuals (S), single infected individuals (I), concordant negative (susceptible-susceptible) couples (SS), discordant (susceptible-infected) couples (SI), and concordant positive (infected-infected) couples (II). The rates of pair formation are based on the numbers of uncoupled susceptible and infected individuals and the pair-formation rate; partnerships can either dissolve into singletons or be transformed into other types of partnerships by infection of one partner. This model also includes extra-pair contact with both uncoupled individuals and individuals in other partnerships (we denote it “pairform+epc”), so that susceptible uncoupled individuals and susceptible partners in any type of partnership can be infected by infected uncoupled individuals or infected partners in any type of partnership.
Specifically, single individuals (S and I) form partnerships at a per capita rate ρ, and partnerships dissolve at a rate c. Infected individuals in a discordant partnership infect their susceptible partner at a rate β (within-couple transmission rate) and susceptible individuals outside the partnership at a rate ce (extra-couple transmission rate). Infected individuals in seropositive (II) partnerships can also infect any susceptible individual at rate ce. Likewise, single infected individuals (I) can infect any susceptible individuals (single individuals S, or susceptible members of SS or SI partnerships) at a rate cu through uncoupled mixing. This parameterization follows Champredon et al.; we have adapted some of the details of their model to a multi-strain scenario, so that we track (for example) a matrix IIij that records the number of concordant, HIV-positive partnerships in which the two partners have log10 SPVL of αi and αj.
Our second model (“pairform”, Fig 2c) only considers within-couple transmission, in which case infection can only occur within a serodiscordant partnership; that is, we set ce and cu to zero.
Our third and fourth models, which are intended to bridge the gap between models with fully explicit pair-formation dynamics and the simpler, implicit models used by Shirreff et al. [9], assume instantaneous partnership formation (“instswitch”). The compartmental structure thus omits the single states S and I, comprising only the three partnered states: SS, SI, and II. Like the first two models, this pair of models differs in their inclusion of extra-pair contact: the third model (“instswitch+epc”, Fig 2b) includes extra-pair contact (now only with individuals in other partnerships, since uncoupled individuals do not exist in this model) while the fourth (“instswitch”, Fig 2a) only considers within-couple transmission. Although these models can also be implemented by setting the partnership formation rate of the explicit partnership models to a high value (we have tested that both methods in fact produce same results), we model instantaneous partnership formation models independently so that scaling the partnership formation rate during model calibration (see Simulation runs below) does not affect the eco-evolutionary dynamics.
The fifth and sixth models represent extreme simplifications of sexual partnership dynamics. The fifth (“implicit”) is an implicit serial monogamy model based on the epidemiological model used by Shirreff et al. [9]. It is a random-mixing model that explicitly tracks only the total number of susceptible and infected individuals. However, to reflect the effect of partnership structure, it uses an adjusted transmission rate derived from an approximation of the basic reproduction number of a serial monogamy model with instantaneous pair formation [16]. The sixth model (“random”) is a simple random-mixing model.
Lastly, we incorporated heterogeneity in sexual activity into the models. Individuals are divided into different risk groups based on their level of sexual activity; we scale all aspects of sexual activity, assuming that sexual activity level in both within- and extra-couple contacts is directly proportional to number of non-cohabiting (extra-couple and uncoupled) partners per year [17] (see S1 Appendix for full model details). We assume random activity-weighted mixing between risk groups [18]. (In the main text we focus on the model with non-instantaneous pair formation, extra-pair contact (“pairform+epc”) and heterogeneous sexual activity, which we denote as “hetero”; Fig D in S2 Appendix presents results on the effect of adding heterogeneity to other model variants.) While this model lacks some important elements, such as age-structured mixing patterns, needed for realistic models of HIV transmission in sub-Saharan Africa, it represents a first step toward assessing the effects of epidemiological complexity. As even the models shown here push the limits of compartmental-based models (the heterogeneity model comprises 24530 coupled ordinary differential equations), adding further complexity will probably require a shift to an agent-based model framework, as well as considerable effort in model calibration [10, 19, 20].
For simplicity (and following Shirreff et al.), all of our base models use an SIS (susceptible-infected-susceptible) formulation, where there is no natural mortality (and no explicit introduction of newly sexually active individuals into the susceptible pool). Individuals who die from AIDS are immediately replaced by an individual in the uncoupled-susceptible compartment. While admittedly unrealistic, this approach is reasonable given that (1) the natural mortality rate is low relative to the epidemiological dynamics and (2) the infectious period is long, so that the overall rates of disease-induced mortality and recruitment of newly sexually active individuals would roughly balance. To check the importance of this assumption, we also built models with vital dynamics where individuals dying from AIDS are removed from the population, with constant recruitment rates and constant low per capita natural mortality rates; this additional structure had only minor effects on the results.
Choosing the initial conditions for the simulations is challenging. In many modeling studies, researchers are primarily interested in equilibria (or other long-term dynamical attractors such as limit cycles) and are exploring models that have a single stable attractor, so initial conditions can be ignored as long as we run models for long enough. In eco-evolutionary dynamics, however, the initial conditions do affect our conclusions. We have no empirical information that would justify a particular choice of the fraction infected and the mean and variance of the distribution of SPVL at the point when the pandemic strain of HIV-1 entered the human population; in any case, the level of realism of our model does not support such a detailed consideration of the early dynamics of HIV. In most cases, we started with an initial log10 SPVL of 3.0, to match the value used by Shirreff et al. [9]. Shirreff et al. use an initial prevalence I(0) = 10−3; because we calibrated parameters based on the initial epidemic growth rate (see “Simulation runs” below), we set I(0) to 10−4 for most runs to ensure that the exponential growth phase lasted long enough for reliable estimation of the initial growth rate.
S1 Appendix provides further details on initial conditions, such as the initial distribution of SPVL around the mean and the distribution of initial infected density across single people and different partnership types.
We ran most of our models across a wide range of parameters, as described in the next (Latin hypercube sampling) section. In several cases, however, we inspected only a few parameter sets, to qualitatively assess the sensitivity of the models to initial conditions or to model structure. In particular, we tested model sensitivity to the initial prevalence, I(0), and initial mean log10 SPVL, α(0), using baseline values of all parameters (Table 1). Using baseline parameter values, we also ran all four basic model structures (Fig 2) with vital dynamics and with heterogeneity in sexual contact, to assess the sensitivity of our results to these phenomena.
Despite considerable effort [15, 16], the parameters determining the rate and structure of sexual partnership change and contact are still very uncertain; this uncertainty led Champredon et al. [15] to adopt a Latin hypercube sampling (LHS) strategy [21] that evaluates model outcomes over a range of parameter values. In order to make sure that our comparisons among models apply across the entire space of reasonable parameter values, and in order to evaluate the differential sensitivity of different model structures to parameter values, we follow a similar protocol and perform LHS over a parameter set including both the early- and late-stage transmission and duration parameters (βP, DP, βD, DD) and contact/partnership parameters (ρ, c, cu/cw, and ce/cw). For the heterogeneity model, the mean and squared coefficient of variation (CV) for the number of non-cohabiting partners are sampled as well. We do not allow for uncertainty in parameters that are directly related to the evolutionary process (βmax, β50, βk, Dmax, D50, Dk, σM), instead using Shirreff et al.’s point estimates throughout [9].
Latin hypercube sampling is done as in Champredon et al. [15]. For each parameter, z, its range is divided into N = 1000 equal intervals on a log scale:
z i = exp log ( z min ) + [log ( z max ) - log ( z min )] i - 1 N - 1 i = 1 , 2 , 3 , … , N . (6)
Random permutations of these vectors form columns in a sample parameter matrix; each row contains a different parameter set that is used for one simulation run.
Table 1 gives the ranges of the model parameters used for LHS. Ranges of parameters controlling contact and partnership dynamics (ρ, c, and ce/cw) are taken from Champredon et al. [15], whereas those controlling infection (βP, DP, βD, and DD) are taken from Hollingsworth et al. [16]. The remaining parameter values are taken from Shirreff et al. [9].
One new parameter in our model, the ratio of uncoupled to within-couple transmission cu/cw, is needed to more flexibly contrast uncoupled and extra-couple transmission dynamics within multi-strain models (see S1 Appendix). Since it appears neither in either Shirreff et al. nor Champredon et al.’s models, we need to pick a reasonable range for it. Champredon et al. [15] assume that the effective within-couple contact rate and effective uncoupled contact rate have the same range of 0.05—0.25. Given Champredon et al.’s parameter range, the possible maximum and minimum values of cu/cw are 5 and 1/5. Therefore, we use 1/5-5 as the range for the parameter cu/cw. Although this adds more uncertainty to the parameter cu—Champredon et al.’s range implies a 5-fold difference whereas ours gives a 25-fold difference—we consider the wider range appropriate, as little is known about the uncoupled transmission rate.
Two parameters, mean and the squared coefficient of variation (CV) of number of non-cohabiting partners, are sampled for heterogeneity in sexual activity. To allow for a wide range of uncertainty, range for the mean number of non-cohabiting partners was taken from unmarried men, as that was the group with the largest variability [17]. Omori et al. [17] give a very wide range for the coefficient of variation (≈ 0—20, corresponding to squared CV range of 0-400): we narrowed this range for CV2 to 0.01-100. At the bottom end of the range, an observation that a group behaves perfectly homogeneously (CV = 0) is likely to be a sampling artifact; at the upper end, the estimate is also likely to be noisy because of the low mean value among married females (who have the largest range of CV). We assume that the number of non-cohabiting partners follows a Gamma distribution.
One of the hardest parts of model comparison is finding parameter sets that are commensurate across radically different model structures. For the most part, our models are too complex to derive analytical correspondences among the parameters for different models. Given a numerical criterion, such as r (initial exponential growth rate) or R 0 (intrinsic reproductive number), we can adjust one or more parameters by brute force to ensure that all of the models match according to that criterion. While R 0 is often considered the most fundamental property of an epidemic, and might thus seem to be a natural matching criterion, here we focus on matching the initial growth rate r for several reasons. First, our primary interest is in the transient evolutionary dynamics of virulence, which are more strongly affected by r than R 0. Second, r is more directly observable in real epidemics; r can be estimated by fitting an exponential curve to the initial incidence or prevalence curves [22], while R 0 typically requires either (1) knowledge of all epidemic parameters or (2) calculations based on r and knowledge of the serial interval or generation interval of the disease [23]. Thus, we scale parameters so that every run has the same initial exponential growth rate in disease prevalence.
In order to allow for all models to have equal initial exponential growth rate, r, we need to pick a parameter, s, such that lims→0 r(s) = 0 and lims→∞ r(s) = ∞. As adjusting either partnership change rate (i.e. partnership formation and dissolution rate) or transmission rate fails this requirement for some of our models, we scaled both partnership change rate and transmission rate by the same factor γ: βadj = γ βbase, cadj = γ cbase, ρadj = γ ρbase. Since transmission rate is scaled by γ, uncoupled and extra-couple transmission rates are adjusted as well. For the instantaneous-switching and implicit models, none of which track single individuals, only the transmission rate and partnership dissolution rate (in this case equivalent to the partnership change rate) are adjusted.
We run each model for each of 1000 parameter sets chosen by Latin hypercube sampling, with fixed starting conditions of mean log10 SPVL of 3.0, standard deviation of log10 SPVL of 0.2, and epidemic size of 10−4. After each run, the initial exponential growth rate is calculated. Then the parameters are scaled as described above so that the initial exponential growth rate is scaled to 0.04 year−1, a value that approximates the growth rates of Shirreff et al.’s original models. When calibrating, we run each model for only 500 years (full simulations are run for 4000 years), which is always long enough to capture the exponential growth phase of the model. We use a 4/5 order Runge-Kutta method (ode45 from the deSolve package [24]) for all simulations. (For the heterogeneous model, approximately 10% of the samples failed due to numerical instability; we discarded these runs.)
For each model we derive the following summary statistics: maximum population mean log10 SPVL; time at which this maximum occurs (corresponding to peak virulence—this is also the time at which the maximum rate of progression and maximum transmission rate occur); equilibrium log10 SPVL; and minimum expected time to progression. Minimum expected progression time is obtained by applying the Hill function (eq 2) to the maximum mean log10 SPVL of each run. Equilibrium log10 SPVL is calculated after 4000 years of simulated time. Although most simulations reach equilibrium much earlier than 4000 years, we set this very long time horizon because a small subset of the simulation runs show very slow evolution rates.
Knowing the peak log10 SPVL, timing of the peak log10 SPVL/peak virulence, and equilibrium log10 SPVL provides sufficient detail to identify the overall shape of the virulence trajectory. In particular, knowing the timing of the peak virulence (how many years into the epidemic the virulence peaks) can help epidemiologists guess whether the virulence of an emerging pathogen is likely (1) to peak early, possibly even before the pathogen is detected spreading in the population, and decline over the remaining course of the epidemic; (2) to increase, peak, and decline over the foreseeable future; or (3) to increase very slowly, peaking only in the far future. To the extent that our simplistic model for HIV reflects reality, we would take the peak time of 150-300 years (Fig 1c) to mean that, in the absence of treatment, the epidemic would probably still be increasing in virulence.
Our simplifications of Shirreff et al.’s model [9] reproduce its qualitative behaviour—in particular, its predictions of virulence dynamics—reasonably well. As we calibrate the parameters to achieve initial epidemic growth rates r ranging from 0.042 year−1 to 0.084 year−1 (the former value matching the initial rate of increase in prevalence in Shirreff et al.’s full model) the initial trajectory of increasing virulence brackets the rate from the original model (Fig 1a). For matching initial growth rates (r = 0.042) the peak log10 SPVL occurs at the same time (≈ 200 years) but the peak virulence is lower than Shirreff’s (≈ 4.3 vs. ≈ 4.6 log10 SPVL), as is equilibrium virulence (≈ 4.25 vs. ≈ 4.5 log10 SPVL).
Changing the initial infectious density (I(0)) has little effect on the virulence trajectory. Decreasing I(0) makes SPVL peak slightly later and higher, because it allows a longer exponential-growth phase before the transition to endemic dynamics (Fig 1b). Decreasing the initial SPVL also leads to progressively later, higher peaks in SPVL (Fig 1c). In this case the delay in the peak is more pronounced than for low I(0), because the rate of SPVL increase is eventually limited by the mutation rate. The peak SPVL is actually larger for a lower starting SPVL, presumably because lower SPVL also allows for a longer epidemic phase. However, the peaks are similar across the entire range of initial conditions, because even in the most limited (high-I(0), high-α(0)) cases HIV can evolve close to its optimal growth-phase SPVL.
Across the entire range of parameters covered by the Latin hypercube samples, all of our models produce qualitatively similar virulence trajectories, which we quantify in terms of population mean log10 SPVL (Fig 3: higher population mean log10 SPVL corresponds to higher virulence). Although the speed of virulence evolution varies, leading to wide variation in the peak log10 SPVL (ranging from 3 to 5.5) because HIV can evolve farther toward its growth-phase optimum before the transition to the endemic phase, virulence peaks between 200 and 300 years in all models.
Our chosen summary statistics (peak time, maximum mean log10 SPVL, equilibrium mean log10 SPVL, and minimum mean progression time) all vary considerably across models (Fig 4). We first consider the models of intermediate realism: implicit, instantaneous-switching with and without extra-pair contact, and pair formation without extra-pair contact. Some parameter sets for these models lead to low equilibrium virulence (2.3-3 log10 SPVL). For these data sets, virulence may either increase from its initial value, reaching an early peak (≈ 200 years) between 3 and 4 log10 SPVL and then declining to a lower equilibrium value, or in extreme cases virulence may decline immediately, leading to a peak virulence (as we have defined it) equal to the starting value of α(0) = 3 log10 SPVL at t = 0 (Fig 5). At the opposite extreme, parameter sets that produce high equilibrium virulence (4.7 log10 SPVL) also produce late peaks (> 200 years) and high peak virulence (5.6 log10 SPVL).
The most striking aspect of the univariate comparisons in Fig 4 (and the bivariate comparisons in Fig 5), is the similarity between the results of the least (random-mixing) and the most complex (pair formation with extra-pair contact and pairform+epc with heterogeneity) models. The random-mixing model has the lowest variability, because it is unaffected by uncertainty in pair formation and extra-pair contact parameters, but otherwise the virulence dynamics of these three extreme models are remarkably similar. This phenomenon is driven by the strong effects of extra-pair contact in the model with explicit pair formation and extra-pair contact (“pairform+epc” in Figs 3–6). When individuals spend time uncoupled between partnerships, and when these single individuals can transmit disease to coupled individuals, the resulting unstructured mixing overwhelms the effect of structured mixing within partnerships, leading to mixing that is effectively close to random. Once unstructured mixing is strong, adding realistic heterogeneity of mixing to the model has little effect other than increasing the variability in the outcomes.
The random-mixing, pairform+epc, and heterogeneous models all predict high population mean log10 SPVL at the virulence peak (median (95% CI) = 4.7 (4.65-4.79), 4.72 (4.37-4.96), 4.72 (4.09-5.03)). In contrast, the implicit model predicts a much lower peak log10 SPVL value: 3.52 (3-4.02) years. The random-mixing, pairform+epc, and heterogeneous models predict rapid progression to AIDS at the virulence peak (median/95% CI = 6.1 (5.7-6.3), 6.02 (5.04-7.7), 6.03 (4.8-9.2)), while the implicit model predicts minimum progression times about twice as long (12.5 (9.6-15.6) years). The corresponding differences in mean within-couple transmission probability at the peak are even more extreme, about a fourfold difference: 0.249 (0.24-0.26), 0.252 (0.19-0.28), and 0.252 (0.15-0.28) per year for the random and pairform+epc models vs. 0.059 (0.02-0.13) per year for the implicit model. (S2 Appendix presents plots showing univariate summaries of expected progression time to AIDS and transmission probability.)
Bivariate relationships (Fig 5) help distinguish the results of different models with similar univariate distributions of dynamical summaries. While the relationship between equilibrium log10 SPVL and peak time is similar for all model structures (top left panel), the other relationships show more variation. In particular, the implicit and pair-formation (without extra-pair contact) models are very similar to each other, but distinct from the other models. We still do not have a convincing explanation for this distinction; we would have expected the implicit model to be most similar to the the instantaneous-switching model without extra-pair contact, which most closely matches its underlying assumptions. However, we note that the implicit model derivation is based on defining the force of infection to match a scaled version of R 0, and as such would be expected to match the equilibrium behaviour but not necessarily the epidemic-phase behaviour of a model with explicit partnership dynamics.
Finally, the sensitivity plot (Fig 6) shows the effects of each parameter on the summary statistics. The most notable difference can be observed by comparing the scaled parameters (e.g. βP, βD, c, ρ) with the unscaled parameters (e.g. DP, DD, ce/cw, cu/cw, κ, μ); the effects of βD and DD are not shown in Fig 6 as they show patterns almost identical to βP and DP, respectively.
For the scaled parameters, the parameter ranges (horizontal axis) are compressed for models without extra-pair contact because these models require a large amount of parameter scaling in order to achieve the specified initial epidemic growth rate (r = 0.04). In contrast, models with extra-pair contact show a wide range of parameters as they can display a wide range of dynamics depending on ce/cw (as well as cu/cw for models with uncoupled mixing) and thus require a wide range of scaling factors to achieve the target growth rate. Parameter ranges for the random-mixing model (especially c) are severely compressed because this model has little flexibility.
For parameters involved in partnership turnover (c and ρ), the figure again shows differences between models with and without extra-pair contact. Models with extra-pair contact show a gradual decrease in peak time, maximum log10 SPVL, and equilibrium log10 SPVL with increasing turnover rates. Increases in the other parameters lead to increases in all three summary statistics. In these models, increased turnover rates diminish the effect of extra-pair contact, thus selecting for lower log10 SPVL. For models without extra-pair contact, increased turnover rates decrease the level of structured mixing (mimicking extra-pair contact models), resulting in selection for higher log10 SPVL.
The implicit model and the instantaneous partnership formation model show similar patterns in scaled parameters. In fact, the effect of partnership dissolution rate, c, on equilibrium log10 SPVL is almost identical in these models (although they can be distinguished in Fig 5). Lastly, increasing in transmission rates (βP and βD) causes the summary statistics to decrease in all models except the random-mixing model.
Surprisingly, once calibration is taken into account, the unscaled parameters have little effect overall. Increase in duration (DP, DD) in the primary and disease stages generally decreases the equilibrium virulence, peak virulence, and peak time, although the models with uncoupled mixing and random-mixing model have high, relatively constant values with respect to these parameters. The ratio of extra-pair to within-pair contact (ce/cw) affects summary statistics in the instantaneous-switching +epc model, but not the pair-formation+epc model (probably because the uncoupled individuals present in the pair-formation+epc model make extra-pair contact by coupled individuals less important). Similarly, increasing the ratio of uncoupled to within-pair contact, cu/cw, increases peak and equilibrium log10 SPVL and delays peak time of the pair-formation+epc model but has almost no effect on the heterogeneous model. Neither the uncoupled contact rate nor the mean (μ) or CV2 of the number of non-cohabiting sexual partners has much systematic effect in the heterogeneous model.
Finally, incorporating additional realism to the model, i.e. combining heterogeneity with all four basic contact structures or allowing for vital dynamics rather than assuming an SIS model, leads to only small differences in the conclusions stated so far (Fig D in S2 Appendix). Relative to our baseline SIS assumption, the effect of adding vital dynamics is to delay the virulence peak slightly and increase both the peak and equilibrium virulence. The changes are small, however: across all models, the maximum increase in time until the virulence peak is 40 years (for the instswitch+epc model), in the peak log10 SPVL is 0.24 units (instswitch), and in the equilibrium log10 SPVL is 0.4 units (pairform). The changes in the most realistic model (pairform+epc) are considerably smaller: an increase of 2 years vs. a decrease of 13 years in the time to the virulence peak for the models with vital dynamics and heterogeneity, respectively; an increase of 0.1 units vs. a decrease of 0.01 units in peak log10 SPVL; and an increase of 0.145 units vs. a decrease of 0.03 units in equilibrium log10 SPVL. Thus, while we can never rule out the possibility of some higher-order interaction among epidemiological phenomena leading to significant changes in our conclusions, we are reasonably confident that the results reported here are robust to additional complexities.
How contact structures are modeled can strongly affect researchers’ conclusions about the evolutionary dynamics of virulence. In particular, a relatively simple, strategic eco-evolutionary model of HIV can predict peak log10 set-point viral loads (over the course of an epidemic) ranging from 3.5 to 4.8 depending on the specific model of sexual partnership behaviour used. This difference in log10 SPVL is epidemiologically significant, corresponding to a twofold difference (12 vs. 6 years) in expected time to progression.
The restriction of transmission within stable partnerships strongly limits eco-evolutionary dynamics by limiting the maximum speed of epidemic growth. An HIV genotype that optimizes SPVL to maximize the speed of spread in a homogeneous population will be sub-optimal in a context where infection can only spread beyond a partnership once it dissolves. This finding echoes a long line of studies that show that population structure leads to the evolution of “prudent” parasites, although most of these studies focus on equilibrium optima rather than eco-evolutionary dynamics [25–28]. The more complex contact structures we modeled mitigate these constraints by allowing HIV to spread among uncoupled individuals (through finite pair-formation) and members of stable partnerships (through extra-pair contact), albeit at lower rates than within partnerships.
Thus, we see the biggest differences not between the simplest and the most complex contact structures, which either ignore pair structure completely or allow for extra-pair contact that reduces its impact, but between the complex contact structures and models of intermediate complexity. These intermediate-complexity models attempt, quite reasonably, to add at least some of the realism of human sexual behaviour, but err by neglecting the apparently insignificant detail of extra-pair contact. If partial complexity may lead to such mistakes, how can modelers do anything but always strive to build the most realistic models possible?
All models must simplify the world. Many constraints—among them data availability, computation time, and code complexity—drive the need for parsimony, with different constraints applying in different contexts. The critical question that modelers must ask is whether the simplified model gives adequate answers, or whether the simplifications lead to qualitative or quantitative errors. This question is especially important for modelers who are hoping that their conclusions will guide management decisions.
In the particular example of HIV virulence eco-evolutionary dynamics and the complexity of contact structures we reach the slightly ironic conclusion that the effort put into building a more realistic model essentially cancels out, putting us back where we started when used a naive random-mixing contact model. However, we are not quite back where we started, as the complex models lead to wider, presumably more realistic confidence intervals on the predictions. In general, unstructured mixing—whether occurring through purely random mixing, or through extra-pair contact and contact among people outside of stable partnerships—tends to drive faster virulence evolution, leading to higher peak virulence and lower times to progression at the peak time.
Taking further steps to make the model even more realistic might add further structure, making the random-mixing model predictions less accurate. For example, our model forms partnerships randomly, and assumes that extra-pair contact is randomly mixing across the population; one could instead model extra-pair contact as arising from multiple concurrent partnerships (some, such as contact with sex workers, of very short duration) and/or more structured partnership formation (by age, ethnicity, or behaviour group). In contrast, the elevated viral load in the early stage of HIV infection, neglected in our model, will likely lead to higher maximum epidemic growth rates and allow more scope for transient viral evolution, although only if extra-pair contact is possible. The effects of other realistic complications such as explicit modeling of two sexes (both in contact structure and differential transmission probabilities), temporal and spatial variation in epidemic processes, or presence of genetic variation in hosts are harder to predict. As mentioned above, our compartmental model already requires tens of thousands of coupled differential equations, which will increase multiplicatively with additional model dimensions such as age, sex, or HIV stage. Thus, further model elaboration will best be done with agent-based models.
Parameterization is one of the biggest challenges of epidemiological modeling. In addition to following Champredon et al. [15] by doing Latin hypercube sampling across a wide range of epidemiological parameters, we calibrated each set of parameters to the same initial epidemic growth rate, chosen to match the results of previous models [9]. Previous models in this area have drawn their parameters from cohort studies from the 1990s [16, 29] rather than doing any explicit calibration to epidemic curves, but they give reasonable order-of-magnitude growth rates (≈ 0.04 year−1) for the early stages of the HIV epidemic (although considerably lower than estimates of ≈ 0.07 − 0.1 year−1 based on population genetic reconstructions [30]). However, our reason for calibrating was not to match any specific observed epidemic, but rather to make sure that we were making meaningful comparisons across a range of models with radically different contact structures, and hence involving different interpretations of the same quantitative parameters. For example, in models with instantaneous switching the partnership dissolution rate c is identical to the partnership formation rate; in models with explicit partnership formation, the partnership formation rate is also c at equilibrium, but might vary over the course of an epidemic. Models with equal parameters but different structures cannot be compared directly; calibration solves this problem.
More generally, any model that wants to be taken seriously for management and forecasting purposes should be calibrated to all available data, using informative priors to incorporate both realistic distributions of uncertainty for all parameters from independent measurements [31] and calibration from population-level observations of epidemic trajectories. Such a procedure would also be an improvement on the common—although not universal—practice, which we have followed here, of assessing uncertainty over uniform ranges rather than using distributions that allow more continuous variation in support over the range of a parameter.
Researchers have documented that HIV virulence and set-point viral load are changing, on time scales comparable to those portrayed here (e.g., compare Fig 3 to Herbeck et al.’s estimated rate of change of 1.3 log10 SPVL per century [95% CI -0.1 to 3] [32]), and have begun to build relatively realistic models that attempt to describe how interventions such as mass antiretroviral therapy (ART) can be expected to change the trajectory of virulence evolution [11, 33, 34]. While these efforts are well-intentioned, we caution that structural details that are currently omitted from these models could significantly change their conclusions.
|
10.1371/journal.pgen.1000250 | The Impact of the Nucleosome Code on Protein-Coding Sequence Evolution in Yeast | Coding sequence evolution was once thought to be the result of selection on optimal protein function alone. Selection can, however, also act at the RNA level, for example, to facilitate rapid translation or ensure correct splicing. Here, we ask whether the way DNA works also imposes constraints on coding sequence evolution. We identify nucleosome positioning as a likely candidate to set up such a DNA-level selective regime and use high-resolution microarray data in yeast to compare the evolution of coding sequence bound to or free from nucleosomes. Controlling for gene expression and intra-gene location, we find a nucleosome-free “linker” sequence to evolve on average 5–6% slower at synonymous sites. A reduced rate of evolution in linker is especially evident at the 5′ end of genes, where the effect extends to non-synonymous substitution rates. This is consistent with regular nucleosome architecture in this region being important in the context of gene expression control. As predicted, codons likely to generate a sequence unfavourable to nucleosome formation are enriched in linker sequence. Amino acid content is likewise skewed as a function of nucleosome occupancy. We conclude that selection operating on DNA to maintain correct positioning of nucleosomes impacts codon choice, amino acid choice, and synonymous and non-synonymous rates of evolution in coding sequence. The results support the exclusion model for nucleosome positioning and provide an alternative interpretation for runs of rare codons. As the intimate association of histones and DNA is a universal characteristic of genic sequence in eukaryotes, selection on coding sequence composition imposed by nucleosome positioning should be phylogenetically widespread.
| Why do some parts of genes evolve slower than others? How can we account for the amino acid make-up of different parts of a protein? Answers to these questions are usually framed by reference to what the protein does and how it does it. This framework is, however, naïve. We now know that selection can act also on mRNA, for example, to ensure introns are removed properly. Here, we provide the first evidence that the way DNA works also affects gene and protein evolution. In living cells, most DNA wraps around histone protein structures to form nucleosomes, the basic building blocks of chromatin. Protein-coding sequence is no exception. Looking at genes in baker's yeast, we find that sequence between nucleosomes, linker sequence, is slow evolving. Both mutations that change the gene but not the protein and those that change gene and protein are affected. We argue that selection for correct nucleosome positioning, rather than differences in mutational processes, can explain this observation. Linker also exhibits distinct patterns of codon and amino acid usage, which reflect that DNA of linker needs to be rigid to prevent nucleosome formation. These results show that the way DNA works impacts on how genes evolve.
| In simple models of molecular evolution, selection on protein coding sequence (CDS) is exclusively devoted to optimizating protein function. As such, we expect amino acid choice to be dictated by protein function alone and synonymous mutations to be neutrally evolving. This is now known to be naïve. The protein's mRNA template can be under selection to maintain favourable mRNA structure [1]–[5] or facilitate speedy and accurate translation through usage of certain synonymous codons [6]–[10]. There is also evidence for selection on regulatory motifs in exons required for correct splicing [11]–[14]. Thus, many stages of the protein production chain are subject to their own particular regimes of selective constraint. But is this also the case when protein-coding information is still stored as DNA in its chromosomal context? In other words, does the way DNA is organized come with its own important requirements on sequence composition, requirements that potentially conflict with optimization of protein function or translation rate optimization or any of the other forces?
One candidate process that might set up selective constraint at the DNA level is nucleosome positioning. Nucleosomes are the elementary units of chromatin organization, at their core comprising a ∼147 bp stretch of DNA tightly wrapped around a histone protein octamer. These core parcels are separated along the chromosome by “linker” regions of variable length [15]. At least two aspects of nucleosome architecture combine to make effects on coding sequence evolution a distinct possibility. First, the histone core has characteristic DNA-binding preferences [16]–[18], governed by the variable bending and twisting attributes of different sequences [19]. Although nucleosomes can form on any stretch of DNA [15], relative affinities can differ by several orders of magnitude [20]. In consequence, nucleosome positioning partly reflects the equilibrium state expected under a model in which energy penalties for coercing rigid DNA into a nucleosome state are minimized [21]. For example, nucleosome-free regions are enriched in rigid poly-A and poly-T runs [22],[23]. Second, selection is likely to favour nucleosomes to be present at particular intra-genic sites and not at others. In particular, well-positioned nucleosomes frequently flank transcriptional start sites thus determining promoter accessibility [23]–[26]. Given that nucleosome formation preferentially occurs on particular sequences, but positioning cannot be entirely opportunistic because it is oriented relative to functional motifs, we might expect coding sequence composition to be biased and its evolution to be constrained to maintain adequate nucleosome architecture.
To examine this expectation we make use of a recent high-resolution (4 bp) genome-wide nucleosome map for Saccharomyces cerevisiae [23]. Based on evidence from codon and amino acid usage as well as comparative rates of evolution we identify nucleosome positioning as a novel layer of selection acting on protein-coding DNA.
Based on the experimentally determined S. cerevisiae nucleosome map of Lee and colleagues [23], we assigned a likely occupancy state (OS) to each coding nucleotide. OSs comprise putatively unoccupied linker region, fuzzily positioned nucleosomes, and well-positioned nucleosomes (see Methods). For intra-specific comparison of compositional differences, genes were then “abridged” so that they only contained codons that were predicted to have the same OS (see Methods). Assuming that occupancy is relatively static over the evolutionary time scale analyzed here, we can also study differences in sequence evolution as a function of OS. S. cerevisiae codons from abridged genes that could be assigned to an orthologous codon in S. mikatae were retained for inter-specific comparison. Results of all orthology-based analyses are largely insensitive to choice of close comparator species, with S. bayanus or S. paradoxus orthologues showing the same trends (data not shown).
Analyzing evolutionary rates solely as a function of nucleosome occupancy is likely to yield misleading results because covariates common to both nucleosome architecture and sequence evolution are not controlled for. Prominently, selection on translational accuracy, speed, and robustness requires attention. Translational selection has been put forward as the single most important cause of between-gene variation in evolutionary rates in yeast [27], where highly expressed genes show reduced rates of non-synonymous [28] and synonymous [27] substitutions as well as substantial codon bias [29]. More acutely, expression intensity is linked to promoter-type [30], which in turn is linked to where, and how, nucleosomes are positioned. Nucleosomes tend to be depleted from promoters [24],[25],[31] but enriched over the coding regions [23] of highly expressed genes. In fact, Shivaswamy and colleagues [26] recently demonstrated that poorly positioned, i.e. fuzzy, nucleosomes over the CDS are associated with high transcription rates.
Considering genes (N = 1718) for which information is available on evolutionary rates, nucleosome occupancy and protein abundance [32], we confirm proportional OS composition as a quantitative marker of expression (Kendall's tau (%linker∼abundance) = −0.24, P≪0.0001; tau (%fuzzy∼abundance) = 0.11, P<0.0001; tau (%wp∼abundance) = −0.07, P<0.0001). Protein abundance is the expectedly strong negative predictor of evolutionary rates (Spearman's rho (abundance∼Ka) = −0.47, P<0.0001; rho (abundance∼Ks) = −0.38, P<0.0001) linking OS composition to Ks (rho (%fuzzy∼Ks) = −0.06, P<0.0001) and, more pertinently, Ka (rho (%fuzzy∼Ka) = −0.1, P<0.0001). Consequently, controlling for expression in analyzing the impact of nucleosome occupancy is imperative.
The ideal approach to eliminate differences in expression between genes is to compare OS-linked evolution within genes. Within-gene analysis suggests that linker sequence exhibits reduced synonymous and non-synonymous evolution (ΔKa(well-positioned v linker): 15%, paired t-test: 4.37, P<0.0001; ΔKa(fuzzy v linker): 7%, paired t-test: 1.61, P<0.11; ΔKs (well-positioned-linker): 10%, paired t-test: 4.64, P<0.0001; ΔKs (fuzzy-linker): 12%, paired t-test: 5.47, P<0.0001; N = 158; see Methods). These results offer preliminary support for the hypothesis that linker sequence is under stronger purifying selection than non-linker sequence at both synonymous and non-synonymous sites.
However, within-gene comparisons can only be carried out for a small number of genes (N = 158) because rarely is there sufficient sequence for all OSs within the same gene to obtain reliable rate estimates. Consequently, this sample is biased towards very long genes (see Methods). Further, within-gene comparisons might still not reflect the true relationship between nucleosome occupancy and sequence evolution if there is intra-genic heterogeneity in substitution dynamics. This is because nucleosomes exhibit promoter-specific architectures, in line with their role in regulating promoter accessibility [23],[25]. As the majority of translational start sites (ATG) in yeast are positioned within one nucleosomal rotation of the transcriptional start site [33], 5′ ends of CDSs show regular occupancy patterns (Figure 1A), which have repeatedly been described in the literature. This intimate association of CDS region and OS only gradually collapses downstream because linker length variation is typically modest [23]. Furthermore, regularities can also be detected across 3′ ends of CDS [26] (Figure 1A). If, then, there existed gene-region distinct evolutionary trajectories, we would expect any analysis of OS-based differences to be biased as a result of the uneven representation of OSs across these regions (Figure 1A bottom panel).
To address the issue of regional biases and increase the amount of available sequence, we chose a concatenation-based approach. Eligible codons were concatenated across all genes ≥906 nt (N = 845) by region (5′, core, 3′) and OS. The terminal 100 codons were taken to represent 5′ and 3′ regions. For the core region, we analyzed the central 100 codons (“restricted core”) as well as all sequence after the termini are removed (see Methods and Table S1). As depicted in Figures 1B&C, there is indeed a marked regional component to coding sequence evolution, with Ks reduced at the CDS periphery and Ka at the centre of genes. That reduced synonymous substitutions at CDS termini can combine with low amino acid substitutions towards the centre of the gene has been observed previously in bacteria [34]. Selection on translational control mechanisms [35]–[37] and Hill-Robertson effects [38] might be the cause of regionally distinct Ks while the explanation for intra-genic variation in Ka is more elusive. Whatever the cause, the result is a spatial bias likely to confound analyses of nucleosome-related sequence evolution by inflating existing trends. In particular, linker sequence evolves particularly slowly at 5′ ends, where it is most prevalent (Figure 1A bottom panel). Importantly, however, OS-linked differences are still manifest within regions (Figure 1B&C, Table S1). Thus, regional biases are insufficient to explain why sequences show distinct evolutionary patterns depending on OS.
From the described results, a contradictory finding emerges. When comparing evolutionary rates within genes, we found Ka and Ks both reduced in linker sequence, yet in the regional analysis Ka and Ks, oddly, disagree. Ka appears reduced for fuzzy sequence (Figure 1C). This discrepancy, however, might be an artefact of fuzzy sequence being enriched in highly expressed genes, which in turn show elevated levels of amino acid conservation [28].
To evaluate this possibility, sequence concatenated by region and OS was further binned by protein abundance (see Methods). Although noise is substantial, Figures 2A&B illustrate for 5′ regions that controlling for expression recreates a more consistent picture of substitution dynamics. Synonymous but also non-synonymous substitution rates are reduced in linker regions (Table 1, Methods) by ∼6% (Table S2). Ks but not Ka is also reduced in core regions (by ∼5%) while we detect no significant differences in substitution rates between OSs across 3′ regions (Table 1). Evolutionary rates of sequence associated with fuzzily or well-positioned nucleosomes are virtually indistinguishable (Table S2). Thus, the reduced Ka for fuzzy sequence observed in Figure 1C is an artifact of the enrichment of fuzzy sequence in highly expressed genes.
Patterns of single nucleotide polymorphisms (SNPs) suggests that whichever factors have caused OS-linked differences in divergence are still a relevant evolutionary force in current populations of S. cerevisiae. Analyzing polymorphism data from a recent re-sequencing effort of over 30 S. cerevisiae strains (see Methods), we found SNP density in the same set of genes to be reduced relative to random expectation at synonymous (chi-square test = 35.61, P = 1.8E-08, enrichment: linker: 0.89, fuzzy: 1.00, well-positioned: 1.02) and non-synonymous sites (chi-square test = 11.48, P = 0.0032, enrichment: linker: 0.95, fuzzy: 1.04, well-positioned: 0.98). These trends become even more clear-cut when expression is controlled for (data not shown).
Although the above results support the notion that purifying selection is stronger in linker than in non-linker, this need not be the correct interpretation. Linker sequence might simply be less mutable. This could be for one of two reasons. First, codons enriched in linker are less mutagenic. Second, regardless of codon composition, linker is somehow protected from mutation.
As regards the first possibility, codons preferentially employed in linker sequence are noticeably AT-rich (see below). As G and C are typically considered more mutable, this alone may explain low evolutionary rates in linker. We control for this scenario in the following way: for every aligned S. cerevisiae linker codon, we randomly select (without replacement) an identical S. cerevisiae codon from the pool of identical codons in the fuzzy and well-positioned concatenated sequences in the same expression/region bin respectively. In the small number of cases where a linker codon could not be matched to a codon in a different OS, a codon was chosen at random. In this way, we end up with sequences of the same length as the linker sequence and virtually identical codon composition. Table 1 reveals that, controlling for codon composition, we find the same pattern of constraints uncovered previously (also see Table S2). We conclude that the low rates of evolution observed for linker sequence are not more parsimoniously explained by an AT-mutation bias.
Could it be that linker sequence is less mutagenic, regardless of codon content? One can imagine mechanistic models in which this might be possible. For example, Kepper et al. [39] recently explored the links between chromatin fiber conformation and nucleosome geometry. Their models, based on mammalian chromatin, suggest that during higher-order organization of nucleosomes into compact chromatin fibers linker sequence is brought into the core of the chromatin fiber upon binding of linker histone, and might be better protected against mutagens as a result. It has also been shown that the binding of linker histone Hho1p inhibits homologous recombination [40]. As homologous recombination in yeast is thought to be mutagenic [41]–[43], reduced rates of substitution might be linked to the protective effects of Hho1p binding.
Aside from the fact that it is unclear whether yeast chromatin is organized in a mammal-like fashion as far as higher order structure is concerned, it seems unlikely that mutational effects can be the sole explanation, not least because linker sequence shows different rates of evolution as a function of intra-gene position even when overall regional biases are taken into account. The proportional reduction of linker Ks to synonymous rates of nucleosome-bound sequence in the same bin tends to be significantly higher at 5′ (median reduction = 0.114) versus 3′ ends (median reduction = 0.026, Wilcoxon test P = 0.04), with the difference to core regions not quite significant (median reduction = 0.057, P = 0.07).
If nucleosome positioning is responsible for elevated linker conservation then we might additionally expect to see skews in patterns of codon and amino acid usage. We compared codon and amino acid composition between OSs within the S. cerevisiae genome. As alignability is not an issue in this analysis, we can exploit a substantially larger number of genes ≥906 nt (N = 1986). Figure 3 shows for core sequence binned by protein abundance that multiple amino acids are depleted or enriched in linker sequence relative to their proportional use across all core sequence (regardless of OS).
These skews appear linked to nucleosome occupancy. First, some amino acids are coded exclusively by nucleotide trimers that are unanimously, albeit sometimes weakly, predictive of either nucleosome binding or exclusion as determined by Peckham and colleagues for genomic sequence [44] (Table 2). If nucleosome positioning was a relevant functional concern, such amino acids should be depleted from linker sequence if all their codons have a positive positioning score, and vice versa, because they have no capacity to negotiate this concern by adjusting their codon usage. This is what we observe. Eight out of eleven amino acids with unanimous positioning score across all codons show skewed usage in the expected direction (Table 2, Table S3), while the remaining three show no significant skews.
This rule of thumb can explain the majority of cases where amino acids are depleted from linker regions. Amino acids most strongly enriched in linker (I, L, N, Y), on the other hand, show the strongest and most consistent evidence for biased usage of certain codons (Table 2), and are therefore probably enriched because one or more of their codons is preferentially employed in linker. We tested non-random enrichment/depletion of synonymous codons across OS for each protein abundance bin independently using Fisher's exact test. Of those amino acids (D, F, I, K, L, N, Y) where we find an overall trend for certain codons to be significantly enriched or depleted (Table 2, Table S3, see Methods on how significance was determined), asparagine (N) codons in particular discriminate remarkably well between OSs, with AAT highly enriched in linker sequence (Genomic ratio: AAT/AAC = 1.44, ratio in nucleosome-bound sequence: AAT/AAC = 1.38, ratio in linker: AAT/AAC = 2.5; determined across all bins and regions).
Finally, we compared codon usage in experimentally determined linker sequence with codon usage in sequences selected for maximum nucleosome exclusion potential from simulated sequences (see Methods) and found them to be in good agreement (Figure 4). In particular, all codons consisting entirely of A and T nucleotides are enriched in both simulated and experimentally determined linker sequence. We identify only one codon, GAT, that is not entirely composed of A or T nucleotides. It is interesting to note here that linker elements proximal to nucleosomes can interact with nucleosome remodeling complexes [46],[47] and that Song et al. [48] recently reported recognition motifs of the GATA family of transcription factors to be enriched in nucleosome-free regions at the fission yeast centromere 2, with the binding consensus being centered around the GATA motif.
The above evidence is consistent with stronger purifying selection acting on linker to maintain correct nucleosome positioning. Could it be, however, that purifying selection is operating, just not as regards nucleosome positioning? We consider two alternatives.
First, might linker sequence be enriched for transcriptional control elements? This seems unlikely for several reasons. Whereas in multicellular eukaryotes it is not unusual for transcription control elements to be located within the open reading frame, transcription regulation in yeast is typically governed by upstream regulatory elements alone [49]. For a handful of genes an effect on expression level upon removal/mutation of specific intra-genic elements has been demonstrated experimentally. However, these elements are mostly located in nucleosome-bound regions (Table S4).
A second possibility is that functional mRNA secondary structure, another cause of sequence conservation and biased composition [1],[4],[50],[51], preferentially maps onto linker sequence. Proposing such a small-scale spatial bias is not unreasonable. We know that nucleosomes are regularly positioned around the promoter, which is also the pivot around which secondary structure facilitating translation initiation is organized [52]. As a result, 5′ regions in yeast are enriched for strong local secondary structures vis-à-vis the remainder of the CDS [51].
Might it be that linker regions and functional secondary structure spatially overlap so that the signature of elevated conservation is really owing to selection on mRNA secondary structure? We find no evidence for this. The window within which hairpin structures downstream of the start codon have an effect on translation initiation (+12–+18 nt [37],[53],[54]) typically fall within the CDS region occupied by the well-positioned nucleosome downstream of the promoter rather than linker sequence (cf. Figure 1A). We also examined a set of strong local mRNA secondary structures (Supplementary Table 1 in [51]), but found no preferential mapping onto linker sequence (Table S5).
The aim of the present analysis was to elucidate whether selection at the DNA level, specifically on nucleosome organization, has affected the evolution of protein-coding sequence. Controlling for intra-genic biases in nucleosome occupancy and, critically, gene expression, we find linker sequence to evolve more slowly, particularly 5′ where constraints are evident on both synonymous and non-synonymous evolution. This is consistent with nucleosome architecture in this region being essential to control gene expression. We estimate that linker sequence across yeast genes evolves approximately 6% slower than sequence bound by nucleosomes. As linker accounts for less than 10% of total genic sequence (with a regional maximum of ∼15% across 5′ regions), the overall reduction in Ks is small (<1%). Note, however, that we almost certainly underestimate the effect of nucleosome positioning concerns on coding sequence evolution. This is because our method of detecting selection is based on differences between OSs. In consequence, if nucleosome-bound sequence were also under selection, as suggested by previous research [26],[55], this would lead to an underestimation of the magnitude of selection.
Even assuming that overall effects are modest, however, the results are nonetheless important for several reasons. First, as nucleosome formation on genic sequence is a universal process, our finding of OS-linked evolutionary patterns across regions and expression levels implies that nucleosome positioning, and thus selection at the DNA level, could affect coding sequence evolution in most if not all other eukaryotes. This potentially has direct implications for estimating the neutral mutation rate from genic regions, although as noted above, the effects are probably weak so unlikely to cause serious errors.
Second, while the overall effects on sequence evolution might be minimal vis-à-vis other determinants of substitution rates, synonymous substitutions might individually be of selective significance. The presence of purifying selection certainly argues that individual synonymous mutations have in the past been weeded out because they introduced sequence-based errors in nucleosome positioning. By implication, and given that nucleosomes are a ubiquitous companion of genic sequence, such mutations might be a novel cause of genetic disease.
Third, these results have an important implication for interpreting local patterns of codon usage. Translationally optimal codons are frequently depleted from linker regions (Table 2). As a result, adaptation for translational efficiency is reduced in linker sequence, as evidenced by a reduced frequency of optimal codons (FOP) (Figure 5; paired t-test for extended core regions: ΔFOP(well-positioned-linker) = 11.20, P<2.2E-16; ΔFOP(fuzzy-linker) = 11.73, P<2.2E-16; ΔFOP(well-positioned-fuzzy) = −3.7, P = 3E-04) and longer runs of translationally non-optimal codons are more likely (Table S6). Previously, runs of non-optimal codons have been considered in the context of selection on translation regulation [56]. Such runs may, for example, induce ribosomal stalling as non-optimal codons tend to be specified by rare tRNAs. This in turn may affect protein folding [57]–[59]. Specification of linker sequence provides a viable alternative hypothesis for a subset of these runs (Table S6).
Finally, the results are consistent with the idea that nucleosome positioning in CDS is in no small part determined by linker-based exclusion signals in contrast to specific nucleosome binding signals, an idea that has recently grown in appreciation [23],[44],[60]. While affinity sequences are more common in coding sequence than expected by chance [55], this signature is relatively weak [26]. If positioning of nucleosomes on CDS is principally achieved by exclusion signals, this is what we expect. Positioning by exclusion might be a particularly beneficial modus operandi for coding sequence, as it restricts constraints to a small proportion of an already highly constrained class of sequence.
Note added during production: the observation that linker sequence evolves more slowly has recently been independently made by Washietl et al. [61]
Likely occupancy states (linker, fuzzily, and well-positioned nucleosomes), across the S. cerevisiae genome were downloaded from http://chemogenomics.stanford.edu/supplements/03nuc/ (Table S5). S. cerevisiae chromosomes were obtained in GenBank format from the Saccharomyces Genome Database (SGD) (ftp://genome-ftp.stanford.edu/pub/yeast/data_download/sequence/genomic_sequence/chromosomes/fasta/archive/genbank_format_20060930.tgz; archived versions from 30/09/2006 to match the data of Lee et al. [23]). Gene models were extracted and filtered so that only genes with a multiple of three nucleotides, proper start and termination codon, no internal stops or ambiguous nucleotides (“n”) were retained. Further, all genes containing introns without consensus splice sites (GT-AG) were eliminated. For each nucleotide in each gene, a likely OS was determined by retrieving all tiling probes (from Lee et al. [23]) containing this nucleotide and determining the dominant call. For example, if covered by 3 probes called as linker, linker, and fuzzy nucleosome, we considered the nucleotide to be in the linker region; regions with 2-probe coverage, where probe calls can be in conflict, were excluded from the analysis, as we had no biological reason to attribute codons to either category. These cases are rare (<0.2% of codons) and thus did not warrant inclusion in a separate category. Only genes in which every nucleotide is covered by at least 2 probes were considered.
For the filtered set of S. cerevisiae genes, orthologues of S. mikatae were obtained from SGD (ftp://genome-ftp.stanford.edu/pub/yeast/data_download/sequence/fungal_genomes/S_mikatae/MIT/orf_dna/orf_genomic.fasta.gz). Filters for likely protein-coding capacity were applied as above. The remaining orthologue pairs were aligned at the protein level using MUSCLE (v3.6) after removal of start and stop codons. Alignments with >5% gaps were discarded. Aligned codons for which S. cerevisiae OS was consistent across all three nucleotides were concatenated by OS across relevant gene subsets as stated in the Results. Ka and Ks were calculated using Li's protocol [62].
Analysis of OS-linked differences in sequence evolution were based on a small number of genes (N = 158) with ≥300 coding nucleotides of each major (linker, fuzzy, well-positioned) OS and a sufficient number of degenerate sites to calculate Ks. Relative rate differentials were calculated as (Ks linker−Ks well-pos)/((Ks linker+Ks well-pos)/2). The analysis was repeated excluding genes with Ks or, more likely, Ka = 0. The results remained qualitatively the same (data not shown). Median gene length is markedly longer (median = 2787 nt) than across all yeast genes (median = 1245 nt, Mann-Whitney U test P<2.2E-16), with likely implications for gene function and expression, so that this sample cannot be considered representative.
Genes ≥906 nt without alignment gaps (N = 845, median CDS length = 1473 nt) were considered in the analysis of regional differences. Start and stop codons were trimmed off and terminal (5′ and 3′) and core 100 amino acids concatenated separately. On average, 11010 linker, 54328 fuzzy, and 50780 well-positioned codons were analyzed per region. We chose 100 amino acids as a convenient cut-off as this a) typically captures well-positioned nucleosomes (plus linker) at the start and end of genes (cf. Figure 1A), for which exact positioning is most likely to be of functional significance, and b) analysis of intra-genic substitution variation in prokaryotes [34] suggests that biases extend at least 50 amino acids into the gene. As we do not know what the causes of this variation are or how substantially they affect yeast, a cut-off of 100 amino acids appears a prudent conservative choice. Defining the core as all sequence left after termini have been removed yields qualitatively identical results (data not shown). As the larger amount of sequence available affords a better resolution when the core is defined in this way, we present results for this definition unless otherwise indicated. Ka and Ks were determined for all aligned concatenates. Significance of differences in evolutionary rates across OSs was tested by repeated random sampling of aligned codon pairs from a region-specific super-concatenate containing all OS concatenates to create 3(OS)×3(regions)×10 000 sequences of the same lengths as the original concatenates. Observing Ka (Ks) values for the original concatenate more than two standard deviations below the mean of the distribution of randomized sequences is taken to be indicative of evolutionary constraint. Concomitant Ka (Ks) values significantly faster than expectation are attributed to the fact that OSs are non-independent. This constraint-guided interpretation is justified because positive selection is expected to be much rarer than purifying selection across the large sample of genes considered here.
Coding sequence concatenated by region and OS was split into expression bins based on protein abundance data from Newman and colleagues [32]. Starting with the gene whose protein was least abundant, sequence from individual genes was allocated to bins of increasing protein abundance. A new bin was generated once the previous bin contained at least 400 codons of the rarest OS, linker. Sequence from any one gene was never split between bins. The results are robust for smaller bins (minimum 250 linker codons) but we decided to prioritize reducing sampling noise for Ka(Ks) estimates rather than achieving equal coverage of successive expression ranges. The final bin (highest protein abundance) was discarded because mean average deviation was disproportionally large and the minimum number of codons criterion frequently violated. Differences in evolutionary rates were assessed by analysis of covariance (ANCOVA). OS-specific slopes were shown not to differ significantly, as a prerequisite for assessing the importance of OS as a covariate (Table S2). Average differences in evolutionary rates were quantified by comparing the intercepts of OS-specific slopes (Table S2).
We tested enrichment/depletion of synonymous codons (Table 2) for each protein abundance/region bin independently using Fisher's exact test. At the p<0.05 level we expect N*0.05 bins to show codon skews by chance. With 64 (73, 32) bins in the 5′ (core, 3′) region, we thus expect to see 3.2 (3.65, 1.6) bins with skewed codon usage by chance. Further, there are multiple codons for which significant skews in both directions are observed. This could be owing to both noise in the data and chances of a codon to function as part of linker sequence being dependent on its sequence context. We therefore took a conservative approach to judging whether codon usage is significantly skewed across OS for any one amino acid in that we required A) the difference between numbers of enriched and depleted bins in the core region, for which most data are available, to be 5 or greater and B) the direction of skews not to be inconsistent across regions, e.g. not to find a codon more often enriched than depleted in 5′ regions but more often depleted than enriched in 3′ regions, regardless of whether the relative enrichment in either region was significant on its own.
To evaluate whether codon usage differences across OSs are parsimoniously explained by nucleosome positioning ruled by intrinsic binding affinities, we generated sequences (k = 10 000) of equal length to the region bound by the histone core (147 bp = 49 codons), picking codons at random according to their approximate genomic usage (http://www.kazusa.or.jp/codon/cgi-bin/showcodon.cgi?species=4932). Nucleosome formation potential of these short sequences was scored by assigning a weight to each sequence based on the additive occurrence of all nucleotide k-mers evaluated for their predictiveness in nucleosome positioning by Peckham et al. [44]. Weights corresponded to the receiver operating characteristic (ROC) scores calculated by Peckham et al. [44]. ROC scores reflect the capacity of a k-mer to discriminate between two sets it is differentially represented in, with k-mers of no discriminative power scoring 0.5, a perfect classifier 1.0 (see Peckham et al. [44] and references therein for a more detailed explanation). Overlapping and embedded k-mers were scored as in the following example: 4-mer AAAA was assigned 4× the score for “A”, 3× the score for “AA”, 2× the score for “AAA”, and once the score for the full motif “AAAA”. The overall score was divided by the number of motifs detected. Cross-validation with an alternative algorithm [63] suggests that this approach does, in fact, recover sequences with high and low nucleosome formation potential (Figure S1). Codon usage was compared between the highest and lowest scoring 5% of sequences using a chi-square test. Chi-square cell values were chosen as an approximate measure of codon bias for individual codons (Figure 4).
Codon usage bias towards translationally optimal codons was calculated as the frequency of optimal codons (FOP) [64] using codonw (J.F. Peden) with S. cerevisiae default parameters.
SNP analysis is based on data from the Saccharomyces Genome Resequencing Project available at http://www.sanger.ac.uk/Teams/Team71/durbin/sgrp/index.shtml.
Table S7 contains gene names for all S. cerevisiae genes used for each major analysis, together with identifiers for orthologous S. mikatae ORFs (if applicable). Custom scripts, for example to map nucleosome calls onto coding sequence, are available on request from the authors.
|
10.1371/journal.pntd.0003070 | Bat Rabies in Guatemala | Rabies in bats is considered enzootic throughout the New World, but few comparative data are available for most countries in the region. As part of a larger pathogen detection program, enhanced bat rabies surveillance was conducted in Guatemala, between 2009 and 2011. A total of 672 bats of 31 species were sampled and tested for rabies. The prevalence of rabies virus (RABV) detection among all collected bats was low (0.3%). Viral antigens were detected and infectious virus was isolated from the brains of two common vampire bats (Desmodus rotundus). RABV was also isolated from oral swabs, lungs and kidneys of both bats, whereas viral RNA was detected in all of the tissues examined by hemi-nested RT-PCR except for the liver of one bat. Sequencing of the nucleoprotein gene showed that both viruses were 100% identical, whereas sequencing of the glycoprotein gene revealed one non-synonymous substitution (302T,S). The two vampire bat RABV isolates in this study were phylogenetically related to viruses associated with vampire bats in the eastern states of Mexico and El Salvador. Additionally, 7% of sera collected from 398 bats demonstrated RABV neutralizing antibody. The proportion of seropositive bats varied significantly across trophic guilds, suggestive of complex intraspecific compartmentalization of RABV perpetuation.
| In this study we provide results of the first active and extensive surveillance effort for rabies virus (RABV) circulation among bats in Guatemala. The survey included multiple geographic areas and multiple species of bats, to assess the broader public and veterinary health risks associated with rabies in bats in Guatemala. RABV was isolated from vampire bats (Desmodus rotundus) collected in two different locations in Guatemala. Sequencing of the isolates revealed a closer relationship to Mexican and Central American vampire bat isolates than to South American isolates. The detection of RABV neutralizing antibodies in 11 species, including insectivorous, frugivorous, and sanguivorous bats, demonstrates viral circulation in both hematophagous and non-hematophagous bat species in Guatemala. The presence of bat RABV in rural communities requires new strategies for public health education regarding contact with bats, improved laboratory-based surveillance of animals associated with human exposures, and novel techniques for modern rabies prevention and control. Additionally, healthcare practitioners should emphasize the collection of a detailed medical history, including questions regarding bat exposure, for patients presenting with clinical syndromes compatible with rabies or any clinically diagnosed progressive encephalitis.
| Bats (Order: Chiroptera) have been implicated as hosts and reservoirs for numerous emerging infectious diseases, and are considered one of the most relevant groups of mammals in the study of disease ecology [1]. Guatemala is home to some of the world's richest bat biodiversity, with over 104 extant species [2]–[4]. As one representative global disease detection site established by the U.S. Centers for Disease Control and Prevention, enhanced rabies surveillance and pathogen discovery over the past five years targeting bats has facilitated the discovery of numerous novel viral and bacterial agents [5]–[8]. Considering the diversity and zoonotic potential of pathogens detected to date, the most pressing zoonotic threat from bats in Guatemala is Rabies virus (RABV), the only Lyssavirus documented in the New World [9]. Rabies is defined clinically in humans that present with an acute progressive encephalitis dominated by hyperactivity or paralytic syndromes that eventually deteriorate towards coma and death in nearly 100% of cases [10].
Rabies epizootiology is well appreciated in countries with an established laboratory-based surveillance network. Combined with molecular epidemiology, enhanced and passive surveillance are used to define the geographic distribution of RABV variants, infer the temporal and spatial spread of infections associated with diverse reservoir hosts, identify spillover infections into humans, and to devise relevant prevention and control based upon such information [11]–[13]. Generally, RABV can be divided into two major clades: one comprising variants associated with carnivores around the globe, and another containing variants associated with bats, raccoons and skunks in the New World. In Latin America, RABV is classically recognized as two broadly distinct epizootiological forms or ‘cycles’, one ‘urban’, in which dogs may serve as the primary reservoir host and vector, and the other as a so-called rural or ‘sylvatic’ cycle, involving wildlife [13].
RABV transmitted by vampire bats (Desmodus rotundus) represents the most apparent economic and public health threat associated with bats [14]. Vampire bats exist only in Latin America, and range from the Tropic of Cancer in Mexico, to the Tropic of Capricorn in Argentina and Chile [15]. Vampire bats include three monotypic genera, and all three species derive nutrition from feeding on the blood of other vertebrates. A consequence of this unique biological adaptation is that vampire bats are highly effective at transmitting RABV to a wide diversity of mammals, primarily livestock, but also humans, if preferred prey are not widely available [16]. In Latin America, human rabies fatalities have been associated with RABV spillover from frugivorous and insectivorous bats. However, the burden of mortality appears negligible compared to the human rabies burden associated with the common vampire bat [17]–[19].
Laboratory-based surveillance is recommended by the World Health Organization in all RABV enzootic countries. Accurate ongoing prevalence measures are essential in rabies prevention and control, to estimate the burden of disease, to monitor trends to evaluate the effectiveness of case intervention, and to ensure appropriate management of outbreaks [20]. The surveillance system for rabies in Guatemala is passive, where samples from suspected human and animal cases are collected, sent, and tested by the direct fluorescent antibody (DFA) test at one of two laboratories: the Laboratorio Nacional de Salud in Villa Nueva (main facility), or Laboratorio de Ministerio de Agricultura, Ganadería y Alimentación (MAGA) in Quetzaltenango. The number of reported human rabies cases has decreased over the past two decades. Dogs remain the primary RABV vector, and are associated with approximately 73% (44 of 60) of human cases from 1994 to 2012. However, 23% (14 of 60) of these cases were due to an unspecified or unknown exposure [21], [22]. In the U.S., numerous reports link insectivorous bats to the majority of human rabies cases without a history of conventional exposure to RABV [23]. During the same time period in Guatemala, despite detection of over 200 cases of rabies in livestock, and suspected association of such cases with vampire bat rabies, there were no reported cases in bats.
Why RABV has not been detected in Guatemalan bats is unclear, considering RABV has been detected in the majority of bats species tested throughout North America [24]. Very few bats are tested based upon the existing passive surveillance system, and monoclonal antibody or genetic characterization of RABV variants infecting humans, domestic animals and wildlife is not performed. Given the nocturnal and somewhat cryptic nature of bats, transmission dynamics are difficult to study in natural populations, and many critical gaps remain for a basic understanding the epizootiology of different RABV variants maintained in specific bat taxa. In the present study, we report results from field studies conducted in Guatemala from 2009–2011. The objective of the study was to test whether enhanced surveillance would: (1) complement passive surveillance, specifically regarding the presence of RABV in bats, and (2) extend information on the geographic distribution of RABV circulation among bats to assess the public and veterinary health risks associated with bats in Guatemala.
All animals were captured and handled in accordance with national guidelines (Guide for the Care and Use of Laboratory Animals) [25]. Protocols for animal capture and use were approved by the CDC Animal Care and Use Committee (USA) protocol number 1843 and 2096, and the Animal Care and Use Committee of the Universidad del Valle de Guatemala (Guatemala).
Guatemala was selected as one major comparative New World study location as part of the U.S. Centers for Disease Control and Prevention's (CDC) Global Disease Detection (GDD) program among ten international locations. The objective of the CDC GDD program is to develop and strengthen global capacity to rapidly detect, accurately identify, and promptly contain emerging infectious threats. Nineteen field sites for sampling bats in Guatemala were selected on the basis of historical outbreaks of rabies, contemporary national surveillance data, known or suspected vampire bat depredation upon human populations, or neurological illness reported in livestock. Figure 1a illustrates the geographical distribution of field sites used in the present study. Bats were collected using mist nets set near fruit trees, confined livestock, or near the entrance of caves. Nets were opened between 19:00 to 0:00, and checked every 30 minutes. Bats were removed from mist nets, placed individually in cloth bags, and transported to a nearby temporary field station, where they were sedated by a 0.05 to 0.1 mg intramuscular injection of ketamine hydrochloride, oral/fecal swabs obtained, and terminal blood samples were collected by cardiac puncture under heavy anesthesia. Following euthanasia, bats were identified to the level of species following a key for bats of Costa Rica [26]. Standard morphological measurements were also collected (e.g. gender, age, mass, and forearm length). A complete necropsy was then performed on all bats, and samples were stored immediately on dry ice in the field, and maintained thereafter at −70°C in the laboratory at the Universidad del Valle de Guatemala until shipment to the CDC Rabies Laboratory in Atlanta, GA. Carcasses were banded and fixed in 10% buffered formalin for several days, then permanently transferred to 70% ethanol, for archival purposes.
Brain impressions prepared from frozen bat tissues were fixed in acetone at −20°C, and RABV antigens were detected by the DFA test, using fluorescein isothiocyanate (FITC)-labeled monoclonal antibody (mAb) conjugate (Fujirebio Diagnostics, Inc., Malvern, PA, USA), as described [27].
The presence of RABV neutralizing antibodies (rVNA) was determined by the rapid fluorescent focus inhibition test (RFFIT) or a modified micro-RFFIT test on sera collected from ten field sites [28], [29]. The rVNA titers of individual bats were calculated by the Reed-Muench method, and were converted to international units (IU/mL) by comparison to a standard rabies immune globulin (SRIG) control containing 2 IU/mL [30]. The SRIG titer was generally higher in the micro-RFFIT test compared to the standard RFFIT. For the objective of this study, positive rVNA titers (≥0.06 IU/mL) were defined by at least 50% neutralization of the RABV challenge virus dose (50 focus forming doses) at a 1∶5 dilution (RFFIT) or 1∶10 dilution (micro-RFFIT). Final titers less than 0.06 IU/mL were considered negative for the presence of rVNA for the purposes of this investigation.
Antigenic characterization was performed by indirect immunofluorescence using eight mAbs directed against the RABV nucleoprotein (N) antigens (C1, C4, C9, C10, C12, C15, C18, C19), supplied by EMD Millipore Corporation (Billerica, MA, USA), as previously described [31]. Positive reactivity pattern results were analyzed using previous described antigenic variant patterns [32], [33].
Virus isolation was attempted on oral and fecal swabs stored in growth medium (MEM-10), and organs, obtained from DFA-positive bats. Homogenates (10% w/v) were prepared in MagnaNA lyser tissue homogenizer tubes containing 1.4-mm (diameter) ceramic beads (Roche Applied Science, Penzberg, Germany), using 1.0 mL of MEM-10 as a diluent. The entire solution from the oral swab, and 0.5 ml from the fecal swab was used to inoculate cells for isolation. For virus recovery, 100 uL of test inoculum was added to 1 mL of MEM-10 containing 5×106 mouse neuroblastoma cells (MNA) in a T-25 tissue culture flask (Corning, NY). Tissue culture flasks were incubated at 0.5% CO2 at 37°C for 72 hours. All cultures were sub-passaged a minimum of four times. For infectivity assessments, Teflon-coated four well slides were seeded with 30 uL of MEM-10 containing 0.5×106 cells per mL, and incubated in a humid chamber at 0.5% CO2 at 37°C for 24 hours. The slides were then rinsed with phosphate-buffered saline (PBS 4550), and fixed in cold acetone at −20°C for one hour. RABV antigens were visualized by use of the DFA test, using optimal working dilutions of FITC-labeled anti-RABV mAb conjugate (Fujirebio Diagnostics, Inc., Malvern, PA, USA) after each passage.
Total RNA was extracted from fecal swabs and organ tissues of RABV-infected bats using TRIZol reagent (Invitrogen, Carlsbad, CA, USA). Fecal swabs were stored in 1 ml of MEM-10 in the field, and 200 µl of the swab suspension was mixed with 1 ml of TRIzol for RNA extraction. Primers were selected within the coding region of the N gene and the initial reaction was performed with sense primer 1066F, GARAGAAGATTCTTCAGRGA (positions 1157–1173), which was also used for reverse transcription and antisense primer 304B, TTGACGAAGATCTTGCTCAT (positions 1514–1533). The hemi-nested reaction was performed with sense primer 1087F, GAGAARGAACTTCARGA (positions 1157–1173). The glycoprotein (G) gene was amplified as two overlapping fragments using primer combinations umf2/994b and 760f/308b, as previously described [34]. All positions are given according to the Street Alabama Dufferin RABV strain genome sequence (GenBank accession number M31046). The RT-PCR was performed as described elsewhere [35]. Positive results were confirmed by nucleotide sequencing. RT-PCR products were purified with Wizard PCR Preps DNA Purification System (Promega, Madison, WI, USA), according to the manufacturer's recommendations and sequenced in forward and reverse directions as described by [35], using the Big Dye Terminator Cycle Sequencing Ready Reaction Kit, version 1.1 on an ABI3770 sequencer (Applied Biosystems, Carlsbad, CA, USA).
The complete N and G sequences were assembled and translated to amino acid sequences using the Bio Edit program [36]. The dataset was supplemented with complete and partial gene sequences available from GenBank and aligned in ClustalX [37]. Table S1 describes details of all sequences used in this study, including accession numbers, country of origin, year, and specimen source. The GTR+I+G model were selected based on the Bayesian factor and Akaike criterion evaluated in MEGA, version 5.1. No molecular clock evaluation was implemented. The analysis was performed using the Bayesian skyline population prior, and two independent Markov Chain Monte Carlo (MCMC) runs were performed with 1,000,000 iterations each. The results were combined in Log Combiner, and the resulting maximum clade credibility (MCC) tree generated with Tree Annotator using 20% for burn in and visualized in Fig Tree, version 1.4.0 [38].
All statistical analyses were performed using JMP version 9.0.2 or SAS version 9.2 (SAS Institute Inc., Cary, NC). The 95% confidence intervals (CI) were calculated for proportions of seropositive bats by species and location of capture. A nested mixed logistic model was used to test the effect of trophic guild on seroprevalence. Species nested within trophic guild was treated as a random effect and trophic guild was treated as a fixed effect. As only a single sanguivorous species was tested, we also compared the model outputs with a data set excluding vampire bats. The antibody prevalence between sex were also compared using χ2 tests, and the level of significance was evaluated at α = 0.05.
During 2009–2011, a total of 672 bats of 31 species were collected from Guatemala. Bats were collected during 2009 (n = 220; Table S1), 2010 (n = 135; Table S2), and 2011 (n = 317; Table S3). Table 1 provides the details on the cumulative frequency of capture by species. Among all captures, 56.3% were male. The most frequently captured species was D. rotundus (n = 200; 30%), followed by the Jamaican fruit bat, Artibeus jamaicensis (n = 128; 19%) (Table 1).
From sera available for testing (n = 398), 28 bats demonstrated detectable rVNA for an antibody prevalence of 7% (95% CI 5–10%). Seroprevalence was highest for insectivorous species (21%), followed by omnivorous (8%) and sanguivorous (9%) taxa. The proportion of rVNA seropositive bats varied significantly across trophic guilds in a complete data set (F3,20 = 4.65, p = 0.04; Species (Diet) = 3.4×10−21) and a data set without vampire bats (F2,20 = 6.97, p = 0.005; Species (Diet) = 4.4×10−19). In both models, pairwise contrasts revealed a significant difference between frugivorous and insectivorous bats. Insectivorous bats were 8.5 times as likely to be seropositive compared to frugivorous bats. No other pairwise contrasts in rVNA seroprevalence between trophic guild levels were significant. Among capture locations, the proportion of seropositive bats was highest for Naranjo (21%), followed by El Jobo (19%), and El Penate (11%). Table 2 provides the details for individual bat species sampled by location and trophic guild. Antibody prevalence between sexes was similar (3.8% of males, 3.3% of females; P = 0.77). Species composition of bats captured varied across sites (Table 2).
RABV antigens were detected in the brain from two common vampire bats. The bats were collected in El Pumpo, near the Pacific Coast in the Department of Santa Rosa, and Palo Seco, near the border with Mexico in the Department of San Marcos (Figure 1b). The antigenic reaction pattern derived from the brain specimens were consistent with RABV antigenic variant V3 associated with Desmodus rotundus (Table 3). Upon capture, one of the two rabid bats demonstrated clinical signs consistent with rabies infection. Additionally, this bat was dehydrated, in poor physical condition, and had evidence of several bite wounds to its body.
RABV was isolated in MNA cells 24 h after inoculation from the kidney of one bat, and from oral swabs of both rabid bats. Additional sub-passages revealed the presence of RABV in the spleen of one bat, and the kidney and lung for both bats. RABV was not isolated from the heart, liver, intestine, and fecal swabs. Viral RNA was detected by hemi-nested RT-PCR in all specimens examined, except the liver of Bat 321. The results of RABV isolation and nucleic acid detection from tissues of rabid bats are presented in Table 4.
In total, two RABV G sequences and two RABV N sequences were generated from the brain tissues of infected bats in this study and deposited into GenBank (accession numbers: KF656696-99).
The N gene sequences of both viruses were 100% identical, whereas the G sequences differed in one non-synonymous substitution (302T,S). The viruses were most similar to one of the lineages of vampire bat rabies viruses circulating in Mexico, and were relatively distant from the lineages circulating in South America, with the exception of Columbian lineages. The most phylogenetically related viruses were described previously from Chiapas, Tabasco and Veracruz states of eastern Mexico [39]. The only available vampire bat RABV N gene sequence from El-Salvador (FJ228492) also clustered within this lineage (Figure 2, G gene reconstruction not shown).
Given the critical ecological importance of bats, especially in tropical regions, novel strategies are necessary for the prevention and control of bat-associated zoonoses. To our knowledge, this is the first report and isolation of RABV from bats in Guatemala. In this study, the prevalence of rabies among all collected bats was low (0.3%), even among vampire bats (1%). The two vampire bat RABV isolated were related phylogenetically to viruses associated with vampire bats in the eastern states of Mexico and El Salvador, which is not unexpected given their geographic proximity. Rabies epizootics and phylogenetic clusters of RABV circulating in vampire bat populations are relatively constrained in space [40], [41]. Previous phylogeographic studies suggested that vampire bat rabies likely originated in the territory of Mexico [39]. If true, these viruses may have been introduced to Guatemala and El Salvador via eastern Mexico. The detection of rVNA from bats in this study demonstrated RABV exposure among multiple species of bats in Guatemala. The overall rVNA prevalence of 7% was similar to other bat RABV surveillance studies conducted in Peru (10.3%), Grenada (7.6%) and Trinidad (12.8%), but less than the 37% detection in Colima, Mexico [42]–[44]. Seroprevalence among collection sites ranged from 0 to 21% among all bats, and 0 to 25% for D. rotundus, respectively. Our results are concordant with previous studies in demonstrating that a substantial fraction of apparently healthy bats have detectible rVNA, indicating previous exposure to RABV, and suggesting clearance of peripheral infection without clinical disease [42]–[44]. The presence of rVNA only demonstrates prior exposure to RABV antigens, and does not provide information about the timing, intensity, or frequency of infection [45]. Furthermore, rVNA negative test results do not guarantee a lack of exposure, and may vary according to the selected cutoff value used for the serology test, as reviewed by Gilbert et al. [46]. At a population level, the rVNA seroprevalence data described here provide information about the cumulative exposure history among all bats collected for each field site.
In experimentally infected bats, rVNA are typically generated only among survivors, and bats that succumb to infection may only seroconvert during late stages of clinical disease [45]. Similar observations have been made regarding rVNA seroconversion in human rabies patients [47]. Combined with antigen detection results, rVNA seroprevalence data can be used to elucidate infection dynamics in bat populations. Detection of RABV excretion can provide meaningful insights for identifying potential transmission pathways that can inform the structure of disease dynamic models. In this study, the highest levels of virus were consistently detected from oral swabs, kidney and lung tissue. These results support primary secretion by oral routes, while suggesting a possible pathway in urine (via the kidneys), which warrants further investigation.
Strategies to control RABV transmitted by vampire bats in Latin America have relied historically on population reduction methods, either by non-specific destruction of roosts or by anti-coagulants applied to cattle or individual bats. When the anti-coagulant paste is applied to vampire bats, it is suspected that the treated bats then return to a roost where is spread to other conspecifics through allogrooming. Culling of vampire bats is thought to benefit agriculture and public health in the short-term by alleviating bat bites on livestock and humans respectively. However, there is little evidence that this method actually targets rabid bats, and the apparent positive effect of culling on seroprevalence, combined with demographic and behavioral responses, may actually increase the proportion of susceptible bats [41], [48]. A recent report also suggests that it may be more economically beneficial to support pre-exposure vaccination of cattle, than rely on vampire bat culling [49].
Rabies surveillance at the national level in Guatemala falls under the jurisdiction of the Ministry of Public Health and Social Welfare [50]. Rabies control efforts in Guatemala are focused primarily on mass vaccination of domestic dogs, which has led to a significant decrease in human rabies cases, though it remains unclear whether current efforts can achieve ultimate elimination of canine rabies [21], [51]. Indirect vaccination of vampire bats with recombinant RABV vaccines have proven immunogenic and efficacious in experimental infection models of D. rotundus, though this strategy has not been tested in the field [52]. Despite possibly reducing RABV infection of cattle, this strategy would not eliminate the behavior or health and economic consequences (e.g. secondary bacterial infections) of vampire bat depredation on cattle. Interventions at the human-bat interface should be directed at decreasing the risk of human exposure to bats by improved educational campaigns on the risk of rabies associated with bats and the importance of laboratory testing and prophylactic treatment following exposures, as well as bat-proofing dwellings where feasible [53]. Routine laboratory-based RABV surveillance is necessary to properly evaluate any intervention strategy and should include the submission and testing of all human and animal cases involving bite contact with bats and other reservoir hosts. For Guatemala, overcoming cold chain and transportation barriers should be a priority to increase the proportion of contact cases that are tested. To further elucidate the relative impacts of the urban or sylvatic cycles on public health and agriculture, routine typing of all positive rabies cases should be implemented. Rapid exchange of information between sectors involved in human and animal rabies surveillance and control is essential.
Unfortunately, despite significant progress in the prevention and control of rabies, once clinical signs manifest in humans the case fatality rate approaches 100%, even with intensive supportive care [47]. However, appropriate post-exposure prophylaxis, including immediate washing/flushing and disinfection of the wound and prompt administration of RIG, and modern cell-culture vaccines according to recommended vaccination schedules assures prevention of rabies if bitten by a rabid bat [54]. Fortunately, modern cell culture vaccines are readily available for both humans and animals in Guatemala, and national authorities should provide pre-exposure prophylaxis guidelines for persons or animals with regular exposure to bats or other potential wildlife reservoirs (e.g. carnivores).
As the first focused rabies study in Guatemala, gaps were evident in our study. Limitations to the current study included a lack of more focal seasonal sampling, a broader spatial scale, and a need for greater engagement of the Ministries of Health and Agriculture in a One Health context to improve routine laboratory based surveillance. The presence of bat RABV in rural communities likely requires new strategies for joint public health and veterinary education and outreach, increased availability for diagnostic laboratory testing in remote areas, and continued enhanced surveillance for rabies prevention and control, as suggested for other developing countries [55], [56]. Additional studies at the human-bat interface would be useful to obtain information on demographic characteristics (e.g. age, gender, education) of persons exposed to bats, circumstances of bat exposures (e.g. bites, scratches, skin contact), actions taken following exposure, and knowledge of bat-borne zoonoses. Previously, based upon ignorance or inaction, preventable human rabies cases have occurred from bat exposure [57]–[59]. Given the implications when such recommendations are not operative, localities with a high risk for exposure to bats, and bat-borne zoonoses, should be targeted in laboratory based surveillance activities for the evaluation of robust, long-term prevention and control strategies.
|
10.1371/journal.ppat.1004264 | A Novel Mouse Model of Campylobacter jejuni Gastroenteritis Reveals Key Pro-inflammatory and Tissue Protective Roles for Toll-like Receptor Signaling during Infection | Campylobacter jejuni is a major source of foodborne illness in the developed world, and a common cause of clinical gastroenteritis. Exactly how C. jejuni colonizes its host's intestines and causes disease is poorly understood. Although it causes severe diarrhea and gastroenteritis in humans, C. jejuni typically dwells as a commensal microbe within the intestines of most animals, including birds, where its colonization is asymptomatic. Pretreatment of C57BL/6 mice with the antibiotic vancomycin facilitated intestinal C. jejuni colonization, albeit with minimal pathology. In contrast, vancomycin pretreatment of mice deficient in SIGIRR (Sigirr−/−), a negative regulator of MyD88-dependent signaling led to heavy and widespread C. jejuni colonization, accompanied by severe gastroenteritis involving strongly elevated transcription of Th1/Th17 cytokines. C. jejuni heavily colonized the cecal and colonic crypts of Sigirr−/− mice, adhering to, as well as invading intestinal epithelial cells. This infectivity was dependent on established C. jejuni pathogenicity factors, capsular polysaccharides (kpsM) and motility/flagella (flaA). We also explored the basis for the inflammatory response elicited by C. jejuni in Sigirr−/− mice, focusing on the roles played by Toll-like receptors (TLR) 2 and 4, as these innate receptors were strongly stimulated by C. jejuni. Despite heavy colonization, Tlr4−/−/Sigirr−/− mice were largely unresponsive to infection by C. jejuni, whereas Tlr2−/−/Sigirr−/− mice developed exaggerated inflammation and pathology. This indicates that TLR4 signaling underlies the majority of the enteritis seen in this model, whereas TLR2 signaling had a protective role, acting to promote mucosal integrity. Furthermore, we found that loss of the C. jejuni capsule led to increased TLR4 activation and exaggerated inflammation and gastroenteritis. Together, these results validate the use of Sigirr−/− mice as an exciting and relevant animal model for studying the pathogenesis and innate immune responses to C. jejuni.
| Research into the key virulence strategies of the bacterial pathogen Campylobacter jejuni, as well as the host immune responses that develop against this microbe have, in many ways, been limited by the lack of relevant animal models. Here we describe the use of Sigirr deficient (−/−) mice as a model for C. jejuni pathogenesis. Not only do Sigirr−/− mice develop significant intestinal inflammation in response to colonization by C. jejuni, but the ability of this pathogen to trigger gastroenteritis was dependent on key virulence factors. We also found that the induction of the inflammatory and Th1/Th17 immune responses to infection in these mice depended on specific Toll-like receptors, principally TLR4, which we identified as the main driver of inflammation. In contrast, TLR2 signaling was found to protect mucosal integrity, with Tlr2−/−/Sigirr−/− mice suffering exaggerated mucosal damage and inflammation. Notably, we found that C. jejuni's capsule helped conceal it from the host's immune system as its loss led to significantly increased activation of host TLRs and exaggerated gastroenteritis. Our research shows that the increased sensitivity of Sigirr−/− mice can be used to generate a unique and exciting model that facilitates the study of C. jejuni pathogenesis as well as host immunity to this enteric pathogen.
| Campylobacter jejuni is one of the leading bacterial causes of gastroenteritis in the world. Although responsible for the majority of food-borne bacterial infections in developed countries, and compared to many other common enteric bacterial pathogens, our understanding of the mechanisms underlying C. jejuni's pathogenesis remains poorly defined [1]. One reason for our limited understanding is that C. jejuni appears to utilize unique pathogenic strategies, as it lacks many of the common toxins, effector proteins and virulence factors found in other pathogenic bacteria [1]. For example, cytolethal distending toxin is the only toxin so far identified within Campylobacter strains [1], [2] yet its potentially toxic role in vivo remains unclear [2]. Furthermore, a number of bacterial factors such as capsular polysaccharides [3], [4], lipo-oligosaccharides [4] and proteins such as CadF [5], Peb1 [6], [7], JlpA [8] and the Campylobacter invasive antigens (Cia) [1], [9]–[11], have all been studied in vitro for roles in C. jejuni cell adhesion and invasion, and in the existing commensal colonization models, however whether they play any role in pathogenicity in vivo is largely unknown.
Indeed, the study of C. jejuni pathogenicity has largely been limited by the lack of relevant and convenient animal models that can be used to replicate human disease [12]. While C. jejuni readily colonizes poultry, it does so in a commensal fashion, causing no disease and thus not providing significant insight into C. jejuni pathogenesis or how the host defends against these microbes. Galleria mellonella larvae, which are a common animal model used for the study of several bacterial pathogens have been applied to C. jejuni [13], [14], but their relevance in modeling vertebrate enteric infection is limited. While colostrum-deprived piglets [15], as well as ferrets [16] have been used to model C. jejuni infection with some success, their use is limited by the difficulty obtaining and maintaining these animals, and a lack of immunologic and genetic tools to aid in studying the host response to infection.
Mice would normally provide a preferred infection model system; however, they have repeatedly proven resistant to pathogenic infection by C. jejuni, and many strains are unable to even be reliably colonized [17]. The basis for their resistance to C. jejuni colonization appears to at least partially reflect active competition from the resident intestinal microbiota, thereby preventing C. jejuni from establishing a niche within the murine gut [17]–[19]. Secondly, the murine immune system has proven very tolerant to the presence of C. jejuni and in wild-type (WT) mice, their presence only rarely elicits any overt intestinal inflammation [20], [21]. To overcome this tolerance, several groups have tested genetically manipulated mice that develop exaggerated inflammatory responses to bacteria, such as IL-10-deficient (Il-10−/−) mice [20]. While Il-10−/− mice can be colonized by C. jejuni, resulting in severe enterocolitis, the loss of IL-10 dramatically alters the murine immune system. As a result, their immune system is unable to effectively clear C. jejuni from the GI tract, leading to chronic colonization rather than the acute infections seen in humans. Moreover the immune systems of Il-10−/− mice are so sensitive that the presence of any commensal microbe can potentially trigger spontaneous enterocolitis [21]. The oral gavage and intraperitoneal injections of MyD88-deficient mice, have also been employed for the study of C. jejuni colonization and dissemination in mice [22]–[24], but encounter the reverse limitation of Il-10−/− mice, where the immune response is attenuated, allowing for colonization of the intestine or systemic sites with limited host responses. This provides utility for the study of colonization, but not immunity and the development of inflammation in response to C. jejuni.
Recently we showed that mice deficient in Single IgG IL-1 Related Receptor (SIGIRR) exhibit increased susceptibility to infection by two natural enteric bacterial pathogens of mice, namely Citrobacter rodentium and Salmonella enterica serovar Typhimurium [25]. In both mice and humans, SIGIRR is highly expressed by intestinal epithelial cells and acts as a negative regulator of MyD88-dependent signaling, thus acting to dampen signaling by most Toll-like receptors as well as interleukin (IL)-1R [25]–[27]. In the absence of SIGIRR, when these receptors are activated, their downstream signaling is increased, resulting in increased innate inflammatory responses [26], [27]. In the context of C. rodentium and S. Typhimurium infections, we found that Sigirr−/− mice not only developed exaggerated forms of infectious colitis, but they were also infected much more rapidly and with much lower infectious doses than WT mice. This heightened susceptibility was shown to reflect exaggerated antimicrobial responses by Sigirr−/− mice that were surprisingly ineffective against pathogens, but instead depleted the competing commensal microbes [25], dramatically reducing the microbiota based resistance to intestinal pathogen colonization.
Based on their heightened susceptibility to natural bacterial pathogens of mice, we examined whether Sigirr−/− mice could potentially serve as an infection model for the human pathogen C. jejuni. Although orally delivered C. jejuni were able to sporadically colonize Sigirr−/− mice, antibiotic pretreatment was found to facilitate pathogen colonization, leading to acute gastroenteritis in infected Sigirr−/− mice. We confirmed that C. jejuni primarily activates the innate receptors TLR2 and TLR4 [28]–[31], and found that TLR4 signaling was responsible for most of the inflammatory changes seen during infection. In addition to the requirement for innate signaling, the development of gastroenteritis was also dependent on the activity and pathogenicity of C. jejuni. In infections with C. jejuni mutants deleted for flaA (flagella) [32] or kpsM (capsular polysaccharide) [31], [33], [34], the ability of C. jejuni to cause gastroenteritis was significantly altered. Together, these results validate the use of Sigirr−/− mice as an exciting and relevant animal model for studying innate immune responses to C. jejuni, as well as for the study of pathogenicity factors governing infection by this microbe.
The murine intestine is thought to be highly resistant to oral infection by C. jejuni, based primarily on the ability of the resident gut microbiota to outcompete any incoming C. jejuni [17], [35]. Our experiments support this concept, as we found infrequent and inconsistent C. jejuni colonization of conventionally housed WT C57BL/6 mice following oral inoculation with our wild-type C. jejuni strain 81–176 (107 CFU) (data not shown). To overcome this barrier to colonization, we pretreated WT C57BL/6 mice with vancomycin by oral gavage prior to inoculation with C. jejuni. Previous research by Russell et al. [36] showed that oral vancomycin treatment depleted Bacteroidetes and Clostridia from the intestines of mice while promoting the overgrowth of Lactobacilli [36]. Vancomycin pretreatment has also been shown to promote S. Typhimurium colonization and colitis in a fashion similar to streptomycin pretreatment [37]. Following oral inoculation with C. jejuni, we found the vancomycin pretreated WT mice exhibited consistent and robust pathogen colonization in their ceca (Figure 1a) and colons (Figure S1a). Despite their high levels of colonization, minimal signs of inflammation were observed (Figure 1b and c), consistent with previously published analyses of C. jejuni-colonized immunocompetent mice.
Considering that Sigirr−/− mice exhibit impaired colonization resistance against several murine enteric bacterial pathogens, we tested their susceptibility to C. jejuni infection/colonization. We again saw only sporadic colonization in some mice, but in contrast to WT mice, we also saw occasional signs of intestinal inflammation and other forms of pathology but the results were insufficiently reproducible to provide a reliable model (data not shown). We therefore tested the impact of pretreating Sigirr−/− mice with vancomycin, as previously described for WT mice. We noted that vancomycin induced a similar change in the intestinal microbiota of Sigirr−/− mice as we had found for WT mice (Figure S2). Four hours after vancomycin pretreatment, we orally infected Sigirr−/− mice along with WT mice with approximately 107 CFU of C. jejuni 81–176. We euthanized the mice at 3 and 7 days post-infection, assessing pathogen burden in the cecum (Figure 1a), colon, ileum, mesenteric lymph nodes (MLN), spleen and feces (Figure S1a–e). Both WT and Sigirr−/− mice were quickly colonized, with both strains of mice reaching cecal colonization levels of approximately 109 CFU/g within 3 days. In Sigirr−/− mice, colonization numbers usually peaked within 7–9 days, and began to drop significantly by 2–3 weeks post-infection, with low levels of C. jejuni (<104 CFUs/g) found beyond 3 weeks (Figure S3a). WT mice maintained high and relatively unchanging pathogen burdens for at least 25 days (Figure S3a). Colonization in the colon was similar to the cecum, whereas relatively fewer C. jejuni were recovered from the ileum (Figure S1b). Fecal samples taken just prior to euthanization, and throughout the infection, proved to be largely representative of the colonization of both the cecum and colon, with the numbers more closely resembling the numbers recovered from the colon (Figure S1e and f). C. jejuni were also occasionally recovered in low numbers from the MLN and rarely from the spleen (Figure S1c, d), indicating that even with a high pathogen burden in the gut, C jejuni did not readily go systemic.
Despite carrying similar pathogen burdens, the macroscopic pathology resulting from C. jejuni colonization was dramatically more severe in the Sigirr−/− mice as compared to WT mice. Although neither mouse strain exhibited significant weight loss (>10%) (Figure S3b) or other severe signs of morbidity, the ceca and proximal colons of the Sigirr−/− mice were overtly inflamed and often devoid of stool contents. In mice at the height of infection, the stool itself often became noticeably softer and sticky. Significant enlargement of the mesenteric lymph nodes was also noted in infected Sigirr−/− mice (Figure S4). In comparison, control mice treated with vancomycin, but not receiving C. jejuni, exhibited no significant signs of intestinal pathology 3 or 7 days post antibiotic treatment (Figure 1c and data not shown). As expected, histology revealed few, if any, signs of cecal inflammation in WT mice at 3 DPI (Figure 1b), and only very mild signs of inflammation at 7 DPI, despite their heavy pathogen burden. In contrast, the cecal pathology and inflammation observed in infected Sigirr−/− mice was very severe at both 3 and 7 DPI, including submucosal edema, crypt hyperplasia and widespread immune/inflammatory cell infiltration (Figure 1c). In some cases, the Sigirr−/− mice developed focal cecal ulcers, accompanied by bleeding into the lumen (Figure 1b). The severe damage was focused within the cecum and proximal colon, with only minimal signs of inflammation appearing elsewhere in the intestine.
When we measured gene transcript levels of several key cytokines, we found that despite the lack of overt inflammation in the infected WT mice, they still showed upregulated gene transcript levels for a number of cytokines compared to uninfected mice, most notably TNFα, indicating that the WT mice were not completely unresponsive to the presence of C. jejuni. However these responses were insufficient to trigger overt signs of inflammation (Figure 2). The infected Sigirr−/− mice also showed elevated cytokine gene transcripts at levels significantly higher than those seen in WT mice. Notably, at both 3 and 7 DPI, the Sigirr−/− mice showed significantly higher mRNA levels for IL-17A, TNF-α and Interferon gamma (IFNγ), indicative of a stronger inflammatory response. We also observed higher transcription of the neutrophil chemoattractant KC, as well as the cytokines IL-1β, IL-18 and IL-22, though the variability between mice prevented the demonstration of statistical significance.
To better define the cause of the exaggerated tissue pathology suffered by infected Sigirr−/− mice, we explored the localization of the colonizing C. jejuni and whether it differed with that in WT mice. Staining for C. jejuni in intestinal tissue sections of WT mice at both 3 and 7 DPI revealed the bacteria were largely limited to the intestinal lumen, with rare C. jejuni found in only 26.5% (70/264) of crypts. We also noted C. jejuni accumulating in the mucus layer, whereas relatively few microbes were found in direct contact with the intestinal epithelium or penetrating the cecal or colonic crypts (Figure 3). Conversely, in the Sigirr−/− mice, C. jejuni were not only found within the intestinal lumen and the mucus layer, but large numbers were also seen penetrating deep within cecal and colonic crypts (Figure 3). In these mice, 56.2% (82/146) of crypts were found to be heavily colonized by C. jejuni.
When we examined the localization of C. jejuni within Sigirr−/− tissues more closely, we observed that large numbers of the bacteria were in direct contact with the intestinal epithelium, particularly within crypts (Figure 3a). By co-staining for C. jejuni antigens along with either β-actin or cytokeratin 19, we could clearly visualize the cytoskeleton of the epithelial cells, relative to the localization of the C. jejuni. In addition to adherent C. jejuni, we also visualized C. jejuni co-localizing with and potentially within the epithelial layer (Figure 4a). To address whether these C. jejuni were intracellular, we examined the stained cells using confocal microscopy, to determine whether they were in fact internalized (Figure 4b and c). Indeed, in the X, Y and Z axes, labeled C. jejuni were present inside epithelial cells, often organized into spherical foci, suggesting their localization within a vesicle or phagosome (Figure 4b and c). Previous studies have identified Lamp-1, a lysosome-associated membrane protein as a marker for intracellular S. Typhimurium containing vacuoles [38], [39] as well as for phagosomes containing C. jejuni [40], [41] inside cultured epithelial cells. To address whether a similar structure was present in vivo, we stained for Lamp-1 [42], and clearly observed internalized C. jejuni within epithelial cells to be surrounded by Lamp-1 positive membrane structures (Figure 4d). Although the precise numbers of internalized C. jejuni present in a tissue section varied, we observed intracellular C. jejuni in all infected Sigirr−/− mice tested.
While our data showed that SIGIRR deficiency facilitated the ability of C. jejuni to adhere to and infect intestinal epithelial cells in vivo, resulting in overt gastroenteritis, it was unclear whether the resulting pathology depended on C. jejuni pathogenicity factors. To test this, we inoculated our WT and Sigirr−/− mice with two previously well-characterized C. jejuni mutants: ΔkpsM and ΔflaA, as well as the complemented strains for each mutant. The kpsM gene encodes the permease of the capsule polysaccharide ABC transporter. This gene deletion results in the loss of the entire capsule surrounding the microbe, which is thought to be a key virulence-associated cellular structure [33], [43]. The ΔflaA flagellar mutant lacks the primary flagellin protein, and although expression of the secondary FlaB flagellin continues, the result is a truncated flagellum and a significant loss of motility [44]. This phenotype has been previously associated with an inability to invade epithelial cells in vitro [32], and defective colonization of chicks [45].
We initially observed significant shifts in colonization for each of the mutant strains tested. Whereas wild-type C. jejuni readily colonized the intestines of Sigirr−/− mice, each of the mutant strains suffered colonization defects (Figure 5a and b, Figure S5a). The ΔkpsM mutant was significantly impaired at 3 DPI, but approached WT numbers by 7 DPI. The complemented version of this mutant was significantly less impaired for colonization at 3 DPI, and more closely resembled the colonization potential of the wild-type strain. Conversely, the ΔflaA flagellar mutant was severely impaired in colonization and was completely lost from the intestine by 3 DPI and remained absent at 7 DPI. The mutant and complemented strains were also assessed for growth in vitro, and neither mutant exhibited growth defects (Figure S5b).
In terms of pathology, each mutant exerted a substantially different effect on the gastroenteritis seen in infected Sigirr−/− mice. As might be expected given the severe colonization defect, the ΔflaA mutant did not elicit any significant inflammation or pathology (Figure 5c). In contrast, despite the ΔkpsM mutant suffering delayed colonization, it still caused overt gastroenteritis at 7 DPI that was in fact significantly worse than that seen following infection with wild-type C. jejuni (Figure 5d). To explore the basis for this exaggerated pathology, we next examined how the immune system is stimulated during in vivo C. jejuni infection.
Previous research has shown that C. jejuni activates several Toll-like receptors (TLR) including TLR2 and TLR4, and that TLR activation may play a key role in regulating host inflammatory responses to C. jejuni [21], [28]–[31]. To confirm that our wild-type C. jejuni strain (81–176) stimulated these TLRs, we used HEK-TLR2 and HEK-TLR4 reporter cells with a NF-κB/AP-1 inducible reporter- SEAP to measure stimulation of TLR2 and TLR4 in vitro. We observed significant stimulation of both receptors by C. jejuni 81–176, consistent with previously published results by Maue et al. [31] (Figure 6). To explore the impact that this activation might play in our infection model, we infected Tlr2−/−/Sigirr−/− and Tlr4−/−/Sigirr−/− mice. Although the Tlr4−/−/Sigirr−/− mice were heavily colonized by C. jejuni, (Figure 7a) they proved largely unresponsive to the pathogen, exhibiting few if any signs of the gastroenteritis seen in infected Sigirr−/− mice (Figure 7b). Notably, these mice showed little response to infection even at the gene transcriptional level. While these mice did exhibit significantly elevated expression of TNFα and IFNγ at 3 DPI, by 7 DPI, their expression of these, and other pro-inflammatory cytokines had decreased to levels similar to those in uninfected controls (Figure 2).
Conversely, the Tlr2−/−/Sigirr−/− mice were significantly more sensitive to C. jejuni infection, even compared to infected Sigirr−/− mice (Figure 7b and c), suffering exaggerated gastroenteritis by 3 DPI that involved worsened edema, crypt hyperplasia and inflammatory cell infiltration, including large numbers of neutrophils. Moreover, there were frequent signs of ulceration in these mice, along with loss of crypt structure and overall loss of epithelial integrity. Pathological scoring of tissues confirmed that the damage suffered by the Tlr2−/−/Sigirr−/− mice was significantly more severe than that seen in WT mice, and even more than that of Sigirr−/− mice at 3 DPI, though the severity of their inflammation was reduced by 7 DPI, leaving it similar in severity to that seen in Sigirr−/− mice at this time point. Consistent with their severe pathology, we observed a dramatic induction of inflammatory cytokine genes within the ceca of Tlr2−/−/Sigirr−/− mice at 3 DPI (Figure 2). This included significantly elevated levels of IL-1β, IL-6, IFNγ, KC, IL-22, IL17 and TNF-α gene transcripts, particularly at 3 DPI, although expression of many of these cytokines dropped by 7 DPI (Figure 2). Interestingly, the localization of C. jejuni was similar amongst all three SIGIRR-deficient mouse strains, with the C. jejuni seen in large numbers deep within cecal and colonic crypts, as well as inside intestinal epithelial cells (Figure 7c and data not shown). C. jejuni colonization of crypts in the Tlr2−/−/Sigirr−/− and Tlr4−/−/Sigirr−/− strains was comparable to the Sigirr−/− mice, with 41.7% (50/120) and 47.1% (66/140) of observed crypts being positively colonized, often with high numbers of bacteria in each crypt (Figure 7c). This indicates that the significant differences in pathology amongst the three SIGIRR deficient mouse strains were governed by the stimulation of the TLRs instead of by changes in the localization of the bacteria. We also assessed Tlr2−/− and Tlr4−/− single mutants for colonization and inflammation. Once again, pathogen burden was not affected by the mouse strain so long as it was accompanied by vancomycin pretreatment (data not shown). Unsurprisingly, Tlr4−/− were completely unresponsive to the presence of C. jejuni, exhibiting no inflammation (Figure S6). Tlr2−/− were much less responsive than Tlr2−/−/Sigirr−/− mice, but did show modest signs of inflammation by 7 DPI (Figure S6). Together, these results demonstrate that the majority of the inflammation seen in this model is driven by TLR4, whereas TLR2 signaling appears to play a protective role.
Based on previous data published by Rose et al. [34] and Maue et al. [31] we expected the capsule to play a role in modulating TLR responses to C. jejuni. To further explore the impact of TLR signaling during C. jejuni infection, we tested the effect of the ΔkpsM mutant on our TLR2 and TLR4 reporter cell lines. The ΔkpsM mutant stimulated both TLR2 and TLR4 to a significantly higher degree than the wild-type 81–176 strain with the complemented ΔkpsM+kpsM strain completely or nearly completely rescuing the mutant phenotype (Figure 6). To test whether these results translated to increased inflammation in vivo, we infected our different mouse strains with this mutant. As shown in Figure 8a, the ΔkpsM mutant elicited a very significant inflammatory response in the Sigirr−/− mice by 7 DPI. Moreover, it also caused exaggerated inflammation and pathology in Tlr2−/−/Sigirr−/− mice as compared to the effects of wild-type C. jejuni, yet once again there was little response in the Tlr4−/−/Sigirr−/− mice (Figure 8a). The localization of the ΔkpsM mutant in vivo was similar to that of wild-type C. jejuni, as it was frequently found in direct contact with the epithelium and deep within crypts (Figure 8b). These results were confirmed when cytokine transcript levels were assessed (Figure S7) Together, these data support previously published in vitro results [31], [34], and for the first time demonstrates that the C. jejuni capsule limits the host innate responses to this pathogen during the course of infection.
A lack of animal models, and in particular mouse models that replicate the gastroenteritis caused by C. jejuni infection in humans, has long been an impediment to the study of C. jejuni pathogenesis. Moreover, improved preclinical models of C. jejuni infection are a necessity to better define those host factors that protect against this pathogen. Here we demonstrate that the antibiotic vancomycin facilitates C. jejuni's colonization of the mouse intestine, presumably through the removal of commensal microbes that promote resistance against C. jejuni colonization. Moreover, our studies validate the use of vancomycin pretreated Sigirr−/− mice as a model for C. jejuni infection and pathogenesis, as these mice develop acute gastroenteritis following infection. We were able to define the role of innate signaling in this model through the testing of Tlr2−/−/Sigirr−/− and Tlr4−/−/Sigirr−/− mice as well as clarify specific aspects of C. jejuni pathogenesis through the testing of mutant strains. Taken together, we demonstrate that by modulating the gut microbiota as well as the innate sensitivity of the murine intestine, we have been able to develop a reliable and exciting new animal model of C. jejuni infection.
WT mice, including the C57BL/6 strain used in this study, have in the past proven resistant to infection by C. jejuni, limiting their utility as an infection model [17], [35]. This limitation in the ability to colonize mice has been linked to the intestinal microbiota, and its ability to out-compete invading C. jejuni. Previous studies have explored how to overcome this barrier by using mice with a “humanized” microbiota [17], as well as germ-free mice and mice carrying a limited microflora [19], [21], [46]. In the current study, we used an approach successfully employed with other bacterial pathogens [37], using a single pre-treatment of the antibiotic vancomycin to disturb the murine microbiota sufficiently to allow C. jejuni to establish in the intestine. This colonization was, however, insufficient to produce an effective model of inflammation, as the colonized mice remained highly tolerant to the presence of C. jejuni, displaying few if any signs of inflammation, even in the presence of a high pathogen load.
Our previous studies identified Sigirr−/− mice as displaying increased susceptibility to infection by the natural mouse pathogens S. Typhimurium and C. rodentium in terms of both the severity of disease as well as pathogen burden [25]. Normally highly expressed by the intestinal epithelium, SIGIRR acts to dampen signaling through MyD88-dependent receptors such as most TLRs as well as IL-1R [26]. Thus SIGIRR expression is thought to help maintain the relative innate hypo-responsiveness of the intestinal epithelium. While the absence of SIGIRR does not lead to spontaneous intestinal inflammation, it does leave the epithelium more sensitive to microbial stimulation through TLRs. Previous studies identified TLR2 and TLR4 as being stimulated by C. jejuni in vitro [31], [34] and our study confirmed these two TLRs actively respond to C. jejuni. Moreover, TLR4 has been identified as being a major driver of inflammation in C. jejuni-infected IL-10−/− mice [21] and the present study found the gastroenteritis seen in infected Sigirr−/− mice is almost completely TLR4 dependent. In contrast, TLR2 signaling was found to protect the intestine from exaggerated injury, potentially by promoting responses in the epithelium that limit the damage caused by the TLR4 driven inflammation.
Taken together, there are several advantages to the use of the Sigirr−/− mouse over other models of C. jejuni infection. While C. jejuni readily colonizes the intestines of newborn chickens, it only does so in a commensal fashion, thus providing little insight into its pathogenesis or how it triggers gastroenteritis. While neonatal piglet and ferret models have been used successfully as models for infection [15], [16], both have significant limitations as these animals are difficult to acquire as well as handle, and there are few immunological or genetic tools available for these species. To circumvent these issues, mice remain one of the preferred animal species for use in research, but their resistance to C. jejuni colonization and disease has limited their utility in the field. The most successful mouse model of C. jejuni infection to date has been the Il-10−/− mouse, which has several features in its favor, including a very strong and reproducible inflammatory response [20], [21], [46]. However, it also suffers from several complications, notably the propensity of the Il-10−/− mice to develop spontaneous colitis as a reaction against their own microbiota [47], forcing researchers to use more inconvenient and costly germ-free conditions [21], [46]. Additionally, IL-10 has been shown to be a key cytokine in the resolution for inflammation following infection, meaning that C. jejuni infection in IL-10−/− mice is ultimately chronic and lethal to the mice, as opposed to the acute, self-limiting infection observed in humans.
Our use of Sigirr−/− mice addresses most of these issues. When orally inoculated with a relatively low dose of C. jejuni, the vancomycin pre-treatment allowed for reliable colonization, while only causing a temporary disruption in the microbiota. Although our previous research has identified a higher inflammatory “tone” in the Sigirr−/− mice, characterized by slightly higher expression of several pro-inflammatory cytokines [25], the Sigirr−/− mice themselves do not develop spontaneous colitis in response to their own microbiota as often occurs in IL-10−/− mice. When infected with C. jejuni they developed only an acute gastroenteritis, in keeping with the clinical effects of C. jejuni infection. Moreover, the gastroenteritis bears several hallmarks of C. jejuni infection, including prominent neutrophil infiltration into the infected tissues and lumen. Our assessment of Tlr4−/−/Sigirr−/− mice determined that the vast majority of the inflammation seen in this model is TLR4 dependent, which is in agreement with previous observations in gnotobiotic IL-10−/− mice [21].
In contrast to the modest responses seen in infected Tlr4−/−/Sigirr−/− mice, when we infected Tlr2−/−/Sigirr−/− mice, we observed a significantly exaggerated and accelerated form of gastroenteritis, especially at the early stages of infection (3 DPI). Correspondingly, these mice suffered increased pathology, including widespread loss of epithelial integrity, loss of crypt structure and frequent ulceration. This was accompanied by substantially increased pro-inflammatory cytokine expression. The most severe pathology was apparent at 3 DPI, with both cytokine expression and pathology dropping substantially by 7 DPI to the point where it was no longer significantly more severe than that seen in Sigirr−/− mice. These findings are intriguing as they suggest that TLR2 plays a protective role, at least during the early stages of C. jejuni infection. Previous studies have identified roles for TLR2 in the maintenance of epithelial tight junctions in the intestine, for example by increasing the production of Trefoil Factor 3 [48], along with other barrier protective proteins [49], [50]. As recently described by our laboratory, innate inflammatory responses in the GI tract appear to reflect a tenuous balance between damaging inflammatory signals and concurrent protective or tolerance inducing innate responses that limit the resulting tissue damage. It appears this is also the case during C. jejuni infection, with TLR2 playing a key role in limiting damage suffered by the host as its immune system tries to clear C. jejuni from the intestine [51].
Aside from exploring the host response to infection, an optimal C. jejuni infection model must be able to distinguish subtle aspects of C. jejuni pathogenicity. To address this issue, we infected our Sigirr−/− mice with C. jejuni strains lacking the ability to form a capsule, as well as a flagellar mutant. Regarding the ΔflaA flagellar mutant, previous studies have found C. jejuni mutants that are non-motile or suffer reduced motility are unable to effectively colonize the intestines of chicks [45], piglets [15], or wildtype mice [52], [53]. We therefore expected the ΔflaA strain to be impaired in colonization, and indeed, the ΔflaA mutant was unable to colonize the Sigirr−/− mice or cause any level of gastroenteritis. Precisely why this mutant was unable to colonize is an interesting question. We predominantly observe C. jejuni colonization in the mucus layer and into the crypts. Both leaving the lumen of the intestine and migration through the mucus layer would presumably require fully motile bacteria. It would appear that the inability of C. jejuni with reduced motility to reach and move through these niches results in a loss of colonization potential, even in mice with reduced microbiota competition.
In the case of the ΔkpsM mutant, it exhibited a delay in colonization as assessed at 3 DPI, but by 7 DPI, its pathogen load had increased to levels similar to WT C. jejuni. This could indicate a greater sensitivity of this mutant to the innate defenses in the gut. Previous work has already linked the loss of capsule to increased sensitivity to environmental factors such as osmotic stress [54], as well as antimicrobial factors [55]. However, the reduced colonization of ΔkpsM was transient, as the mutant quickly recovered to WT levels. Despite this initial delay, the ΔkpsM strain was able to elicit significantly more severe gastroenteritis than that seen with the wild-type C. jejuni. When we tested the ΔkpsM mutant in our in vitro reporter system, we found it stimulated both TLR2 and TLR4 to a significantly higher degree than that seen with WT C. jejuni. This result is consistent with previous findings [31], [34], suggesting a role for the capsule in reducing the exposure of pathogenic bacteria to the host's immune system by masking some of the TLR activating PAMPs. For example, Rose et al. determined that the presence of a capsule reduces cytokine expression by dendritic cells exposed to C. jejuni in vitro [34], and a similar observation was made by Maue et al. in an epithelial reporter cell line [31]. Here, for the first time in vivo, we demonstrate that the capsule does help conceal C. jejuni from the host's immune system, potentially as a means to limit host driven defenses as well as perhaps limit collateral tissue damage to the host.
One of the most intriguing findings from our study involved the localization of C. jejuni in the intestines of the Sigirr−/− mice. In WT mice, C. jejuni were found predominantly in the lumen, along with a clustering at the luminal surface of the mucus layer. Notably, relatively few C. jejuni were found in direct contact with the cecal or colonic epithelium, and few were seen penetrating the crypts. In contrast, in the Sigirr−/− mice we found large numbers of C. jejuni, not only within this mucus layer, but also penetrating and accumulating in large numbers at the base of the intestinal crypts. Similar observations were made in infected Tlr2−/−/Sigirr−/− mice, although in these mice, the exaggerated damage they suffered often left C. jejuni mixed within sloughed epithelial cells as well as phagocytosed inside neutrophils at sites of ulceration. This phenotype of crypt colonization was not restricted to mice developing severe inflammation, since Tlr4−/−/Sigirr−/− mice also displayed large numbers of bacteria within their cecal crypts. Thus the factor within Sigirr−/− mice that permits C. jejuni to colonize the crypts is not a result of overt inflammation, in contrast to recent studies of S. Typhimurium colonization of the murine intestine [56], but may be due instead to more subtle differences in the microenvironment of the crypts.
Aside from penetrating intestinal crypts, we also determined that C. jejuni could invade the intestinal epithelial cells of our Sigirr−/− mice. Intracellular C. jejuni were observed in both the cecum and colon of infected mice, and were usually seen in the more mature epithelial cells, at the top of crypts rather than at their base. Co-staining with β-actin or cytokeratin 19 confirmed that the C. jejuni were inside epithelial cells, while staining for LAMP-1 localized the internalized bacteria within LAMP-1 positive vesicles. Interestingly, the presence of intracellular C. jejuni did not on its own drive significant inflammation as internalized bacteria were also found in infected Tlr4−/−/Sigirr−/− mice, which exhibited no significant signs of inflammation. This indicates that intracellular C. jejuni can exist without causing overt inflammation or pathology; however, it remains possible that this cellular invasion can play a triggering role for the overt inflammation seen in infected Sigirr−/− and Tlr2−/−/Sigirr−/− mice.
Together, these studies provide new insight into the pathogenicity of C. jejuni, and how colonization by this microbe triggers an inflammatory reaction by its host. In conventional WT mice, the commensal microbiota provide colonization resistance against C. jejuni by outcompeting the invading pathogen. Through vancomycin treatment, we were able to readily disrupt this protection, but the WT mice remained substantially tolerant to the presence of C. jejuni, resulting in almost no inflammatory response. In contrast, in the absence of SIGIRR, the murine immune system proved dramatically more responsive to C. jejuni, potentially by increasing the sensitivity of epithelial expressed TLRs. Overall, the degree of inflammation that developed in the infected intestines of the Sigirr−/− mice appeared to correlate with the invasion by C. jejuni of the intestinal crypts, and appeared almost totally dependent on the actions of TLR4. In conclusion, we present the Sigirr−/− mouse as an effective and exciting new model for the study of C. jejuni infection and pathogenesis. We speculate that our demonstration that Sigirr−/− mice can indeed be infected in a relevant fashion by C. jejuni will provide an impetus for further study, to better elucidate both the host factors and pathogenesis that drive gastroenteritis.
The wild-type C. jejuni strain used in this study is the commonly used 81–176 lab strain and all mutant and complemented strains were constructed on this background. The bacteria were routinely grown on Mueller-Hinton agar plates or broth, supplemented with the selective antibiotics Chloramphenicol and/or Kanamycin as required. Additionally, during mutant and complement construction, plates and broth were routinely supplemented with vancomycin (10 µg/mL) and trimethoprim (5 µg/mL) to prevent contamination. Cultures were routinely grown under microaerophilic conditions using anaerojars and CampyGen sachets (Oxoid) at 42°C.
To construct deletion mutants in the genes flaA and kpsM, each gene was PCR amplified with iProof (Bio-Rad) from C. jejuni 81–176 with the appropriate primers listed in Table S1. The product was polyA tailed and ligated to pGEM-T (Promega). Inverse PCR was performed on the resulting plasmid, deleting 1248 bp or 514 bp from the flaA or kpsM genes respectively. The flaA and kpsM inverse PCR products were digested with KpnI and SpeI, or KpnI and XbaI respectively, then ligated to the non-polar kanamycin resistance cassette (aphA-3) digested out of pUC18K-2 [57]. The construct was verified by sequencing and naturally transformed to C. jejuni 81–176. Mutant strains were selected by kanamycin resistance, and verified by sequencing.
To complement each of these mutants, flaA or kpsM was PCR amplified from C. jejuni 81–176 genomic DNA, digested with SpeI and MfeI, or XbaI and MfeI respectively, and inserted into pRRC [58] digested with XbaI and MfeI. The resulting construct was verified by PCR and sequencing, and naturally transformed into the corresponding mutant. Complemented strains were selected on chloramphenicol and verified by PCR and sequencing.
Additional confirmation of the phenotypes of both mutant and complemented strains were undertaken to ensure they corresponded to previously published data for these mutants. The ΔflaA mutant and complement were tested for motility, indicating the mutant was only approximately 25% as motile as the WT or complemented strain [59]. The ΔkpsM mutant was tested for NaCl sensitivity [54] and hyper-biofilm formation [60], and the complement was confirmed to restore the wild-type phenotype for both. In vitro growth curves to confirm equal growth potential between both ΔflaA and ΔkpsM mutant and complemented strains were conducted in MH broth, at 37°C under microaerophilic conditions with samples taken at 6, 24, and 48 hours post inoculation.
The C57BL/6 (WT), Sigirr−/−, Tlr2−/−, Tlr2−/−/Sigirr−/−, Tlr4−/−, and Tlr4−/−/Sigirr−/− mouse strains used in this study were all bred in-house and kept under specific pathogen-free conditions at the Child and Family Research Institute (CFRI). The combined TLR and SIGIRR deficient mice were created by cross breeding single knockout strains as described previously [25]. Mice at 6–10 weeks of age were orally gavaged with 100 µl of a 50 mg/ml vancomycin solution suspended in PBS (dose per mouse of ∼5 mg). Four hours later, each mouse was inoculated with an overnight culture of ∼107 CFUs of C. jejuni 81–176 or one of the above mentioned mutant strains. The weight of each mouse was recorded before antibiotic treatment and inoculation, and each mouse was weighed again every two days to check for weight loss/gain. Fecal samples were collected 1, 3, 5 and 7 DPI, were weighed, homogenized, serially diluted and plated onto Campylobacter agar plates containing Karmali selective supplements (Oxoid). Three and seven days post infection, mice were anaesthetized with isofluorane and euthanized by cervical dislocation. The mice were immediately dissected and their ileum, cecum, colon, mesenteric lymph nodes and spleen were isolated. Cecal and proximal colonic tissues were fixed in 10% neutral buffered formalin (Fisher). Cecal tissues were also washed to remove luminal contents and then suspended in RNAlater (Qiagen) for subsequent RNA extraction. The remainder of the cecum (including luminal contents), and other isolated tissue sections were suspended in 1 ml sterile PBS (pH 7.4) for viable cell counts. Tissue samples were homogenized, serially diluted and plated onto Campylobacter agar plates containing Karmali selective supplements (Oxoid). Following 48 hours incubation, at 42°C under microaerobic conditions colonies were enumerated, and the pathogen burden (CFUs/g of tissue) was calculated. Statistically significant differences were determined using a non-parametric Mann-Whitney test, with a p value below 0.05 used as the threshold for significance.
To monitor colonization of C. jejuni over a 25 day timeframe, three experimental groups comprising 13 WT and 15 Sigirr−/− mice total were inoculated with C. jejuni 81–176. Weights and fecal samples were taken every two days from 1 DPI to 25 DPI. CFUs present within the fecal samples were enumerated as described above and statistical significance was determined using multiple t-tests (p<0.05).
All animal experiments were performed according to protocol number A11-290, approved by the University of British Columbia's Animal Care Committee and in direct accordance with the Canadian Council of Animal Care (CCAC) guidelines. Mice were monitored for mortality and morbidity throughout their infection and euthanized if they showed signs of extreme distress or more than 15% body weight loss.
Tissues previously fixed in 10% formalin were paraffin embedded and cut for further histological analysis. The paraffin embedded tissue sections were stained with haematoxylin and eosin, and then photographed, and then used for pathological scoring. The scoring was done by two blinded observers according to previously established criteria [25]. Each tissue section was assessed for: (1) submucosal edema (0-no change, 1- mild, 2- moderate, 3- severe), (2) crypt hyperplasia (0-no change, 1: 1–50%, 2: 51–100%, 3: >100%), (3) goblet cell depletion (0-no change, 1-mild depletion, 2-severe depletion, 3-absence of goblet cells), (4) epithelial integrity (0-no pathological changes detectable, 1-epithelial desquamation (few cells sloughed, surface rippled, 2-erosion of epithelial surface (epithelial surface rippled, damaged), 3-epithelial surface severely disrupted/damaged, large amounts of cell sloughing, 4-ulceration (with an additional score of 1 added for each 25% fraction of tissue in the cross-section affected up to a maximum score of 8 (4+4) for a tissue section that had entirely lost its crypt structure due to epithelial cell loss and immune cell infiltration, (5) mucosal mononuclear cell infiltration (per 400× magnification field) (0-no change, 1- <20, 2- 20 to 50, 3- >50 cells/field), (6) submucosal PMN and mononuclear cell infiltration (per 400× magnification field) (1- <5, 2- 21 to 60, 3- 61 to 100, 4- >100 cells/field). A maximum score under this scale is 24. Statistical significance (p<0.05) was determined using a two-way ANOVA, with a Bonferroni post-test.
The paraffin embedded, formalin-fixed tissue sections were also used for immunofluorescent staining using variations on established protocols [25], [61]. Briefly, tissue sections were deparaffinized by heating for 8 minutes, clearing with xylene, rehydrating with 100%, 95%, and 70% ethanol, followed by dH2O. Antigen retrieval of the tissue sections was conducted with sodium citrate buffer (pH 6.0), in a steam bath for 30 minutes. Blocking was done with an endogenous Biotin-blocking kit (Molecular Probes) following manufacturer protocols, followed by 1 hour blocking with donkey serum blocking buffer (donkey serum in PBS containing 1% bovine serum albumin (BSA), 0.1% Triton-X100, 0.05% Tween 20, and 0.05% sodium azide). The primary antibodies used were for Actin (goat polyclonal, Santa Cruz Biotechnology), Cytokeratin 19 (goat polyclonal, Santa Cruz Biotechnology), and Campylobacter jejuni (Biotin-rabbit polyclonal, Abcam). Each was visualized using Alexa Fluor 488-conjugated donkey anti-goat IgG (Invitrogen) or Alexa Fluor 568-conjugated streptavidin (Molecular Probes). The tissues were mounted using ProLong Gold antifade reagent containing DAPI (Invitrogen). The stained slides were viewed using a Zeiss AxioImager Z1, photographed using an AxioCam HRm camera with AxioVision software. Confocal imaging was conducted with a Leica TCS SP5 system, using the Leica Application suite software.
Slides stained for C. jejuni and DAPI were used to assess crypt colonization. We used slides of formalin fixed, 7 day infected cecal tissues from WT, Sigirr−/−, Tlr2−/−/Sigirr−/−, and Tlr4−/−/Sigirr−/− mice to count the number of crypts containing visible numbers of C. jejuni. In total, 264, 146, 140, and 120 crypts were counted for each mouse strain respectively, from three slides each, each of which contained at least three tissue sections.
Tissue samples previously isolated from infected or control mice were preserved in RNAlater at −20°C for later use. RNA was extracted using a Qiagen RNeasy kit (Qiagen) according to the manufacturer's protocol. The final RNA samples were eluted from the columns in sterile, RNAse free dH2O and quantified using an ND-1000 spectrophotometer (Nanodrop). cDNA was synthesized from the RNA using an Omniscripts RT kit (Qiagen) and Oligo-dT (Applied Biological Material Inc.). Quantitative real-time PCR was carried out using an MJ mini-opticon Real-Time PCR system (Bio-Rad) using IQ SYBR Green Supermix (Bio-Rad). The primers used have been described previously [25] and are listed in Table S1. Quantification of the qPCR results was performed using Gene Ex Macro OM 3.0 software (Bio-Rad) and ANOVAs were used to determined statistical significance of the results.
HEK TLR reporter cell lines, HEK-Blue hTLR2 and HEK-Blue hTLR4, were purchased from InvivoGen (San Diego, CA, USA). HEK-Blue hTLR2 were obtained by co-transfection of hTLR2 and hCD14 co-receptor genes into HEK 293 cells, while HEK-Blue hTLR4 were obtained by co-transfection of hTLR4 and hMD-2/CD14 co-receptor genes. The cells were transfected with the secreted embryonic alkaline phosphatase (SEAP) gene and stably express SEAP under the control of a promoter inducible by NF-κB and activator protein 1 (AP-1). Thus, stimulation of hTLR2 or hTLR4 will lead to the production of extracellular SEAP in the culture medium proportional to the level of NF-κB/AP-1 activation. Cells were grown in High Glucose DMEM (HyClone, Logan, UT, USA) with 2 mM L-glutamine, 10% heat-inactivated FBS (HyClone), 100 µg/ml Normocin (InvivoGen) and selective antibiotics (1×HEK-Blue selection, InvivoGen) according to the manufacturer's instructions.
The activation of TLR2 or TLR4 was assessed by measuring the SEAP activity using QUANTI-Blue (InvivoGen) colorimetric assay. The reporter cells (5×104/well) were seeded in a 96-well plate (BD Bioscience, Mississauga, ON, Canada). The next day, cells were treated with fresh media (without selective antibiotics) containing wild type, ΔkpsM, or ΔkpsM+kpsM C. jejuni strains for 4 h. Cells treated with culture medium only, TLR2 ligand Pam3CSK4 (100 ng/mL, InvivoGen) and TLR4 ligand lipopolysaccharide (LPS, Escherichia coli K-12, 100 ng/mL, InvivoGen) serve as the negative and positive controls, respectively. For each experiment, all conditions were done in triplicate. After 4 h incubation, culture media were collected and centrifuged to remove bacteria. The supernatants (20 µl) were then incubated with QUANTI-Blue solution (180 µl) in a 96-well flat-bottom plate at 37°C for 16–18 h to allow the color development. The color change of the substrate solution corresponds to the activation of NF-κB/AP-1, which can be quantified by optical density (λ = 655 nm) measurement using a SpectraMax 384 Plus plate reader (Molecular Devices, Sunnyvale, CA, USA).
|
10.1371/journal.ppat.1002266 | Tri6 Is a Global Transcription Regulator in the Phytopathogen Fusarium graminearum | In F. graminearum, the transcriptional regulator Tri6 is encoded within the trichothecene gene cluster and regulates genes involved in the biosynthesis of the secondary metabolite deoxynivalenol (DON). The Tri6 protein with its Cys2His2 zinc-finger may also conform to the class of global transcription regulators. This class of global transcriptional regulators mediate various environmental cues and generally responds to the demands of cellular metabolism. To address this issue directly, we sought to find gene targets of Tri6 in F. graminearum grown in optimal nutrient conditions. Chromatin immunoprecipitation followed by Illumina sequencing (ChIP-Seq) revealed that in addition to identifying six genes within the trichothecene gene cluster, Tri1, Tri3, Tri6, Tri7, Tri12 and Tri14, the ChIP-Seq also identified 192 additional targets potentially regulated by Tri6. Functional classification revealed that, among the annotated genes, ∼40% are associated with cellular metabolism and transport and the rest of the target genes fall into the category of signal transduction and gene expression regulation. ChIP-Seq data also revealed Tri6 has the highest affinity toward its own promoter, suggesting that this gene could be subject to self-regulation. Electro mobility shift assays (EMSA) performed on the promoter of Tri6 with purified Tri6 protein identified a minimum binding motif of GTGA repeats as a consensus sequence. Finally, expression profiling of F. graminearum grown under nitrogen-limiting conditions revealed that 49 out of 198 target genes are differentially regulated by Tri6. The identification of potential new targets together with deciphering novel binding sites for Tri6, casts new light into the role of this transcriptional regulator in the overall growth and development of F. graminearum.
| Our knowledge of mechanisms involved in the activation and biosynthesis of DON comes largely from in vitro culture studies. Cumulated knowledge suggests that the physiological status of the fungus and the availability of nutrients are the main determining factors for DON production. Integration of various environmental cues to coordinate expression of secondary metabolic genes is thought to be mediated by a combination of global and pathway-specific transcription factors. While the global transcriptional factors respond to broad range of environmental cues such as the availability of carbon and nitrogen, the pathway-specific transcriptional factors regulate genes within a gene cluster. In F. graminearum, the transcriptional regulator Tri6 is encoded within the trichothecene gene cluster and regulates genes involved in the synthesis and transport of DON. In this report, we utilized ChIP-Seq to demonstrate that Tri6 can potentially bind to promoters and regulate genes not involved in the synthesis of DON and furthermore, many of these non-trichothecene genes are involved in various aspects of cellular metabolism, including transport and energy. Expression profiling revealed that many of the target genes are differentially regulated by Tri6, thus validating our hypothesis that Tri6 is a global regulator involved in cellular metabolism.
| Fusarium graminearum Schwabe [telemorph Gibberella zeae (Schwein.) Petch] is the causal agent of Fusarium head blight (FHB), one of the most destructive crop diseases in temperate climes throughout the world. In addition to yield reduction, FHB is often associated with the accumulation of the secondary metabolite DON in grain [1]. DON belongs to the trichothecene family of secondary metabolites; it binds to the peptidyltransferase of ribosomes thereby inhibiting protein synthesis [2]. DON accumulates in infected plant tissues, is phytotoxic, and poses considerable health risk to consumers [3]. Although considerable effort has been expended to both detect and regulate the amount of this mycotoxin in the infected cereals, there is less information with regard to the regulation of DON biosynthesis.
Our knowledge of mechanisms involved in the biosynthesis of DON and other secondary metabolites comes largely from in vitro culture studies. A considerable amount of evidence gathered over many years suggested that the physiological status of the fungus and the availability of nutrients are the main contributors for secondary metabolite production [4], [5]. Other growth conditions such as pH have also been shown to influence the production of secondary metabolite in numerous fungi including F. graminearum [6], [7]. Recent studies have assessed the role of various carbon and nitrogen sources in the induction of DON [5], [8]. For example, while the products of the polyamine biosynthesis pathway such as agmatine and putrescine strongly influenced DON biosynthesis, the study also revealed negative effects of other nitrogenous compounds on the induction of DON [5]. Other studies have also demonstrated the importance of carbon sources in the regulation of DON biosynthesis [8]. Growth conditions modified by the addition of salt solutions, hydrogen peroxide, and various phytochemicals and fungicides have also been shown to influence DON production [9].
In addition to the physiological conditions, factors affecting fungal developmental also impact the synthesis of secondary metabolites [4]. In Aspergillus species, the canonical heterotrimeric G protein/cyclic AMP/protein kinase A signalling pathway involved in diverse cellular responses including cell division, morphogenesis and pathogenic development affect the production of the secondary metabolites penicillin and sterigmatocystin. For example, mutations in fadA, a gene encoding for the Gα subunit of the heterotrimeric G protein, negatively impacts aflatoxin biosynthesis [10], [11]. Conversely, a dominant active fadA mutant inhibited expression of the transcription factor AflR, resulting in the blockage of sterigmatocystin synthesis. Interestingly, introduction of the same dominant active fadA mutant in F. sporotrichioides resulted in elevated levels of T-2 toxin, suggesting conservation of signalling pathways between filamentous fungi in the regulation of secondary metabolite synthesis [12].
This level of complexity, integrating various environmental inputs to fungal development leading to the activation of secondary metabolic gene clusters, has led to the current working model which proposes a multilevel regulation of secondary metabolism by both global and pathway-specific transcription factors [6]. In this scenario, the global transcription factors would respond and integrate disparate environmental cues such as temperature, pH and various carbon and nitrogen sources. Examples include AreA which mediates nitrogen catabolite repression in Aspergillus, and PacC which mediates pH regulation of the penicillin and trichothecene gene clusters in Aspergillus,and F. graminearum, respectively [13], [14]. One of the characteristic features of the global regulators is that they possess Cys2His2 zinc-finger domains, important for DNA binding and regulating gene expression [6]. The pathway-specific regulators on the other hand have a characteristic Zn(II)2Cys6 zinc binuclear cluster and positively regulate expression of a specific gene cluster. This is best exemplified by AflR, which regulates the sterigmatocystin biosynthetic gene cluster in Aspergillus, [15]. In F. graminearum and F. sporotrichioides, the transcriptional regulator Tri6 is encoded within the trichothecene gene cluster and regulates expression of structural genes involved in the synthesis of DON and T-2 toxin, respectively [16]–[18]. Targeted disruption of Tri6 in both Fusarium species established its role as a positive regulator of trichothecene genes [16], [18], [19]. Although this criterion designated Tri6 as a pathway-specific transcriptional regulator, evidence accumulated over the past few years have suggested that Tri6 may be representative of a global transcription factor whose expression is influenced by a variety of environmental factors. In addition to possessing Cys2His2 zinc-finger domains, Tri6 is influenced by the pH of the growth media and by a large range of nitrogen and carbon compounds [5], [8]. For example, polyamines such as agmatine and putrescine induced novel genes regulated by Tri6 [20]. Additionally, expression profiling of FHB-infected tissues identified more than 200 genes that are not part of the trichothecene gene cluster differentially regulated by Tri6 [16]. This included genes in the isoprenoid biosynthesis pathway which produces farnesyl pyrophosphate, an immediate precursor for trichothecenes, and genes involved in transport and virulence [16]. These observations suggested that genes outside of the trichothecene gene cluster are subject to Tri6 regulation.
To investigate the possibility that Tri6 is a global transcriptional regulator responsive to both environmental and developmental cues, a genome wide ChIP-Seq experiment was undertaken to identify potential new targets of Tri6. Therefore, ChIP-Seq was performed with Fusarium grown in nutrient-rich conditions, a condition optimal to detect targets involved in both growth and development. We identified 198 potential new targets of Tri6 and functional categorization associated them with energy, metabolism and other cellular processes. Expression profiling of Fusarium grown in nitrogen-deprived conditions, a condition optimal for the production of trichothecenes showed that 47 of 198 targets were differentially regulated in the tri6Δ strain. The over expression of Tri6 in the tri6Δ strain confirmed that Tri6 can auto-regulate its own expression under nutrient rich conditions. Detailed analysis of the Tri6 promoter revealed a new tandem GTGA DNA binding site, located adjacent to the previously described binding site for Tri6. This finding, together with the identification of new targets, signifies a broader regulatory role for Tri6.
Tri6 has been defined as a pathway-specific transcription factor that regulates genes of the trichothecene gene cluster under nitrogen-deprived conditions. However, to characterize Tri6 as a global transcriptional regulator, we sought to identify targets of Tri6 by performing a genome wide ChIP-Seq in F. graminearum grown in nutrient-rich conditions. The ChIP-Seq was performed in the Tri6-HA complemented strain and was compared to the Tri6Δ strain. It should be noted that the addition of the HA epitope to the C-terminus of Tri6 did not compromise its function [21]. Moreover, the presence of Tri6 protein in the Tri6-HA complemented strain was confirmed by immunoblot blot analysis using HA antibodies (Fig. S1A). ChIP DNA samples from the Tri6-HA and the Tri6Δ strains were sequenced by Illumina Genome Analyzer as 38 base tags. The software Novoalign was used to map the tags/reads to the reference genome (F. graminearum, PH-1; NRRL 31084) and the software Site Identification from Short Sequence Reads (SISSRs) was used to identify potential binding sites [22]. A browser shot of the output from the SISSRs analysis is displayed in Fig. 1. Among the 1491 enriched binding sites in the Tri6-HA complemented strain, we identified a total of 198 protein-coding genes with at least one binding site 1 Kb upstream of the ORF as potential targets of Tri6, distributed in all four chromosomes (Fig. 1). The binding site was defined by a high stringent criterion with a minimum of 120 tags and the number of tags per given target is proportional to the affinity of Tri6 to its target genes [22]. This is highlighted by the substantial enrichment of region in chromosome 2, where Tri6 is located (dotted box, Fig. 1). The 198 target genes with their tags are listed in Table S1.
The analysis of the 198 target genes by the MIPS F. graminearum FunCat database (http://mips.helmholtzuenchen.de/genre/proj/DB/Search/Catalogs/searchCatfirstFun.html) and Kyoto encyclopaedia of genes and genomes (http://www.genome.jp/kegg/kegg1.html) categorized them into various aspects of metabolism and cellular processes (Table 1). For example, genes involved in nitrogen metabolism, such as pyridoxal decarboxylase (FGSG_08249) which decarboxylates L-glutamate into GABA and ornithine aminotransferase (FGSG_02304) which transaminates L-ornithine into glutamate- γ -semialdehyde were identified (Table S2) [23], [24]. Genes involved in lipid metabolism were also identified as potential targets of Tri6. For example, triacyl glycerol lipase (FGSG_02082) and acyl-CoA thioesterases (FGSG_03286 and FGSG_02848) yield free fatty acids, which are used in the β-oxidation pathway to produce energy [25]. In addition, acetyl-CoA, the by-product of thioesterase activity, is assimilated into the energy generating TCA cycle (Table S2) [25]. The data also revealed that Tri6 has the highest affinity towards its own promoter and to other genes of the trichothecene biosynthesis pathway, namely Tri1, Tri14, Tri3, Tri12 and Tri7 (Table 2). Since the Tri genes are normally induced in nutrient-limiting conditions, the discovery of these genes as targets of Tri6 in nutrient-rich conditions suggested a new role for Tri6.
In addition to the structural genes involved in metabolism, the targets of Tri6 also included genes involved in regulatory functions and signal transduction processes (Table S2). Many of the transcription factors are classified as zinc-binding proteins and some are known to be involved in nitrogen regulation, including two genes (FGSG_05942 and FGSG_10774) with NmrA domains. Genes with NmrA domains with Rossman fold structures can act as negative regulators of nitrogen catabolite repression [26], [27]. Two members of the RAS family of GTP binding proteins (FGSG_01649 and FGSG_06209) and a homologue of GIT1, a member of the adenyl cyclase associated family of proteins (FGSG_01923), were identified as targets of Tri6. RAS has been shown previously to regulate growth and pathogenesis in Fusarium while GIT1 in S. pombe is an essential component of the cAMP signalling pathway that primarily responds to glucose [28], [29].
In summary, the genome-wide ChIP-Seq performed in nutrient-rich conditions identified new targets involved in various aspects of metabolism. The targets encompassed not only regulatory genes, but also genes involved in primary and secondary metabolism, energy, and transport. The analysis also identified genes of the trichothecene gene cluster. This was particularly interesting given the fact that these genes are activated only in nutrient-deprived conditions. This suggested that that under nutrient-rich conditions, Tri6 could potentially exert transcriptional control over itself and other Tri genes.
The ChIP-Seq data suggested that Tri6 had high affinity to its own promoter so we were interested to know if Tri6 would bind to its own promoter and self-regulate its expression. To demonstrate that Tri6 protein binds to its own promoter, the DNA-Tri6 complex was immunoprecipitated with HA antibodies from both the Tri6Δ and the Tri6-HA complemented strains grown in nutrient-rich conditions and PCR was performed using the primers spanning the upstream region of Tri6 ORF. As shown in Fig. 2, the primer set Tri6-Prom F/R (Table S3) amplified a product of 1.2 kb only from the samples immunoprecipitated from the Tri6-HA complemented strain (Lanes 1-3, Tri6-HA, Fig. 2). We could not amplify a 1.2 kb band in the samples immunoprecipitated from the tri6Δstrain (Lanes 1-3, tri6ΔFig. 2), even from 25 ng of input DNA (Lane 1, tri6ΔFig. 2). Genomic DNA was used as control to monitor the size of the PCR fragment (Lanes 1–3, Genomic, and Fig. 2). These in vivo results validated the Chip-Seq results and indicated that Tri6p can bind to its own promoter.
To demonstrate that Tri6 can bind to its own promoter and regulate its expression in nutrient-rich conditions, Tri6 expression was monitored in the wildtype strain, the tri6Δstrain, and the strain that over expressed Tri6 in the tri6 mutant background strain (tri6ΔTri6). We designed two distinct primer sets (Table S3) to monitor Tri6 transcripts. The first primer set (Tri6-ORF F/R) was designed in the coding region of Tri6 (vertical open box, Fig. 3A) and as shown, over expression of Tri6 in the tri6Δ strain (tri6ΔTri6) led to a significant increase of Tri6 transcripts compared to the wildtype strain (52±4, Tri6-ORF, Fig. 3B and 77±7; Tri6-ORF, Fig. 3C). As expected, no expression of Tri6 was detected with these primers in the tri6Δstrain (Fig. 3B and 3C). The second primer set was designed to overlap the 5′UTR region and the coding region of Tri6 (Dotted vertical box, Fig. 3A) which allowed us to monitor Tri6 transcripts originating only in the wildtype and the tri6Δstrains. As shown in the Fig. 3B and Fig. 3C, a significant increase of Tri6 expression (Tri6-5′UTR, 4.4±0.4, Fig. 3B and Tri6-5′UTR, 7±0.9, Fig. 3C) was observed in the tri6Δstrain, compared to the wildtype strain. However, over expression of Tri6 in the tri6Δstrain (tri6ΔTri6) resulted in decreased Tri6 expression (Tri6-5′UTR, 0.72±0.1, Fig. 3B and Tri6-5′UTR, 1±0.06, Fig. 3C). This suggested that Tri6 acts as a repressor, regulating its own expression in nutrient-rich conditions. We did not observe any significant change in the expression of other Tri genes in the tri6ΔTri6 strain (tri6ΔTri6; Fig. 3B and 3C), suggesting that additional factors are required for the expression of these genes.
A previous study performed in F. sporotrichioides suggested TNAGGCC as a DNA binding site for Tri6 protein [17]. A recent study that examined the promoters of genes differentially regulated by Tri6 during the F. graminearum infection process also suggested a similar DNA binding motif [16]. The results described here confirmed that Tri6 is able to bind its own promoter in vivo (Fig. 2) and regulate its own expression (Fig. 3). Since the promoter of Tri6 (−836 to −768) harbours two RNAGGCC (where R = G or A) binding sites (Fig. 4A), we employed EMSA analyses to further delineate the binding site for Tri6. All the EMSA assays were performed with purified recombinant Tri6 protein (Fig. S1B). First, we tested the probe which contained two of the RNAGGCC motifs and as the results indicated, Tri6 did not bind to this probe (Tri6-1, Fig. 4B). This prompted us to examine the region that surrounds this motif. The probe Tri6-2 which included sequences proximal to the binding sites (Fig. 4A) also did not bind the Tri6p (Fig. 4B). However, a probe (Tri6-3) which included sequences distal to the binding sites was able to bind Tri6 protein. The specificity of the binding to the Tri6-3 probe was confirmed by the addition of 10-fold excess non-labelled Tri6-3 probe in the EMSA assays (Tri6-3, competitor, ‘+’, Fig. 4B). In addition, we also used a probe which did not contain this consensus sequence as a negative control (Tri6-NS, Fig. 4B). These results suggested that RNAGGCC motif is not involved in Tri6 binding in vitro. To further confirm this, nucleotides GGCC within the motif were mutated in the Tri6-3 probe and the results indicated that the mutations did not abolish Tri6 binding (Tri6-3-1, Fig. 2). These results confirmed that the YNAGGCC motif is not required for binding of Tri6, but suggested that nucleotides outside of this motif in the Tri6-3 probe are involved in Tri6 binding. Outside of the RNAGGCC motif, three domains in the Tri6-3 probe were recognized that could potentially bind Tri6 (Fig. 5A). We designated CTGA sequence which partially overlaps the AGGCC site as Domain I and the two GTGA repeats separated by six nucleotides as Domain II and III, respectively (Fig. 5A). As the EMSA results indicated, individual nucleotide mutations within Domain I did not have any noticeable effect on Tri6 binding (Fig. 5B). However, individual mutations in Domain II and Domain III of the GTGA sequences, respectively, dramatically decreased Tri6 binding (Fig. 5C). Furthermore, combined mutations in both Domain II and III completely abolished the Tri6 binding (Fig. 5D). These results suggested that either a single GTGA sequence or GTGA repeats in the promoter of Tri6 are required for Tri6 binding.
The 198 Tri6 targets identified in nutrient-rich conditions included several genes of the trichothecene gene cluster (Table 2). Since the Tri genes are expressed only under nitrogen-limiting conditions, we were interested to know how many of the non-Tri gene targets are co-regulated with the Tri genes. To address this, we performed a genome-wide expression analysis of wildtype and tri6Δstrains grown in nitrogen-limiting conditions. As the microarray analysis indicated, we identified a total of 1614 genes that were differentially regulated in the tri6Δstrain (2-fold cut off; Table S4). Among the 870 down-regulated genes, 18 of the Tri6 targets were represented and another eight were represented in the 744 up-regulated genes (Fig. 6). Top five genes from each of the up and -down-regulated genes from the microarray analyses were selected for validation by RT-qPCR analyses (Fig. 6). If the expression threshold was set at 1.5, 49 of the 198 target genes (∼25% of the ChIP targets) were shown to be differentially regulated by Tri6 in this nutrient-limiting condition (Table S5). Out of 49 targets, 23 genes or ∼53% were annotated as unclassified by the MIPS functional annotation program and although others were classified into four major groups, only one category associated with cellular transport comprised of eight target genes were enriched to a significant level (enrichment of 19% vs 10% in the genome, p-value 0.06). Thus expression profiling provided a strong evidence to suggest that Tri6 extends its regulatory control beyond the trichothecene cluster and additionally, Tri6 can act both as a positive and negative regulator.
Much of our understanding regarding the transcription regulator Tri6 stems from earlier studies that identified its control of the mycotoxin pathway leading to trichothecene biosynthesis and designated it as a pathway-specific transcriptional regulator [17], [18]. Since then, evidence from numerous laboratories have challenged this perception and thus prompted us to hypothesize that Tri6 is a global transcriptional regulator whose influence extends beyond the trichothecene gene cluster [16], [20]. Accordingly, the ChIP-Seq experiments performed in nutrient-rich conditions demonstrated that Tri6 could bind to promoters of both structural and regulatory genes associated with many aspects of nitrogen, carbon and lipid metabolisms (Table S2). One of the intriguing revelations was that even under these conditions Tri6 was bound to the promoters of genes involved in the biosynthesis of the secondary metabolite, trichothecene 15-ADON (Table 2). Since this cluster is only activated in Fusarium grown under nitrogen-deficient conditions, binding of Tri6 to the promoters of Tri genes suggested that Tri6 could potentially act as a negative regulator, suppressing these genes under nutrient-rich conditions. The results from the RT-qPCR clearly demonstrated that with the exception of Tri6, none of the other Tri genes are subject to negative regulation by Tri6 (Fig. 3). Since the switch from nutrient-rich to deprived conditions led to differential regulation of the target genes, we conclude that Tri6 can act both as a positive and a negative regulator. Furthermore, since the regulation by Tri6 extended beyond the trichothecene gene cluster, we propose that Tri6 is a global transcriptional regulator.
Nitrogen catabolite repression, or NCR, refers to genes that are repressed when a preferred source of nitrogen such as ammonia or glutamine are present in the environment [24]. NCR is mediated by the action of global nitrogen regulators which regulate the expression of structural genes required to metabolize alternate nitrogen sources. AreA is a prototypic NCR gene regulator and has been identified in many fungi, including A. nidulans, homologues Nut1 from M. grisea and AreA-GF from F. fujikuroi [30]–[33]. Under nutrient-rich conditions, expression of AreA is repressed by the binding of the repressor protein NmrA to the promoter of AreA [31], [32]. However, under nitrogen-starving conditions, AreA repression is relieved and target genes involved in the utilization of non-preferred sources of nitrogen are expressed [31], [32]. Coincidentally, some of the known AreA target genes such as the branched amino acid transferase (FGSG_05696; ∼2.2 –fold, Table S4) and the general amino acid permease (FGSG_05574; ∼4-fold, Table S4) that are normally expressed in the wildtype Fusarium strain under nitrogen-deprived conditions are repressed in the tri6Δstrain [34]. This leads one to speculate that Tri6, similar to AreA, is involved in NCR. Studies in F. fujikuroi have also shown that AreA, in addition to regulating those genes involved in NCR, also regulates expression of genes that synthesize secondary metabolites like gibberellins and the pigment bikaverin [35], [36]. An additional feature that is shared between Tri6 and AreA is that they are both subject to auto-regulation [31]. Taken together, our evidence suggests one possible scenario where Tri6 regulates functions of AreA through NmrA in F. graminearum. In support, we identified two genes (FGSG_05942 and FGSG_10774, Table S2) with Rossman-fold domains as targets of Tri6. Rossman-folds are distinguishing features of NmrA proteins [26]. A further link between NmrA and virulence was recently established in F. oxysporum where the deletion of the transcription factor MeaB, which regulates Nmr1, an orthologue of NmrA resulted in the repression of the virulence gene Six1 [37].
The EMSA studies demonstrated that mutations in either one of the GTGA elements in the promoter of Tri6 reduced the binding of Tri6. However, combined mutations in both the GTGA elements completely abolished this binding (Fig. 5D). This suggested that the presence of two GTGA elements in close proximity to each other in the Tri6 promoter likely contributes to the affinity of Tri6 to this site. Studies with other global regulators, specifically those responding to nitrogen, such as AreA in A. nidulans or its counterpart Nit2 in N. crassa showed that binding sites located within 30bp of each other in either orientation enhances their DNA binding affinity [38]. When these criteria were applied to the analysis of Tri6 target genes, 109 of the 198 target genes harboured multiple GTGA/TCAC binding sites in their promoters, separated by eight nucleotides or less, suggesting a common mechanism of binding and regulation by Tri6 (Fig. 7). It should be noted that GTGA/TCAC motifs separated by more than eight nucleotides are represented in the promoters of all the Tri6 target genes (Table S1). The GTGA/TCAC motif described in this study is distinct from the previously reported YNAGGCC for Tri6 binding [17]. Unlike this study, Hohn et al. did not use purified Tri6 protein and furthermore, their EMSA analysis with regions of the Tri3 promoter which contained the putative YNAGGCC consensus binding site did not result in a shift [17]. This led them to speculate an alternate binding site for Tri6. It is noteworthy to indicate that Tri3 is one of the genes identified as targets of Tri6 with the GTGA repeats in its promoter.
Studies with global regulators have outlined several mechanisms that influence both binding affinity and specificity [31]. For example, proteins that binds to direct repeats most likely bind as dimers. As an example, Dal80 protein in S. cerevisiae, involved in nitrogen catabolism, binds to two closely spaced GATA elements as a dimer in either a tail-to-tail or head-to-head orientation [39]. Interactions between global and pathway-specific regulators also influence both DNA binding specificity and affinity. For example, some of the genes involved in nitrate assimilation are regulated by the interaction between the global regulators AreA and NirA in A. nidulans, while other genes are regulated by the global regulator Nit2 and the pathway-specific regulator Nit4 in N. crassa [24]. Since we observed closely spaced GTGA repeats in the Tri6 promoter of F. graminearum, we could envisage a scenario, where Tri6 could function as a dimer and repress the expression of Tri6 in nutrient-rich conditions. However, under nutrient-deprived conditions, where the expression of Tri6 increases, Tri6 could interact with other known regulators such as Tri10 or PacC to provide specificity and regulate expression in condition-specific manner [14], [19]. This interaction may also provide a platform for the Tri6-complexes to bind to an alternate site, such as the YNAGGCC proposed in several studies. Such a scenario could account for the differences between the genes regulated by Tri6 in this study from those seen in a recent study of F. graminearum infection in planta [16]. It is noteworthy to mention that the authors in that study also observed stark differences between the genes regulated by Tri6 in culture and in planta, albeit the comparison was made between two different Fusarium strains. One of the intriguing findings from the in planta study was that the promoters of genes, specifically the genes of the isoprenoid pathway essential for trichothecene production in both in planta and culture, showed enrichment of the YNAGGCC motif only in F. graminearum and not in other Fusarium species unable to make trichothecenes or in other related fungi [16]. It was proposed that co-regulation of the isoprenoid genes and the trichothecene genes represent an evolutionary adaptation specific to F. graminearum. A logical extension of this argument is that the source of nutrition, determined by the environment such as culture conditions or the type of host that this fungus infects, is the driving force behind this adaptation.
In conclusion, this study has identified new or additional role for the transcriptional regulator, Tri6. The findings from this study together with earlier studies have led us to conceive a new mode of action for Tri6. In this model, Tri6 recognizes and binds the GTGA/TCAC elements under one condition such as the nutrient-rich condition where it regulates itself and other genes involved in various aspects of metabolism. However, when the environmental conditions change, be it in culture or in planta, activation of other regulators including Tri10 and PacC may change the dynamics of Tri6p-complex and initiate new interactions with the DNA, which could involve recognition of an alternate site such as the YNAGGCC motif. Although, we have no evidence to show that Tri6 directly interacts with either Tri10 or PacC, isolation and identification of Tri6 and Tri10 protein complexes in vivo will shed more light on both the dynamics and the links between these regulators.
Fusarium graminearum wildtype strain GZ3639 (NRRL 38155) was provided by C. Babcock of the Canadian Collection of Fungal Cultures (CCFC), Agriculture and Agri-Food Canada, Ottawa. F. graminearum and transformants were grown on PDA (Sigma Chemical Co. USA) plates.
Constructs for the tri6Δand Tri6-HA complement strains have been described (Fig. S1; http://apsjournals.apsnet.org/doi/suppl/10.1094/MPMI-09-10-0210). To construct the Tri6 over expression vector, we first PCR amplified Tri6 ORF with the primer set Tri6-GUE-F/ Tri6-GUE-R (Table S2) and F. graminearum genomic DNA as template. The PCR fragment was cloned into the vector pSW-GU. The backbone of pSW-GU vector is identical to pRF-HUE [40] except that the selection marker Hygromycin gene was replaced with Geneticin, encoded by the gene neomycin phosphotransferase [40]. Transformation of the tri6Δstrain with the Tri6 over expression vector was performed according to Rasmus et al [40]. The transgenic Fusarium strains were verified by PCR (Fig. S2). PCR conditions: 200 µM Primers, 200 µM NTP's, 1 unit of Expand long Taq polymerase (Roche), 95°C for 30 s, 52°C for 30 s for annealing, 68°C for extension for 37 cycles. All PCR products were purified using Qiagen's PCR purification kit.
To induce trichothecene production in liquid culture, a two stage media protocol, modified from Miller and Blackwell was employed [41]. 20,000 spores of wildtype, tri6Δand tri6ΔTri6 over expression strains were inoculated into 4 mL of first stage growth media and incubated in Falcon Multiwell 6-well culture trays (Fig. S2). The culture trays were affixed to an orbital shaker and the spores were grown for 24 hr at 28°C in the dark with constant shaking at 170 rpm. Following 24 hr growth, the mycelial solids were washed with water and resuspended in 4 mL of second stage media (pH 4.0) [41] and then transferred to the 6-well culture trays. The mycelium was grown in second stage media under the same conditions as described previously. The supernatant was collected after 24 hr for trichothecene analyses. Trichothecenes were analysed on an AKTA 10 purifier (GE Healthcare, Canada) with direct injection of 100 µL of the culture filtrate into a 150×4.6 mm, 5 µm Hypersil ODS column (Thermo-Electron Corp.), using a methanol: water gradient from 15∶85 to 60∶40 over 25 min at a flow rate of 1 ml/min. Under these conditions, 15-ADON elutes at a retention time ∼10 min, monitored by UV 220nm.
Spores from the tri6Δand the complemented (Tri6-HA) strains were used to inoculate first stage media and grown for 19 hr as described before. The mycelia were washed with water and filtered (sterile 1MM; Whatman). The ChIP-enriched DNA was prepared according to Saleh et al. [42] with modifications. The mycelial pellet was incubated in the cross-linking buffer for 30 min with continuous shaking. After thorough washing, the pellet was ground in liquid N2 and resuspended in the lysis buffer (250 mM, HEPES pH 7.5, 150 mM NaCl, 1 mM EDTA, 1% Triton, 0.1% DeoxyCholate, 10 mM DTT and Protease inhibitor (Roche complete Mini) and incubated for 1 hr on ice. To obtain uniform fragments of cross-linked DNA, the suspension was sonicated 6x for 15 sec (60% amplitude, Misonix 4500 Sonicator). A small aliquot was used to confirm a dominant diffused band between 500 and 1000 bp. Approximately 750 µg of protein was incubated with 50 µL HA-magnetic beads (Miltenyi Biotech) and left overnight at 4°C. Immunoprecipitates were washed six times with lysis buffer and eluted with 250 µL of freshly prepared elution buffer (0.5% SDS and 0.1 M NaHCO3) with incubation at room temperature (RT) for 15 min with gentle agitation. The elution step was repeated once with 30 min of incubation at RT. To reverse the cross-linking, 20 µL of 5 M NaCl was added to the 500 µL elutes and incubated overnight at 65°C. The samples were further incubated for 1.5 hr at 45°C with 1 µL of Proteinase K (20 mg/mL), 30 mM Tris-HCl pH 6.5 and 10 mM EDTA to digest the proteins. DNA from these samples was precipitated according to Saleh et al. [42].
ChIP- enriched DNA was prepared for sequencing using the Illumina ChIP sample prep protocol with minor modifications. In brief, 2 µg of ChIP-enriched DNA was end repaired, and an adenosine overhang added to the 3′ ends. Standard Illumina adapters were ligated to the DNA and 240 bp fragments size selected from a 1.5% agarose gel. The eluted products were enriched using 18 cycles of amplification with Illumina PCR primers. Libraries were validated on the Bioanalyzer 2100 (Agilent) and Qubit fluorometer (Invitrogen, Carlsbad, CA). Each sample was run on two lanes of the Illumina GAII sequencer at the Centre for the Analysis of Genome Evolution and Function (CAGEF, University of Toronto). For each sample, 6 pmol was loaded in one lane of a standard flow cell, and 8 pmol was loaded in a second lane. The GAII was run for 38 cycles.
In total, 24,201,991 out of 29,933,903 tags from Tri6-HA complemented strain and 15,756,703 out of 22,010,411 tags from tri6Δstrain were normalized and uniquely mapped to the genome by Novoalign (http://www.novocraft.com) [43], [44]. The rest of the tags were either of low quality or mapped to multiple locations. Of the mapped tags, ∼50% were mapped to sense strand and another 50% were mapped to anti-sense strand. The mapping data were analyzed by the Site Identification from Short Sequence Reads (SISSRs) software for identification of the binding sites with tags originating from the tri6Δstrain as control [22]. The following parameters were used to run the program: 36,898,000 base pairs for the genome size, 0.8 for the fraction of genome mappable by reads, 1 for the E-value (minimum number of directional tags required on each side of the inferred binding site), 0.1 for the p-value (for fold enrichment of ChIP tags at a binding site location compared to that at the same location in the control data), 2 for the scanning window size, 240 base pairs for the average DNA fragment length, and 350 base pairs for the upper bound DNA fragment length. The switch -u was turned on to allow for identification of the binding sites supported only by reads mapped to one strand. Under these conditions, we were able to identify 1491 potential binding sites with at least 120 tags (Tag#, the sum of the tags mapped to positive strand of the left half of the binding site and the tags mapped to negative strand of the right half of the binding site) and at least 2.15 fold more tags in treatment vs control. We then compiled a list of 198 protein-coding genes with at least one binding site in 1 Kb upstream of the start site. The sum of the Tag# indicates the relative affinity of Tri6 to the binding sites in the promoter of the gene.
The two-stage media was employed to grow F. graminearum as described before. Mycelia grown for 5 hr in the second stage (nitrogen-limiting) media were filtered and ground to a powder in liquid nitrogen. The RNA was isolated from 0.25 gm of ground mycelia with 1 mL of Trizol (Invitrogen, USA) according to manufacturer's instructions. The RNA was purified further with the InviTrap Spin Cell RNA mini kit (Invitek, Germany) according to the manufacture's instructions. cDNA for the quantitative RT-PCR Analysis (RT-qPCR) experiments was synthesized from 1 µg total RNA using random hexamers using the high-capacity cDNA reverse transcription kit (Applied Biosystems, USA). RT-qPCR was performed using the Applied Biosystems StepOne Plus Real Time PCR system (ABI, Foster City, USA). Standard curves were created for the house keeping genes (β-tubulin; FGSG_09530 and Gapdh; FGSG_06257) and the genes of interest according to Relative standard curve method outlined in StepOne 2.1 software. Primers used in the RT-qPCR are listed in supplementary Table S2. Using the relative standard curve, the relative quantity (RQ) was determined by comparing target quantity in each sample to the reference sample. P values <0.05 were considered to be statistically significant.
Tri6 was cloned into pDEST17 vector (Invitrogen, USA) and expressed in BL21 pLys E.coli and the Tri6 protein was purified by the His Trap FF affinity column (GE Health care, Sweden). Double stranded DNA probes (100 ηg/mL) were labelled using T4 Polynucleotide Kinase (Fermentas MBI) and 20 µci of γP32 (Perkin Elmer, USA). The labelled probes were purified using the Quick spin column (Roche, USA) and specific activity varied from 4–7×104 cpm/ηg. Binding reactions for the EMSA were performed in 15 µL volume with 2 µg purified Tri6-HA recombinant protein, labelled probes (20,000 cpm; 2ηg) for 30 minutes at room temperature in a buffer containing 20 mM HEPES pH 7.9, 2 mM DTT, 5% (V/V) Glycerol and 1 µg poly(dIdT). The reaction was run on 6% polyacrylamide gels (acrylamide/bis ratio of 19∶1 w/w) in 50 mM Tris-HCl, pH 8.0, 50 mM sodium borate and 1 mM EDTA. The competition experiments were performed by adding unlabelled oligonucleotides to the reaction at 10X molar excess (20 ηg) of the labelled probe. The shift obtained by the Tri6 was scanned by the Storm 840 densitometer (GE healthcare, USA) and quantified by Image Quant TL 7.0 software (GE healthcare, USA). All the DNA probes used in EMSA were synthesized by Sigma Genosys (SigmaAldrich, Canada).
Wildtype and the tri6Δstrains were grown in the second stage media for 5 hrs and the RNA was extracted as described before. The integrity of RNA was confirmed by an Agilent 2100 Bioanalyzer (Agilent, Canada). RNA was then converted into cDNA using the Agilent Quick Amp Labeling Kit and converted back into labeled RNA using T7 RNA polymerase and cyanine 3-labeled CTP or 5-lableled CTP from the Agilent two-color RNA Spike-in kit. The cRNA was then hybridized to a custom F. graminearum 4X44K oligomer microarray (Agilent Technologies, CA, USA), using an Agilent gene expression hybridization kit. Both dye combinations were performed for each biological replicate. Each array consisted of 1,417 spike-in and negative controls and up to three 60-mer oligos designed for each of 13,918 predicted F. graminearum genes (NCBI GEO, Platform Accession #GPL11046). The hybridizations were scanned using the GenePix Professional 4200A scanner, and the signals quantified using GenePix Pro 6. The microarray data was transferred into Acuity 4.0 and the data was normalized using Lowess Normalization. Data points with low intensities were removed, and the dye swapped replicates were combined and expressed in log2 ratio. ANOVA was used to determine the consistency between samples for each hybridization (P<0.05). The data points were then averaged first between three biological replicates, then within the three average hybridizations. The data was then back transformed from the log scale as fold-expression. Separate datasets were then generated that contained genes that were positively identified in the chromatin immunoprecipitation data, as well as the genes that are up and down regulated in the entire genome by either 1.5-fold or greater and 2.0-fold or greater. Raw data can be accessed at NCBI GEO, Accession # GSE30892.
|
10.1371/journal.pcbi.1003796 | Emergence of Assortative Mixing between Clusters of Cultured Neurons | The analysis of the activity of neuronal cultures is considered to be a good proxy of the functional connectivity of in vivo neuronal tissues. Thus, the functional complex network inferred from activity patterns is a promising way to unravel the interplay between structure and functionality of neuronal systems. Here, we monitor the spontaneous self-sustained dynamics in neuronal cultures formed by interconnected aggregates of neurons (clusters). Dynamics is characterized by the fast activation of groups of clusters in sequences termed bursts. The analysis of the time delays between clusters' activations within the bursts allows the reconstruction of the directed functional connectivity of the network. We propose a method to statistically infer this connectivity and analyze the resulting properties of the associated complex networks. Surprisingly enough, in contrast to what has been reported for many biological networks, the clustered neuronal cultures present assortative mixing connectivity values, meaning that there is a preference for clusters to link to other clusters that share similar functional connectivity, as well as a rich-club core, which shapes a ‘connectivity backbone’ in the network. These results point out that the grouping of neurons and the assortative connectivity between clusters are intrinsic survival mechanisms of the culture.
| The architecture of neuronal cultures is the result of an intricate self-organization process that balances structural and dynamical demands. We observe that when the motility of neurons is allowed, these neurons organize into compact clusters. These neuronal assemblies have an intrinsic synchronous activity that makes the whole cluster firing at unison. Clusters connect to their neighbors to form a network with rich spontaneous dynamics. This dynamics ultimately shapes a directed functional network whose properties are investigated using network descriptors. We find that the networks are formed such that preference in connectivity between clusters is based on the similarity between their activity, a property that is called assortative mixing in networks' language. This particular choice of connectivity correlations must be rooted to basic survival mechanisms for the neurons constituting the culture.
| The theory of complex networks [1]–[5] has proven to be a useful framework for the study of the interplay between structure and functionality in social, technological, and biological systems. A complex network is no more than a specific representation of the interactions between the elements of the system in terms of nodes (elements) and links (interactions) in a graph. The analysis of such resulting abstraction of the system, the network, provides clues about regularities that can be connected with certain functionalities, or even be related to organization mechanisms that help to understand the rules behind the system's complexity. Particularly, in biological systems, the characterization of the emergent self-organization of their components is of utmost importance to comprehend the mechanisms of life [6]–[8].
One of the major challenges in biology and neuroscience is the ultimate understanding of the structure and function of neuronal systems, in particular the human brain, whose representation in terms of complex networks is especially appealing [9], [10]. In this case, structural connectivity corresponds to the anatomical description of brain circuits whereas the functional connectivity is related to the statistical dependence between neuronal activity.
Network theory and its mathematical framework have provided, through the analysis of the distribution of links, statistical measures that highlight key topological features of the network under study. These measures have facilitated the comprehension of processes as complex as brain development [11], learning [12] and dysfunction [13], [14]. Particularly, these measures have unfolded new relationships between brain dynamics and functionality. For instance, synchronization between neuronal assemblies in the developing hippocampus has been ascribed to the existence of super-connected nodes in a scale-free topology [15]; efficient information transfer has been associated to circuits with small-world features [16], such as in the the nematode worm C. elegans [17] or the brain cortex [18], [19]; and the coexistence of both segregated and integrated activity in the brain has been hypothesized to arise from a modular circuit architecture [20]–[22].
A network measure that has recently caught substantial attention is the assortativity coefficient, which quantifies the preference of a node to attach to another one with similar (assortative mixing) or dissimilar (disassortative mixing) number of connections [23], [24]. Assortative networks have been observed in both structural [20] and functional [25] human brain networks. It has been proposed that assortative networks exhibit a modular organization [26], display an efficient dynamics that is stable to noise [27], and manifest resilience to node deletion (either random or targeted) [23], [28]. Resilience is ascribed to the preferred interconnectivity of high-degree nodes, which shape a ‘connectivity backbone’ [29] that preserves network integrity. The existence of such a tightly interconnected community is generally known as the ‘rich-club’ phenomenon [19], [22], [30]. On the other hand, disassortative networks, such as the ones identified in the yeast's protein-protein interaction and the neuronal network of C. elegans [23], are more vulnerable to targeted attacks. However, in these disassortative networks, the tendency of high degree nodes to connect with low degree ones results in a star-like topology that favors information processing across the network.
The assortativty coefficient is usually calculated through the Pearson correlation coefficient between the unweighted degrees of each link in the network [23]. To account for effects associated to large networks, the Spearman assortativity measure was introduced [31] and, later, weighted assortativity measures were proposed to include the weight in degree-degree dependencies [32].
To better understand the importance of these network measures in describing neuronal networks, in vitro preparations in the form of neuronal cultures have been introduced given their accessibility and easy manipulation [6], [33]. Two major types of cultured neuronal networks are of particular interest. In a first type, neurons are plated on a substrate that contains a layer of adhesive proteins. Neurons firmly adhere to the substrate, leading to cultures with a homogeneous distribution of neurons [34]–[37]. In a second type, neurons are plated without any facilitation for adhesion. Neurons then spontaneously group into small, compact assemblies termed clusters that connect to one another [38]–[41].
The formation of a clustered architecture from an initially isotropic configuration is an intriguing self-organization process [40], [41]. This feature has made clustered networks attractive platforms to study the development of neuronal circuits as well as the interplay between structural and functional connectivity at intermediate, mesoscopic scales [39], [42]–[45]. Moreover, the existence of a two-level network, one within a cluster and another between clusters, has made clustered cultures appealing to study dynamical and topological features of hierarchical [40], [41], [45], [46] as well as modular networks [46]–[49].
In this work we use spontaneous activity measurements in clustered neuronal cultures to render the corresponding directed functional networks and study their topological properties. We introduce a novel theoretical framework that uses the propagation of activity between clusters as a measure of “causality”, although strictly speaking we should refer to as a sequence of delayed activations, giving rise to functional connections that are both directed and weighted. Based on this weighted nature of the network, we propose a new measure of assortativity that explicitly incorporates the weight of the links. We observed that all the studied functional networks derived from clustered cultures show a strong, positive assortative mixing that is maintained along different stages of development. On the contrary, homogeneous cultures tend to be weakly assortative, or neutral. Finally, in combination with experiments that measure the robustness of network activity to circuitry deterioration, we show that the strongly assortative, clustered networks are more resistant to damage compared to the weakly assortative, homogeneous ones. Our work provides a prominent example of the existence of assortativity in biological networks, and illustrates the utility of clustered neuronal cultures to investigate topological traits and the emergence of complex phenomena, such as self-organization and resilience, in living neuronal networks.
We used rat cortical neurons in all the experiments. As described in Methods, neurons were dissociated and seeded homogeneously on a glass substrate. Cultures were limited to circular areas in diameter for better control and full monitoring of network behavior. The lack of adhesive proteins in the substrate rapidly favored cell-to-cell attachment and aggregation, giving rise to clustered cultures that evolved quickly (Figure 1A). By day in vitro (DIV) , cultures contained dozens of small aggregates, which coalesced and grew in size as the culture matured. Connections between clusters as well as initial traces of spontaneous activity were observed as early as DIV . Cultures comprised of interconnected clusters by DIV , and were sufficiently stable and rich in activity for measurements. Although the strength of the connections in the network and its dynamics evolved further, we observed that the size and position of the clusters remained stable. We therefore measured dynamics already at DIV , and studied cultures up to DIV .
The example shown in Figure 1A corresponds to a culture at DIV . Clusters appear as circular objects with an average diameter of and a typical separation of . Connections between clusters are visible as straight filaments that contain several axons.
We monitored spontaneous activity in the clustered network through fluorescence calcium imaging (Figure 1B). Fluorescence images of the clustered network were acquired at a rate of frames per second, and with an image size and resolution that allowed the monitoring of all the clusters in the network with sufficient image quality (see Movie S1). Activity was recorded for typically hour, which provided sufficient statistics in firing events while minimizing culture degradation due to photo-damage.
The analysis of the images at the end of the measurement provided the variations in fluorescence intensity for each cluster and the corresponding onset times of firing (see Methods). As shown in the top panel of Figure 1C, the average fluorescence signal of the network is characterized by peaks of intense cluster activity combined with silent intervals. The accompanying raster plot reveals that this activity actually corresponds to the collective ignition of a small group of clusters, which fire sequentially in a short time window on the order of few hundred milliseconds. We denote by bursts these fast sequences of clusters' activations. Young cultures (DIV ) exhibited an activity of about , while maturer cultures (DIV ) displayed about (see also Table 1).
We observed that the time spanned between two consecutively firing clusters typically ranged between and (see Methods), as also observed by others [47],[48]. These times are fairly large compared to the eventual scale of signal integration-propagation between single neurons (), and is related to the large time scales associated to integration of the intra-clusters information. No consecutive activations were observed above , signaling the termination of a burst. We therefore use this value of as a cut-off to separate a given burst form the preceding one. Then, two clusters that fired above 200 ms cannot be influenced by one another and therefore are not causally connected.
Bursts occurred every 30 s on average for the experiment shown in Figure 1C and, as illustrated by the yellow bands in this figure, each burst typically encompassed a subset of clusters rather than the entire network. In general, however, the number of participating clusters within a burst depended on the details of the culture. Although in a typical experiment the collective firing comprised between and clusters (see Movie S1), in some experiments the entire cluster population lighted up in a single bursting episode.
The analysis of the onset times of firing provides the cluster's activation sequence within each burst. As an example, Figure 1D depicts a highly active region of the network shown in Figure 1B. This region contains clusters, and of them form a subset that regularly fires together. The series of frames show the progress in clusters' activation, revealed by the changes in fluorescence. Activity starts at the top-left cluster and progresses downwards. The time-line of sequence activation after image analysis is shown in Figure 1E, and the actual fluorescence traces are shown in Figure 1F. With our temporal resolution we could resolve well the propagation of activity from a cluster to its neighboring ones (black arrows in Figure 1E). However, and for about of the cases, the time delay between clusters' activation was either too short for detection or activation occurred simultaneously. The clusters associated to these ‘simultaneous’ events are marked in yellow in Figure 1E, and their inter-relation was treated as a bi-directional link (yellow arrow), since no causality can be inferred.
A typical recording provided on the order of bursting episodes. Some of them included the same group of clusters, although the precise sequence of activation could vary. An illustrative example is shown in Figure 1F, which depicts the fluorescence traces of clusters along two consecutive bursts. The first sequence corresponds to the sketch of Figure 1E. The orange box at the bottom of the plot indicates the relative activation time of each cluster within the window, with two clusters treated as simultaneous. To introduce the construction of the directed functional network that is described later, we note that, intuitively, the firing of cluster #9 is most likely caused by #8 and therefore both clusters are (functionally) strongly coupled. At the other extreme, cluster #1 most likely did not trigger #9, and therefore their mutual coupling is very weak. For the second burst, we note that the activation sequence is very similar, but the relative delay times differ, therefore modifying the cluster's coupling strengths. Indeed, cluster #1 and #9 are now functionally disconnected given their long temporal separation.
We carried out measurements in different clustered networks, and labeled them with capital letters as networks ‘A’‘O’. In order to compare their properties with the ones from cultures with a distinct structure, we applied the same measuring protocols and data analysis to cultures characterized with a homogeneous distribution of neurons (see Methods and Figure S1), and labeled them as networks ‘P’‘U’.
The above sequences of clusters' activations, extended to all the clusters and bursting episodes of the monitored culture, convey information on the degree of causal influence between any pair of clusters in the network. For instance, cluster #5 in Figures 1E-F can fire because of the first order influence of clusters #3 and #4, but also because of the second and third order influences of clusters #2 and #1, respectively. Hence, a realistic functional network construction should take into account these possible influences from the upstream connected clusters to build a network whose links are not only directed, but also weighted by the time delays in activation. This weighted treatment of the interaction between clusters is the major novelty of our work and the backbone of our model.
More formally, the interaction between any two clusters follows the principle of causality, i.e. the firing of cluster immediately after cluster eventually implies that cluster has induced the activity of at that particular time. The likelihood of this relation between clusters is weighted according to its frequency along the full observational time, allowing to a statistical validation. Indeed, cluster could induce the activity of various clusters, if all of them activate in a physically plausible short time window after cluster . Such a construction is illustrated in Figure 2.
To construct the directed functional networks for each studied culture we proceed as follows. First of all, we divide the entire firing sequence into the bursts of clusters' activity (Figure 2A) using the cut-off of introduced in the previous section. Once the bursts have been detected, we compute the frequency distribution of time lags between pairs of consecutive firings (Figure 2B). This frequency distribution informs about the characteristic times expected between two consecutive firings within the same burst, and hence it is a good proxy of the causal influence of a cluster on another. We will use this information to weight the causal influence of firing propagation. The frequency distribution presents a good fit to a universal Gaussian decay () in all the analyzed cultures, although the variance is specific for each culture. We indeed observed (Figure S2) that decreases with the culture age in vitro (correlation coefficient , significance ), and increases with the number of clusters present in the network (, ).
The last step in the construction of the directed functional networks consists in linking the interactions within each burst, and weighting them according to the previous frequency distribution (Figure 2B). The rationale behind this process is as follows: we hypothesize that every cluster influences other clusters (posterior in time) within a burst and, the larger the time after a cluster has fired the lower the influence we expect in the activation of another cluster (simply because the signal fades out). Then, the weighting of the interaction by the expected frequency observed in the distribution conveys the functional influence between clusters. The weights are reinforced every time the same pair of clusters' sequence is observed. After processing the full sequence we obtain a peer-to-peer activation map that is our proxy of the functional network.
We proceeded identically to construct the directed functional networks for homogeneous cultures (see Methods), with the only difference that the cut-off time corresponds to . We tested for both clustered and homogeneous cultures that the obtained functional networks were stable upon variations of the cut-off times (see Figure S3 and Discussion).
We computed the functional networks of the (‘A’ to ‘O’) realizations of clustered cultures, as well as the (‘P’ to ‘U’) homogeneous ones, and analyzed some major topological traits. Firstly, for each culture we obtained the number of nodes, the number of edges, the average degree of the networks, and its average strength (see Methods). The investigated networks and their topological measures are summarized in Table 1. Although young cultures display a richer activity, in general all networks present a similar number of nodes and a comparable functional connectivity, which is described by the number of edges, the average degree and the average strength.
Representative examples of the investigated functional networks for the clustered configuration are shown in Figure 3 (see Figure S1 for an example of the homogeneous ones). The position of the nodes and their size are the same as the actual clusters for easier comparison. Edges in the directed network are both color and thickness coded to highlight their importance, with darker colors corresponding to the highest weights. This representation reveals those pairs of clusters that maintain a persistent causality relationship over time. Nodes are also color coded according to their strength, i.e. the total weight of the in- and out-edges.
The functional networks exhibit some interesting features. First, there are groups of nodes that form tightly connected communities. These communities actually reflect the most frequent bursting sequences. Second, nodes preferentially connect to neighboring ones with some long-range connectivity, and often following paths that are not the major physical connections. This indicates that the structural connectivity of the network cannot be assessed from just an examination of the most perceivable processes. And third, as shown in Figure S4A, we observed that there is no correlation between the width of the physical connections and their weight (, ), or the size of the nodes and their strength (, , Figure S4B), and indicates that the dynamical traits of the network cannot be inferred from its physical configuration, stressing the importance of the functional study.
We also observed that the size of the clusters did not correlate with their average activity (, , Figure S4C), i.e. small and big clusters displayed similar firing frequencies, and of on average. However, since some clusters are initiators of activity and others just followers, we also computed the relative contribution of a given cluster size to initiate activity in the network. We found no significant correlation between initiation and cluster size (, , Figure S4D). These results strengthen the conclusion that one cannot predict the clusters that will initiate activity, or the most persistent sequences, by just a visual inspection of cluster sizes and their distribution over the network.
These analyses are important in the context of the work by Shein-Idelson and coworkers [41], who studied the dynamics of isolated clusters similar to ours, and observed that their firing rate increased from to as the clusters' radii escalated from to . This remarkable difference in the dynamics between ‘isolated’ and ‘networked’ clusters reflects the dominant role of the network circuitry in shaping its dynamics.
Finally, to crosscheck that the results found for the functional networks presented here are robust to the inference method, we have also performed a classical mutual information analysis to construct functional networks for the same cultures (see Methods). The results obtained with the mutual information analysis are totally in agreement with the constructed functional networks using time delays.
We determined the values of the weighted formulation of assortativity, both for the Pearson and Spearman correlations, with values in (see Methods for the generalization of assortativity to directed weighted networks). Positive values of the weighted assortativity indicate that nodes with similar strength tend to connect to one another, while negative values mean the preferred interconnectivity of nodes with different strength. In Table 1 we can observe that all clustered networks (labeled ‘A’-‘O’) exhibit a positive weighted assortativity, in the range for the Pearson construction and for the Spearman one. Although the values fluctuate across different cultures, the two assortativity measures provide the same value within statistical error, and reflect that network size corrections provided by the Spearman's treatment have little influence in strongly assortative networks.
To assess the importance of the measured assortativity values, we have also computed the weighted rich-club [50]. The rich-club phenomenon refers to the tendency of nodes with high degree to form tightly interconnected communities, compared to the connections that these nodes would have in a null model that preserves the node's degree but otherwise is totally random. Given the positive assortativity found, we analyzed whether this finding is also reinforced by the existence of rich-club structures.
The weighted formulation for the rich-club takes into account the node's strength instead of the degree, and is particularly useful in situations in which the weights of the links can not be overlooked [51]. The evaluation of the rich-club is performed by computing the ratio between the connectivity strength of highly connected nodes and its randomized counterpart, and for gradually higher values of the strength threshold . The detailed calculation is described in the Methods section, and the results of the analysis for representative networks is shown in Figure S5. Ratios larger than indicate that higher strength nodes are more interconnected to each other than what one would expect in a random configuration. On the contrary, a ratio less than reveals an opposite organizing principle that leads to a lack of interconnectivity among high-degree nodes. After the calculation of the ratios for all the studied clustered networks, we found a positive tendency towards the creation of rich-clubs in all of them (Figure S5), which is in good agreement with the observed values of assortativity.
The above network measures were also analyzed in experiments with a homogeneous distribution of neurons (labeled ‘P’‘U’). The results are summarized in Table 1. Interestingly, the assortativity values are much lower (by an order of magnitude on average) than the ones for clustered cultures, in the range for Pearson's and for Spearman's. Accordingly, the rich-club phenomenon for the homogeneous cultures vanishes (Figure S5).
Several studies highlight the importance of assortative features for network resilience to damage. Given the strong assortativity of our clustered cultures, we carried out a new set of experiments to investigate the concurrent presence of resilient traits. As described in Methods, we considered two major ‘damaging’ actions to the network. In a first one, we gradually weakened the excitatory network connectivity by means of the AMPA-glutamate antagonist CNQX, and measured the decay in spontaneous activity as connectivity failed. In a second one, we continuously exposed a culture to strong fluorescence light, therefore inducing photo-damage to the neurons. This action resulted in random neuronal death across the network and hence a progressive failure of its spontaneous dynamics. The rate of activity decay upon radiation damage provided an estimation of the resistance of the network to node deletion. These investigations were carried out at the same time in clustered cultures (strongly assortative) and in homogeneous ones (weakly assortative or neutral). Their comparison provided a first reference to relate assortativity, network topology and resistance to damage.
Figure 4A shows the results for the application of CNQX to clustered cultures. We first monitored each cluster individually in the unperturbed case, and measured its average firing activity along . We then applied a given drug concentration, measured the firing activity for another , and computed the relative changes in activity respect to the unperturbed case, as . The protocol was repeated until activity ceased. Two illustrative examples of the action of CNQX on network activity are provided in Figure 4. In a clustered cultured and for weak CNQX applications () the activity in some clusters increases, while in some other decreases, and on average the network firing rate remains stable (). As [CNQX] is increased to , we observe that most of the clusters have reduced their activity, although there are still some that maintain a high activity or even increase it. This different behavior from cluster to cluster suggests that clustered networks are highly flexible, and that they may have mechanisms to preserve activity even with strong weakening of the connectivity. Conversely, homogeneous cultures (Figure 4B) lose activity in a more regular and faster way. These networks are characterized by a highly coherent dynamics [36], [37], and therefore all neurons in the network reduce activity similarly as CNQX is applied. Interestingly, for the shown homogeneous culture has almost completely silenced (), while the clustered culture is still highly active. We repeated this study on different realizations of each culture type and observed that, on average, the critical concentration at which activity complete stopped was for clustered and for homogeneous networks (Figure 4C).
Figure 4D shows the results for the resistance of the networks to node deletion as a consequence of direct photo-damage to the neurons. As can be observed, homogeneous cultures decay in activity much faster than the clustered ones, pinpointing the general resistance of clustered cultures to structural failure.
Clustered neuronal cultures have a unique self-organizing potential. An initially isotropic ensemble of individual neurons quickly group to one another to constitute a stable configuration of interconnected clusters of tightly packed neurons. The formation of the clustered network is primarily a passive process governed by the pulling forces exerted by the neurites. Interestingly, aggregation occurs even in the absence of glial cells and neuronal activity [40], and is maintained up to the degradation of the culture [40], [52], [53]. Our work shows that this self-organizing process drives the network towards specific dynamic states, which shape a topology of the functional network that is distinctively assortative. We note that the number of clusters and their distribution are initially random. Therefore, a wide spectrum of physical circuitries and functional topologies are in principle attainable. However, in all the studied cultures, the network drives itself towards markedly assortative topologies with a ‘rich-club’ core. The emergence of these distinct topological traits, concurrently with a stronger network's resilience to activity deterioration, pictures a self-organizing mechanism that enhances network survival by procuring a robust architecture and dynamic stability.
We remark that the link between assortativity and resilience is based on the comparison between the response of clustered and homogeneous cultures upon the same perturbation. To obtain conclusive evidences that assortativity confers resilience traits exclusively from topology, we would require an experimental protocol in which we could arbitrarily ‘rewire’ the connectivity between clusters, or shape in a control manner different circuitries while preserving the number of nodes in the network. Although these strategies are certainly enlightening, they are of difficult development and a major experimental challenge.
We infer the functional connectivity maps of the clustered networks from their spontaneous dynamics. We considered small-sized networks to simultaneously access the entire population ( clusters). The approach that we have used to characterize this functional connectivity is based on the analysis of the time delays between consecutive clusters' activations. The uniqueness of our approach is to use these time delays to provide a direct measure of causality, giving rise to a functional network that is both directed and weighted, with the weights given by a decaying function that follows the frequency of the delay between pairs of clusters. Our formulation is simple and naturally derives from the intrinsic dynamics of the network.
We used two main parameters to quantitatively construct the directed functional network, namely the cut-off time for causality, and the variance of the Gaussian-like weighting function. The cut-off time is set to , two times the maximum measured time delay between consecutive activations. The importance of the cut-off is first to discriminate two successive bursting episodes, and second to exclude individual firing events from an actual cascade of activations. Although these individual firings account for less than 2% of the total activations, they may occur in regions of the culture that are physically distant -though temporary close- from an actual sequence, and therefore they would add spurious, long-range functional connections to the network.
On the other hand, the variance is obtained from a Gaussian fit of the distribution of consecutive activation delays within bursts. The value of is specific for each culture to take into account particular differences in the dynamics of the network, specifically the culture days in vitro or the number of clusters (Figure S2), parameters that could affect the delay times of activation. Young cultures for instance exhibit longer time delays between pairs of clusters, leading to a distribution shifted towards higher values and therefore a larger .
We tested that the obtained functional networks were stable upon variation of the above parameters. In particular, to examine whether the choice of the cut-off does or does not substantially affect the features of the generated functional network, we performed a sensitivity analysis on this parameter. As the process of generating the network from the sets of bursts is deterministic, we analyzed the influence of the cut-off value on the formed groups of firings. To quantify the variation on the bursts generated for different values of the cut-off, we calculated the variation of information [54] between the grouping of bursts at a certain cut-off value and the previous one as a measure to assess their difference (Figure S3). In the case of clustered cultures, we found that for values of cut-off of the variation of information is, on average, on the order of . In the homogeneous case, for cut-off values of ms, this value is on the order of . This means that varying the cut-off values in these regions does not substantially change the grouping of the bursts, and therefore the generated networks are equivalent.
To assess the goodness of our construction in inferring the functional connectivity of the clustered networks, we compared our connectivity maps with those procured by information theoretic measures, such as Mutual Information or Transfer Entropy, applied to the original fluorescence recordings. These approaches have been used to draw the topological properties of neuronal networks in vitro, both in electrode recordings [55], [56] and calcium fluorescence imaging [57], [58]. The comparison of our method with these theoretic measures showed that the identified functional links were fundamentally the same, with small quantitative differences associated to the particular weighting procedures.
Our functional networks consistently maintained high assortativity values, and along a wide range of days in vitro. We also observed that, by contrast, the assortativity analysis in homogeneous cultures procured neutral or low assortativity values, a result that is supported by other studies in homogeneous networks similar to ours [55]. In our study, we have seen that the clustered, assortative networks exhibit a higher resilience of the network to damage compared to the homogeneous, non-assortative ones. Different studies also highlighted the importance of assortativity and the ‘rich-club’ phenomenon on higher-order structures of the network, in particular resilience, hierarchical ordering and specialization [10], [30].
Several studies in brain networks advocate that the functional connectivity reflects the underlying structural organization [59]–[61]. To shed light on this interrelation in our cultures, we would need the identification of all the physical links between clusters. The top images of Figure 3 indeed suggest that some structural connections could be delineated by a simple visual inspection. However, we observed by green fluorescence protein (GFP) transfection that physical connections have long extensions and may easily link several clusters together, and not just in a first-neighbor manner as seen in the images. Since the images provide a very poor subset of the entire structural layout, a complete description of the physical circuitry must be carried out before comparing the structural and functional networks. Such a detailed identification is difficult, and requires the use of a number of connectivity-labeling techniques. Nevertheless, for the connections that we could visualize, we draw two major conclusions. First, that neither the width of the physical connections nor the size of the clusters were related to a particular trait of the functional links, such as the weight of the connections or the strength of the nodes (Figure S4). And, second, that our construction inferred strong functional links between clusters that were not directly connected in a physical manner, highlighting the importance of indirect paths as well as long-range coupling in the flow of activity.
The identification of the full set of structural connections would certainly provide invaluable information to investigate the interplay between structure and function in our networks. In this context, the recent work by Santos-Sierra et al [52] is enlightening. They analyzed some major structural connectivity traits in clustered networks similar to ours, and observed that the networks were strongly assortative as well. Assortativity emerged at early stages of development, and was maintained throughout the life of the culture. Hence, in clustered cultures, the combined evidences of this study and ours hints at the existence of assortative properties in both structure and function.
An interesting peculiarity of our experiments is that, in most of the studied clustered cultures, the spontaneous bursting episodes comprised of a small subset of clusters rather than the entire network. This activation in the form of groups or moduli is often referred as conditional activity. It contrasts with the coherent activity of homogeneous cultures, where the entire network lights up in a short time window during a bursting episode. Given the acute differences in assortativity between clustered and homogeneous cultures, we hypothesize that the modular dynamics by itself increases or reinforces assortative traits in the functional network.
We finally remark that our neuronal cultures are spatial, i.e. embedded in a physical substrate, which imposes constraints to the layout of connections and, in turn, their assortative characteristics [52], [62]. Spatial networks have caught substantial interest in the last years to understand the restrictions—or advantages—that metric correlations impose on the structure and dynamics of complex networks [63], in particular brain circuits [64]. Vértes et al showed that spatial constraints delineate several topological properties of functional brain networks [65], and Orlandi et al showed that the initiation mechanisms of spontaneous activity are governed by metric correlations inherited by the network during its formation [36]. Strong spatial constraints in clustered networks can be attained by anchoring the neuronal aggregates in specific locations, for instance through carbon nanotubes [39]. The comparison of the functional maps of such a forced organization with our free one is enlightening, and would shed light on the importance of structural constraints in shaping functional connectivity.
To conclude, we have presented a simple yet powerful construction to draw the directed functional connectivity in clustered neuronal cultures. The construed networks present assortativity and ‘rich-club’ features, which are present concurrently with resilience traits. Our analysis has been based on spontaneous activity data, and may certainly vary from evoked activity. Hence, the combined experimental setup and functional construction can be viewed as a model system for complex networks studies, specially to understand the interplay between structure and function, and the emergence of key topological traits from network dynamics. Also, the spatial nature of our experiments may also procure invaluable data to understanding the role of short- and long-range connections in network dynamics; or to investigate the targeted deletion of the high degree nodes that shape the backbone of the network. The latter is a powerful concept that may assist in a detailed exploration of resilience in neuronal circuits, for instance to model the circuitry-activity interrelation in neurological pathologies.
All procedures were approved by the Ethical Committee for Animal Experimentation of the University of Barcelona, under order DMAH-5461.
In our experiments we used cortical neurons from day old Sprague-Dawley rat embryos. Following standard procedures [36], [66] dissection was carried out in ice-cold L- medium enriched with glucose and gentamycin (Sigma-Aldrich). Cortices were gently extracted and dissociated by repeated pipetting.
Cortical neurons were plated onto glass coverslips (Marienfeld-Superior) that incorporated a poly-dimethylsiloxane (PDMS) mold. The PDMS restricted neuronal growth to isolated, circular cavities in diameter. Prior plating, glasses were washed in nitric acid for 2 h, rinsed with double-distilled water (DDW), sonicated in ethanol and flamed. In parallel to glass cleaning, and following the procedure described by Orlandi et al. [36], several diameter layers of PDMS thick were prepared and subsequently pierced with diameter biopsy punchers (Integra-Miltex). Each pierced PDMS mold typically contained to cavities. The PDMS molds were then attached to the glasses and the combined structure autoclaved at , firmly adhering to one another. For each dissection we prepared identical glass-PDMS structures, giving rise to about cultures of in diameter. Neurons were plated in the PDMS cavities with a nominal density of , and incubated in plating medium at 37°C, CO2 and humidity. Plating medium consisted in of foetal calf serum (FCS, Invitrogen), of horse serum (HS, Inivtrogen), and 0.1% B27 (Sigma) in MEM Eagle's-L-glutamate (Invitrogen). MEM was enriched with gentamicin (Sigma), the neuronal activity promoter Glutamax (Sigma) and glucose.
Upon plating, the absence of adhesive proteins in the glass substrate rapidly favored cell-cell attachment and, gradually, the formation of islands of highly compact neuronal assemblies or clusters that minimized the surface contact with the substrate. Clustered cultures formed quickly. By day in vitro (DIV) 2 the culture encompasses dozens of small aggregates that coalesce and grow in size as the culture matures. Spontaneous activity and connections between clusters were observed by DIV . Clusters at this stage of development also anchored at the surface of the glass and, although they continued growing and developing connections, their number and position remained practically stable along the next weeks. At the moment of measuring, each PDMS cavity contained an independent culture formed by interconnected clusters.
Clustered cultures were maintained for about weeks, as follows. At DIV the medium was switched from plating to changing medium (containing FUDR, Uridine, and HS in enriched MEM) to limit glial cell division. Three days later, the medium was replaced to final medium (enriched MEM with HS), which was then refreshed periodically every three days.
Overnight exposure of the glass coverslips to poly-l-lisine (PLL, Sigma) provided a layer of adhesive proteins for the neurons to quickly anchor upon seeding, leading to cultures with a homogeneous distribution of neurons over the substrate. The remaining steps in the preparation and maintenance of the cultures were identical as the clustered ones, i.e. we used the same nominal neuronal density for plating, we included PDMS pierced molds to confine neuronal growth in cavities in diameter, and we refreshed the culture mediums in the same manner.
The pharmacological protocols described below were used identically in clustered and homogeneous cultures.
The acquired images (recorded at a typical speed of fps) were first analyzed with the Hokawo 2.5 software to extract the fluorescence intensity of each cluster as a function of time. The regions of interest (ROIs) were chosen manually and typically covered an area of pixels, each ROI corresponding to a single cluster. As illustrated in Figure 1C and Figure 1F, activity is characterized by a stable baseline (resting state) interrupted by peaks of fluorescence that correspond to clusters' firings. At the onset of firing, the fluorescence signal increases abruptly due to the fast intake of ions. Fluorescence then reaches a maximum, and slowly decays back to the baseline in s.
The algorithm that we used to detect the onset of firing for each cluster was as follows. We first corrected the fluorescence signal from small drifts, and calculated the resting fluorescence level by discarding the data points with an amplitude two times above the standard deviation (SD) of the signal. The corrected signal was then expressed as . We next took and computed its derivative in order to detect fast changes in the fluorescence signal. Finally, the onset of ignition in cluster was defined as the time where a maximum in was accompanied by values of two times above the SD of the background signal, and for at least 5 frames.
Recordings in homogeneous cultures provided the activity of neurons in an circular area in diameter. Neurons were marked individually as regions of interest in the images and the corresponding fluorescence time traces extracted using custom-made software. Ignition times for each neuron were next obtained by using the sub-frame resolution method described above (detailed in Ref. [36]), and that consisted in fitting two straight lines to the fluorescence data, a first fit encompassing the points in the background region prior to firing, and a second fit including the points during the fast rise in fluorescence that follows ignition. The crossing point of the two lines provided the onset of firing.
The extension of this analysis to all the active neurons within a burst, and along all the bursts, finally provided the entire set of ignition sequences. The construction of the directed functional networks for the homogeneous cultures was then carried out identically as the clustered ones.
Recordings in clustered cultures typically lasted for h and contained between bursts in the quietest networks and bursts in the most active ones. To test whether bursts sufficed to draw the functional networks, we carried out a control experiment in which we monitored spontaneous activity along h in a standard clustered culture, measured at DIV and containing nodes (Figure S6). We then analyzed the data using two different procedures. In the first one we drew the functional connectivity using the data extracted from the entire recording, and determined its assortativity values. In the second procedure, we separated the recorded sequence in three blocks, each long, built the functional connectivity for each block, and computed the respective assortative values. The studied culture fired in a sustained manner at a rate of , and procured a total of bursts. Thus, each block typically contained about bursts.
The results (Figure S6) led to two major conclusions. First, that the functional connectivity is very similar among the blocks, and between any of the blocks and the entire recording, providing assortativity values that are compatible within statistical error. And second, that the first 40 min of recording (with 45 bursts only) sufficed to shape the major traits of the functional network, therefore validating our strategy of using h of acquisition to procure a reliable estimate of the functional connectivity of the network and its assortative traits.
Here we describe the calculation of the assortativity coefficients, assortativity errors and the rich-club distributions. In the process, we have to define the assortativity for directed weighted networks.
Mutual information [56], [72] is a particular case of the Kullback-Leibler divergence [73], an information-theoretic measure of the distance between two probability distributions. In fact, the mutual information between two stochastic variables and provides an estimation of the amount of information gained about when is known.
Let us indicate by the time series corresponding to the -th cluster, with and the total number of time frames involved in the observation process. The time series adopted for the successive analysis are obtained by mapping the observed train of cluster activations to another time series termed walk, defined by (17)
In the specific case of our analysis, the mutual information between two time series and , corresponding to two different clusters, is interpreted as the amount of correlation between the dynamics of cluster and . In general, the time scale of the correlation between two time series is not known a priori. Such a time scale corresponds to the time delay required to maximize the gain of information. Therefore, in the spirit of Fraser and Swinney [74], we define the time delayed mutual cross information between and by (18)where and are indices running over some partition of the observed time series. In Eq. (18), indicates the probability to find a value of time series in the -th interval, is the probability to find a value of time series in the -th interval, whereas denotes the joint probability to observe a firing from the -th cluster falling in the -th interval and a firing from the -th cluster falling in the -th interval exactly time frames later.
For the sake of simplicity, in the following we will adopt the more concise notation to indicate the time delayed mutual cross information. Finally, in order to gain the highest amount of information about the dynamics of cluster by observing cluster , we consider only the maximum value of with respect to the time delay .
We estimate the importance of the observed amount of correlation by performing the above analysis on surrogate data. Surrogates adopted in this study are time series generated by randomly reshuffling the temporal observations of the firing series , for each cluster separately. Such a procedure destroys any correlation between pairs of time series while preserving the empirical probability distribution, thus allowing to test the null hypothesis that the observed correlation is obtained by chance.
We indicate by the walk corresponding to the surrogate obtained from time series and with the time delayed mutual cross information between and . We perform 200 independent random realizations of surrogates for each pair and we estimate the corresponding expected value of the maximum mutual cross-information, as well as the root mean square of the underlying distribution.
Hence, we fix a priori the significance of the hypothesis testing and we estimate the -score corresponding to each pair by . Therefore, the observed correlation between cluster and is said to be statistically significant if , where is the standard error function. Finally, we obtain the functional network of clusters by building the weight matrix whose elements are defined by if , and if .
|
10.1371/journal.pcbi.1004915 | Bridging Mechanistic and Phenomenological Models of Complex Biological Systems | The inherent complexity of biological systems gives rise to complicated mechanistic models with a large number of parameters. On the other hand, the collective behavior of these systems can often be characterized by a relatively small number of phenomenological parameters. We use the Manifold Boundary Approximation Method (MBAM) as a tool for deriving simple phenomenological models from complicated mechanistic models. The resulting models are not black boxes, but remain expressed in terms of the microscopic parameters. In this way, we explicitly connect the macroscopic and microscopic descriptions, characterize the equivalence class of distinct systems exhibiting the same range of collective behavior, and identify the combinations of components that function as tunable control knobs for the behavior. We demonstrate the procedure for adaptation behavior exhibited by the EGFR pathway. From a 48 parameter mechanistic model, the system can be effectively described by a single adaptation parameter τ characterizing the ratio of time scales for the initial response and recovery time of the system which can in turn be expressed as a combination of microscopic reaction rates, Michaelis-Menten constants, and biochemical concentrations. The situation is not unlike modeling in physics in which microscopically complex processes can often be renormalized into simple phenomenological models with only a few effective parameters. The proposed method additionally provides a mechanistic explanation for non-universal features of the behavior.
| Dynamic systems biology models typically involve many kinetic parameters that reflect the complexity of the constituent components. This mechanistic complexity is usually in contrast to relatively simple collective behavior exhibited by the system. We use a semi-global parameter reduction method known as the Manifold Boundary Approximation Method to construct simple phenomenological models of the behavior directly from complex models of the underlying mechanisms. We show that the well-known Michaelis-Menten approximation is a special case of this approach. We apply the method to several complex models exhibiting adaptation and show that they can all be characterized by a single parameter that we denote by τ. The scenario is similar to modeling complex systems in physics in which a large number of microscopically distinct systems are mapped onto relatively simple universality classes characterized by a small number of parameters. By generalizing this approach to dynamical systems biology models, we hope to identify the high-level governing principles that control system behavior and identify their mechanistic control knobs.
| Complexity is a ubiquitous feature of biological systems. It is both the origin of the richness of biological phenomena and a major hurdle to advancing a mechanistic understanding of that behavior. Mathematical models, formulated as differential equations of biochemical kinetics for example, supply many tools for improving our understanding of complex biological systems. Systems biology is largely concerned with identifying mechanistic explanations for how complex biological behaviors arise [1–3]. However, mathematical models are never a complete representation of a biological (or physical or chemical) system. Indeed, one of the advantages to mathematical modeling is the ability to apply simplifying approximations and abstractions that provide insights into which components (or collection of components) of the system are ultimately responsible for a particular behavior [4]. A mathematical model, therefore, reflects the judicious distillation of the essence of the complex biological system into a more manageable representation. A good mathematical representation, while not complete, will be both complex enough to convey the essence of the real system and sufficiently simple to reveal useful mechanistic insights that enable the prediction of the system behavior under new experimental conditions, i.e., “as simple as possible, but not simpler.”
Biological research has collected a wealth of knowledge about gene regulatory networks, epigenetic controls, and biochemical reactions from which systems-level behavior derives. While this enterprise is not complete, it is sufficient in many cases to motivate models that are reasonably accurate surrogates of the real system. Exhaustive pathway maps are nearly overwhelming in their complexity [5]. Such models are often very complex, reflecting both the wealth of information available and the intricacies of the underlying mechanisms. This complexity is manifested, for example, in the high-order dynamics of the model, the number of interacting heterogeneous components, or the nontrivial topology of the network structure. These models typically have a large number of parameters that are unknown and which are left to be inferred from data.
The problem of parameter estimation has consequently received considerable attention in the systems biology community. Over-parameterized models are often “sloppy,” i.e., leading to extremely ill-posed inference problems when fitting to data [6–12]. Identifiability analysis is useful for determining which parameters’ values can be estimated from data [13–16], and optimal experimental design methods judiciously choose experiments that can most efficiently produce accurate parameter estimates [15, 17–25]. This enterprise is in many respects the natural continuation of the program of cataloging the complex web of gene regulatory networks and protein signaling cascades. Unknown parameters represent a gap in our knowledge of a specific biological system that ought to be filled.
The present work looks to answer an orthogonal question. A parameterized model can be interpreted as class of potential biological systems. Different parameter values correspond to distinct members of this class that have a related structure but differ in the microscopic specifics, i.e., parameter values. For example, parameter values may vary depending on cell-type, developmental stage, species, or many other factors. Rather than estimate all the parameters for specific biology systems, we seek a characterization of the biologically relevant behavior for all systems in the model class. Because parameter inference problems are ill-posed there are many members of the model class that exhibit identical systems-level behavior. We therefore expect that a minimal model with many fewer parameters exists that reproduces the same behaviors as the family of biological systems. In other words, we would like to characterize the class of microscopic models with indistinguishable macroscopic behavior. In addition, we would like to identify which combination of microscopic components controls the collective behavior.
Our approach to this problem is a non-local parameter reduction method known as the Manifold Boundary Approximation Method (MBAM) [12, 16, 26]. Model reduction is an active area of research and there are many techniques available. Common methods involve exploiting a separation of scales [27–29], clustering/lumping similar components into modules [30–32], or other methods to computationally construct a simple model with similar behavior [33, 34]. Many methods have been developed by the control and chemical kinetics communities focused on dynamical systems [27–29, 34, 35]. Systems biology has been a popular proving ground for new methods [33, 36–39].
Most model reduction methods suffer from two problems that make them unsuitable for the present work. First, many techniques, particularly automatic methods, produce “black box” approximations that are not immediately connected to the complicated, mechanistic model. In contrast, MBAM connects the microscopic to the macroscopic through a series of limiting approximation that provide clear connections between the macroscopic control parameters and the microscopic components from which they are derived. Second, most methods make “local” approximations, in the sense that they find computationally efficient approximations to a single behavior. However, we seek a (semi-) “global” approximation that can reproduce the entire behavior space of a model class. This is a challenging problem; brute force exploration of the parameter space is impossible because of its high-dimensionality. MBAM solves this problem by using manifold boundaries in behavior space as approximate models [26]. Manifold boundaries are topological features and therefore characterize the global behavior space [16].
Finding a minimal, “distilled” version of a complicated model has many practical applications. It identifies the system’s control knobs that could effectuate a change in the system’s behavior, reducing the search space for effective control methods. It highlights the “design principles” underlying the system and inspires approaches for engineering synthetic systems. Finally, it leads to conceptual insights into the system behavior that deepen the understanding of “why it works.”
In this paper we show that the well-known Michaelis-Menten approximation is a simple case of the MBAM. We then use this method to derive minimal models of adaptation discovered by Ma et al. [40] and a more complex mechanical model of EGFR signaling due to Brown et al [7]. Our primary result is that adaptation can be characterized by a single dimensionless parameter, τ, the ratio of the activation and recovery time scales of the system. We express these time scales as nonlinear expressions of the microscopic, mechanical parameters. Any adaptive system can be easily characterized by its value of τ from simple measurements. We discuss the advantages and limitations of this approach. We also consider more profound implications for modeling and understanding complexity in biology and how it relates to similar questions in the physical sciences.
Technical details of the Manifold Boundary Approximation Method (MBAM) are outlined in the materials and methods section. Briefly, the method assumes a parameterized model that makes predictions for a specific set of experimental conditions, known as Quantities of Interest (QoIs). Generally, the QoIs will be a subset of all the possible predictions that a model could make. Using information theory and computational differential geometry, the MBAM makes a series of approximations that remove the parameters from the model that would have been least identifiable if the experiments corresponding to the QoIs were actually performed. The refinements to the model take the form of limiting approximations. For example, the equilibrium and quasi-steady state approximations familiar to Michaelis-Menten reactions are a special case as we now show.
Many biological reactions take the form of an enzyme catalyzed reaction in which an enzyme and a substrate combine reversibly to form an intermediate complex which can then disassociate as the enzyme and a product: E + S ⇌ C → E + P. These reactions can be modeled using the law of mass action as:
d d t [ E ] = - k f [ E ] [ S ] + k r [ C ] + k c [ C ] (1) d d t [ S ] = - k f [ E ] [ S ] + k r [ C ] (2) d d t [ C ] = k f [ E ] [ S ] - k r [ C ] - k c [ C ] (3) d d t [ P ] = k c [ C ] . (4)
These equations have two conservation laws
E 0 = [ E ] + [ C ] (5) S 0 = [ S ] + [ C ] + [ P ] , (6)
so that the system in Eqs (1)–(4) has only two independent differential equations. We take the initial conditions of the enzyme and substrate to be E0 and S0 respectively and those of the intermediate complex and final product to be zero.
Consider the scenario in which E0 and S0 are fixed to 0.25 and 1 respectively and the three rate constants kf, kr, and kc are allowed to vary. In Fig 1 (top) we see many of the possible time series for the fractional concentration of the final product. If we take as QoIs, the fractional concentration of product at times t = 5, 10, 15, then Fig 1 (bottom) shows the corresponding model manifold. Because the model has three parameters, the model manifold is a three dimensional volume. The two colors (red and green) are two faces that enclose this volume and correspond to two possible reduced models that we consider shortly.
Notice that the model manifold, in this case a three-dimensional volume, is highly anisotropic. There is clearly a dominant, long axis, a second thinner axis, and a third axis that is much thinner still. MBAM exploits this low effective dimensionality in order to construct a model with an equivalent range of behavior with fewer parameters.
We now consider the phenomenon of adaptation. More specifically, we consider the problem of “adaptation to the mean of the signal” which is the ability of a system to reset itself after an initial response to a stimulus as illustrated in Fig 5 [45]. Throughout this work, we follow the problem statement in reference [40]: A system is given a step-function stimulus at time t = 0 and the response is observed.
In this section we consider two minimal topologies exhibiting adaptation due to Ma et al. [40]. We then consider a more complete mechanistic description of EGFR signaling [7], a real system known to exhibit adaptation. We will identify the EGFR pathway as being equivalent to one of the two minimal adaptive topologies. Finally, we will show that each of these adaptive systems can be represented by a single parameter model.
We note that it is possible to choose inputs other than a single step function. In fact, different adaptive systems are known to respond differently to different types of inputs [46, 47]. We here restrict ourselves to single step inputs as those the conditions described in references [7, 40] and because it is the most natural context for defining adaptation. If responses to other inputs are biologically relevant and controlled by different microscopic parameters, other choices for QoIs could be considered.
We now consider a model of EGFR signaling due to Brown et al. [7] that has been used extensively as a prototypical “sloppy model” for purposes of sensitivity analysis [6, 7, 9] and experimental design [21, 23]. The model describes the system response to two external stimuli, extra-cellular EGF and NGF hormones. The differing responses to these stimuli ultimately determine the differentiated cell type. The authors applied the MBAM to this model in reference [26] where the quantities of interest were taken to be the experimental conditions of the original analysis. From the original 48 parameter model, a 12 parameter model was constructed that could fit all of the data in the original experiments.
In the current context the model is interesting because the level of ERK activity (the final protein in the signaling cascade) exhibits adaptation behavior in response to EGF stimulus but long-term sustained ERK activity in response to NGF. We therefore seek a hybrid mechanistic/phenomenological description of this dual response. This requires a different set of QoIs from those in reference [26]. We here consider how the reduced model varies as the quantities of interest change. We will see that by systematically coarsening the QoIs, we can bridge the mechanistic and phenomenological descriptions of the system and gain a deeper understanding for the relationship between the structure of the model’s components and the resulting phenomenology.
Specifically, we consider the effect of four successive coarsening of the QoIs. First, we preserve the predictions of all species in the model under the same experimental conditions as reference [7] and deduce an 18 parameter model. Next, we consider only those species experimentally observed in reference [7], in which case we recover the 12 parameter model of reference [26]. Third, we consider only the response of ERK activity to EGF and NGF stimulus, reducing the model further to 6 parameters. Finally, we consider only the response of ERK to an EGF stimulus and recover a four parameter model exhibiting a minimal negative feedback loop topology characterizing the system’s adaptation and spanning the same phenomenological degrees of freedom in Fig 5.
Fig 9 shows the FIM eigenvalues for the entire reduction process. The initial reduction process from 48 to 18 parameters is summarized in Fig 9 (top left). The initial 48 parameter model exhibits the characteristic “sloppy model” eigenvalue spectrum in which the eigenvalues are logarithmically spaced over many orders of magnitude [6–9]. Observe that each iteration of MBAM removes the smallest FIM eigenvalue from the model while the remaining eigenvalues are approximately unchanged. Thus, the resulting approximate model is not sloppy; the eigenvalues cover fewer than four orders of magnitude. At this point the remaining parameter combinations are precisely those phenomenological parameters necessary to explain the important features of the QoIs; further reductions would sacrifice statistically significant model flexibility.
We can also consider the effect of the reduction process on the model’s network structure as summarized in Fig 10. Observe the condensation of the network between the top left and right panels in Fig 10. Many of the nodes in the network exhibit similar behavior; the reduction naturally clusters these nodes and highlights the emergent, effective topology governing the system.
Using this 18 parameter model as a starting point, we next coarsen the QoIs by ignoring those species for which experimental data was not available in reference [7]. The remaining observed species are Ras, Raf1, Rap1, B-Raf, Mek1/2, and Erk1/2. The eigenvalues of the 18 parameter model in top right panel of Fig 9 therefore correspond to the same parameters as those in the top left of the same Figure. This is the eigenvalue spectrum that would have resulted if the 18 parameter model had been fit to the original data. Notice that three eigenvalues are now zero (numerical zero ∼10−16). These correspond to the three remaining parameter of the EGF/PI3K/Akt cascade for which there were no observations in reference [7]. The data allow no predictions for these unobserved species.
Two other eigenvalues are dramatically smaller after coarsening the QoIs (λ ∼ 10 - 4). One parameter corresponds to the relative activity level of P90/Rsk (exactly analogous to the limit leading to Eq (27)). The other parameter is the unbinding rate of NGF from NGFR. The dramatic decrease in these eigenvalues upon coarsening the QoIs indicate that these QoIs contain practically no information about these parameters. These parameters are therefore irrelevant for explaining the system behavior. Additionally, one other parameter can be removed which lumps MEK and ERK as a single dynamical variable. These approximations are further reflected in the condensed network (Fig 10 bottom center). Model predictions that depend strongly on these parameters could not be constrained by the original data.
The activity level of ERK is the quantity of primary biological interest in this model as it signals to the nucleus the presence of extra-cellular EGF or NGF and ultimately determines cell fate. Therefore, we next consider only the level of ERK activity in response to EGF and NGF stimuli (Fig 9 bottom left and Fig 10 bottom center). These QoIs can be explained by a six parameter model. Of these six parameters, two are associated with the C3G cascade which is only activated by NGF stimulation. Coarsening the QoIs to only include an EGF stimulus therefore reduces the model to four parameters (Fig 9 bottom right) and a minimal negative feedback loop (Fig 10 bottom right) analogous to that in Fig 6 (left).
In Fig 11 we illustrate the sensitivities of the ERK adaptation curve to each of the four coarse-grained parameters. The sensitivities of parameters 1 and 4 are very similar in that they both increase the over-all level of ERK activity through the time series. Unlike parameter 4, parameter 1 is also characterized by a narrowing of the response peak.
It is interesting to compare these sensitivities with those in Fig 7. Parameters 2 in both models have the same functional effect, controlling the turnover point for the adaptation. Similarly, parameters 4 in both models control the over scale of the time series.
In contrast, parameters 1 and 3 in the minimal EGFR model have a different functional role from parameters 1 and 3 in the simple negative feedback loop above. However, by tuning an appropriate combination of parameters 1 and 3 in the minimal EGFR model, it is possible to control only the final steady state of the model without affecting the transient peak, directly analogous to parameter 3 in Fig 7. Likewise, another combination can be chosen to be functionally equivalent to parameter 1 in Fig 7. Although the mechanism by which these degrees of freedom are controlled are different in the two models, they ultimately span the same four degrees of freedom summarized in Fig 5.
We have seen that all three adaptation models can be simplified to four phenomenological parameters. These four parameters span the same four degrees of freedom illustrated in Fig 5. The four parameter models can fit artificial adaptation data generated from the full models, and the systematic errors due to approximations in the model are indistinguishable from the artificial noise. However, removing more parameters results in statistically significant errors when the models are fit to data. That is, further simplifications result in observable systematic errors. However, it is possible to remove additional parameters and still preserve the qualitative behavior of the system. For example, by increasing error bars for the QoIs, additional parameters can be removed. The resulting models still exhibit adaptation, but are unable to fit the exact curvature of the true model’s time series.
In general applications, the level of granularity in the final model will be driven by many factors, and it may be preferable to consider several models of varying levels of complexity. We illustrate this for the adaptation models considered above. In all three cases, the qualitative adaptation behavior can be approximated by models with two parameters. Although these minimal models are not quantitatively accurate they provide insight into the governing mechanisms.
The equations governing the two parameter negative feedback model are
d d t [ C ] = k A C I Θ ( 1 - C ) - B ˜ [ C ] (37) d d t B ˜ = k C B k B C K C B K B C [ C ] . (38)
Those governing the two parameter incoherent feed forward loop model are
d d t [ C ] = k A C I Θ ( 1 - C ) - B ˜ [ C ] (39) d d t B ˜ = k A B k B C K B C I . (40)
In both cases B ˜ = [ B ] ( k B C / K B C ). The only difference between these two adaptation mechanisms is how in the stimulus information is transmitted to the buffer node, either indirectly through the adaptive node C in the case of negative feedback, or directly from the input in the case of feed forward.
In both models, one parameter defines the time unit of the system. In particular, the models are invariant to the transformation t → αt, kAC → kAC/α, kCB → kCB/α, kBC → kBC/α, kAB → kAB/α. By choosing units in which kAC = 1, i.e., the initial slope of the rising portion of the curve, the models are reduced to a single parameter. The lone remaining parameter controls the time scale for recovery from the initial inputs. Adaptation can therefore be universally characterized by the dimensionless ratio τ of these two scales:
τ N F B L B = k A C K C B K B C k C B k B C (41) τ I F F L P = k A C K B C k A B k B C . (42)
The time series for various values of τ are given in Fig 12 for both mechanisms. While the curves are similar, notice that negative feedback loop generally achieves better sensitivity, i.e., height of the peak in response to the input. The incoherent feed forward loop, in contrast, achieves better precision (i.e., final steady state closer to zero) after the initial transient has faded. Fig 13 shows the time to achieve a maximal response and the value of the maximal response for various values of τ for the two mechanisms.
In going from the four phenomenological parameters in Fig 5 to the single parameter τ, the models have lost some flexibility. It is important to remember that the sensitivities in Figs 7, 8 and 11 are based on a local analysis. An actual adaptive system can vary its parameters to make small adjustments to all four phenomenological degrees freedom. However, the primary adaptation response is characterized by the value of τ as in Fig 12. Notice that the phenomenological interpretation of τ does not correspond directly to any one of the four phenomenological parameters in Fig 5. From Fig 12 we see that increasing τ corresponds to an increase in parameters ϕ1, …, ϕ4. This correlation is common to both mechanisms and indicates a universality in the types of adaptation curves that can be constructed in nature. There will be small small variations from these universal curves from system to system that represent fine-tuning of less important parameter combinations.
The equations governing the two parameter EGFR model are
d d t [ Erk ] = θ 1 [ EGF ] - P 90 ˜ [ Erk ] (43) d d t P 90 ˜ = θ 2 [ Erk ] , (44)
which are identical to those governing the negative feedback loop. The phenomenological parameters have expressions in terms of the structural parameters:
θ1=(KmRasGapKmdErkKmdMekKmdRaf1KEGFKRasToRaf1KSoskpMekCytoplasmickpRaf1KmEGFKmpMekCytoplasmicKmpRaf1kRasGapkdErkkdMekkdRaf)×(Mek SosPP2A2Raf1PPtase RasGap) (45) θ2=(kdSoskpP90RskKmdSosKmpP90Rsk)(P90Rsk) (46) P90˜=KKmpP90RskkpP90RskP90Rsk, (47)
with values θ1 ≈ 1.558 and θ2 ≈ 0.977. The dimensionless parameter characterizing the EGFR system for the rat model from reference [7] is therefore τEGF ≈ 1.6.
The control mechanisms underlying adaptation in both the negative feedback and incoherent feed-forward loops has been discussed extensively in the literature, particularly in reference [40]. It is therefore interesting and instructive to consider these analyses in light of the minimal models derived above.
First, consider the steady state values for the four-parameter negative feedback loop in Eqs (28)–(30):
A * = 1 (48) B * = K F B k A C k C B k B C F B k F B K C B K B C (49) C * = F B k F B k A C K C B K B C K F B k C B k B C . (50)
Of particular interest is the case of “perfect adaptation” in which node C returns very nearly to its pre-input value (zero in this case). Precision refers to the discrepancy between the final steady state of node C and the its pre-input value. Eq (50) identifies a combination of parameters that control this system behavior. Note, that one way to accomplish this is for the parameter KCB to become very small, consistent with one of the findings of reference [40].
At first, this result appears to contradict the limit (kCB, KCB) →∞ was used in deriving the equations for the negative feedback loop. However, this limit should not be interpreted to mean that kCB and KCB are really large in the full model. Rather, it means that the model predictions do not require these parameters to be finite so long as the ratio kCB/KCB has the appropriate value. In a real system KCB will certainly be finite and decreasing its value will affect the the system behavior. The effect decreasing KCB has on the outputs of the full model is preserved in the reduced system through the ratio kCB/KCB.
Eq (50) also predicts that large values of KFB are preferable for improved precision. Interestingly, reference [40] found that KFB was often small. These results are not necessarily in contradiction. Eq (50) allows for high precision with small KFB provided other parameter compensate accordingly. Reference [40] reports on a global search over all parameter space, i.e., allowing other parameter values to float as well. However, holding all other parameters fixed, precision can be improved by increasing KFB, a result that we confirm numerically.
In reference [40], the mechanism of the incoherent feed-forward loop was explained as an “anticipation” by directly monitoring the input node A. This was confirmed by demonstrating a proportionality between the steady state values of node A and node B so that “Node B will negatively regulate C in proportion to the degree of pathway input” [40]. This result can be seen readily in the reduced model in Eqs (34)–(36) for the entire dynamics. Assuming a constant input (as we have done), the equations for A and B can be integrated exactly to give (for times before saturation)
A = k I A I t (51) B = ( 1 / 2 ) k A B k I A I t 2 = ( 1 / 2 ) k A B t A . (52)
Both the negative feedback and incoherent feed-forward loops share a more general integral control mechanism. For the simple three node models, the topology of these networks is preserved by the reduction process so that previous analyses specific to the topology still apply to the simplified models [40]. In many cases of practical importance, however, the relevant control mechanism is embedded in a large network with many more than three nodes that has many potential control mechanisms. Consider, for example, the full network of in Fig 10 (top left) that contains both extended negative feedback and incoherent feed-forward loops as well as many other interconnections. In such a case, it is desirable to condense the network into a minimal mechanistic model in order to identify the relevant control mechanism. This is what is done by the MBAM. Strikingly, this relatively complicated network was reduced to exactly the same functional form as minimal negative feedback topology.
We have presented the Manifold Boundary Approximation Method specialized to the context of differential equation models of biochemical kinetics. We have shown that MBAM is capable of deriving simple phenomenological models of system behavior directly from a microscopic, mechanistic description. Because it was derived directly from the microscopic, the resulting simplified model is not a black box but provides real insights into how the microscopic mechanisms govern the emergent system behavior.
MBAM connects the microscopic to the macroscopic through a series of limiting approximations that are automatically identified and rigorously justified in a specific context defined by the Quantities of Interest (QoI). The parameters of the reduced model are therefore given as (often nonlinear) expressions of microscopic parameters that are exactly the identifiable combinations relative to the specific QoIs. It therefore becomes possible to identify how microscopic perturbations, such as gene mutations, over-expression, or knockout, will alter the macroscopic phenomenological parameters.
Selecting appropriate QoIs is an important component of the MBAM; however, the results are usually robust to many changes in the QoIs. The question of how the MBAM results are dependent on the QoIs has begun to be explored in reference [16]. Changing the QoIs will change the Fisher Information and by extension the geometric properties of the manifold. First, consider changes to the QoIs such as changing which time points are considered or the time dependence of the inputs. These changes effectively “stretch” or “compress” portions of the manifold, i.e., transform the model in a differentiable way–transformations known as diffeomorphisms. Because the boundaries of the model manifold are singularities of the FIM, the relationship among the boundaries are invariant to these diffeomorphisms. In other words, the boundaries are a feature of the differential topology of the family of manifolds generated by varying the QoIs. MBAM is therefore robust to changes in the QoIs because it is identifying a topologically invariant feature of the parameter space. MBAM uses geometric operations (e.g., geodesics) find these topological invariants, so that the QoIs are incidental to the process, but the details of the QoIs are not critical to the final result.
More drastic changes to the QoIs, such as changing which chemical species are observed, are not necessarily differentiable changes to the model manifold. Indeed, we have seen for the case of the Brown et al. model, that observing fewer species had a dramatic effect on the final reduced model as summarized in Fig 10. Other cases are considered in reference [16] where it is observed that changing the QoIs can lead to folding/unfolding of the manifold or even a “manifold collapse” along some dimensions. By systematically coarsening the QoIs, we have seen how the microscopic mechanism can be connected to the simple effective description.
In many cases it may not be obvious which QoIs should be chosen. Drastically different choices in QoIs will lead to different reduced models. While MBAM cannot say which choice is correct, it does provide way to systematically study the implications of different choices and generate testable hypotheses about how some intermediate behaviors may or may not influence larger-scale phenomena such as phenotype.
MBAM requires a model that is a more-or-less complete microscopic description as a starting point. Of course, any real model is never complete in the reductionist sense. However, microscopic models that can be used effectively with MBAM have made approximations that do not affect the important dynamics of the system. For example, the Brown et al. model is already a dramatic simplification over a comprehensive pathway map [5]. In many cases, however, little to nothing is known about the microscopic mechanisms. Although beyond the scope of this paper, we speculate that MBAM could be used to reverse engineer mechanisms when the microscopic model is unknown.
It is instructive to compare the MBAM with another common approach to parameter identification in complex biological models. Many parameter values are often fixed based on educated guesses found for example from in-vitro experiments. The small number of remaining parameters are fit to data. If there are only a few effective degrees of freedom in the model, this procedure will succeed if the remaining parameters have components along the stiff direction of the complete model. While this procedure will reduce the number of fitting parameters in the model, the model is not made conceptually simpler. Furthermore, it is difficult to know a priori how many or which parameters to fix and which to fit. After fixing several parameters, the remaining degrees of freedom in the model are generally misaligned with the true long axes of the model manifold. The restricted model will therefore not encompass the full range of possible model behavior of the original model. In other words, this procedure gives a local approximate model. For different regimes in the model’s parameter space, it will be necessary to fix a different set of parameters.
In contrast, the MBAM is a semi-global approximation scheme. Boundaries are a global, topological feature of a manifold [16]. By construction, the parameters of an MBAM simplified model are aligned with the true principal axes of the original model manifold and naturally follow its curvature. The MBAM approximation will generally be valid over a much broader range of the original parameter space than a model in which a handful of parameters are fixed. Furthermore, the boundaries represent structurally simplified approximate models that lead to conceptual insights about collective behavior while retaining an explicit connection to the microscopic mechanisms.
The key insight that enables this semi-global approximation scheme is an empirically observed correlation between local information, i.e., the eigenvalues of the FIM, and the global structure of the manifold, i.e., manifold widths [48, 49]. This observation allows the geodesic to find a path to the nearest model boundary using the eigenvalues of the FIM calculated at some point in the interior. In order for this to work, it is generally necessary for the parameters to be dimensionless and in the natural units of the QoIs. This is the reason we recommend using log-transformed parameters (see Materials and Methods section).
In our experience, the procedure of identifying limits from a single geodesic generally works; however, it is not fool-proof. On some occasions, the geodesic may encounter a region of high-curvature and bend away from the desired boundary and become lost–analogous to a spaceship experiencing a gravitational slingshot around a planet. In these cases, it will be necessary to guide the method by hand. In our experience, calculating a few geodesics starting from either nearby points or oriented along different directions in the sloppiest subspace (i.e. two or three eigendirections with smallest eigenvalues) will eventually identify the desired limit. For most models, the curvature has been demonstrated to be small, so this is a rare occurrence. We encountered it twice in our reduction of the EGFR model and once in our reduction of the adaptation models.
Because MBAM is a nonlinear approximation, it is involves considerably more computational expense than other local approximations. Fortunately, as mentioned above, the correspondence between FIM eigenvalues, manifold widths, and the existence of boundaries greatly reduces the computational cost associated with finding a semi-global approximation. Here, we have applied the method to a model of 48 parameters and 15 independent differential equations. However, we estimate that the method could be reasonably applied to models with several hundred parameters given standard simulation methods common in systems biology.
“Phenomenological” and “mechanistic” are two adjectives often used to describe models as well as general modeling philosophies. These two approaches reflect a dichotomy that pervades nearly all scientific disciplines between top-down, phenomenological models and bottom-up, mechanistic models [12, 50–55]. Both approaches have relative strengths and weaknesses. Phenomenological models reflect the relative simplicity of the collective behavior, automatically including the appropriate number of parameters to avoid over-fitting but lacking mechanistic explanations. Phenomenological models exploit correlations among observed data to make predictions about statistically similar experiments. In contrast mechanistic models are constructed to reflect causal relationships among components. These models are often complex and consequently susceptible to over-explaining behavior or over-fitting data. Because they model causal relationships, mechanistic models have a type of a priori information about the system behavior. Mechanistic descriptions are therefore an important ingredient for enabling new engineering and control applications that directly manipulate microscopic components.
A precise delineation between “mechanistic” and “phenomenological” modeling is difficult to define. Here, we take the difference between phenomenological and mechanistic models to be the model interpretation with respect to physical reality (in the reductionist sense). For example, the EGFR model summarized in Fig 10 (top left) is mechanistic because the modeler claims that there really is a biochemical agent known as Ras, for example, that really does respond to mSos and really does influence Raf1 and PI3K. In contrast, consider the phenomenological models derived from time series data by Daniels and Nemenman [54, 55]. In this case, the S-systems that make up the model components are not claimed to correspond to any real microscopic components.
The models derived in this work have properties of both phenomenological and mechanistic models. The original EGFR model of Brown et al. is mechanistic, but what about the minimal, condensed, negative feedback loop of Fig 10 (bottom right)? We claim that this mechanism reflects the reality of the collective biological system. Similarly, we interpret the components of this minimal model as representing real biological components.
In some sense, the parameter τ is phenomenological; it can be easily determined from experimental data without regard to microscopic mechanisms. However, because the expression for τ was derived incrementally from a mechanistic description, expressions such as Eqs (41)–(42) and Eqs (45)–(47) explicitly identify the mechanisms that control its value. In principle it would be possible to use these expressions to predict the value of τ from the microscopic parameter values. This is an important conceptual advance because it bridges the high-level phenomenological description and the low-level mechanisms. Indeed, these expressions identify which information about the microscopic components are necessary to predict a macroscopic behavior or conversely, which information about microscopic mechanisms can be inferred from systems-level observations.
Expressions directly relating microscopic and phenomenological parameters allows one to easily predict the effect on phenomenology (i.e., τ) in response to changes in any of the microscopic parameters (such as gene-knockout, over-expression, etc.) without the need to directly explore the large microscopic parameter space. Compressing parameter space in this way reduces the potential for over-fitting and over-explaining system behavior and significantly simplifies the ensuing statistical analysis.
In many cases of interest, mechanistic explanations are elusive. Although we have not explored the possibility here, we believe the current approach may be useful in these situations as well. For example, given several candidate mechanistic models, understanding how each mechanistic would hypothetically explain a system-level behavior could be useful in motivating experiments to distinguish among competing hypotheses by providing insights into competing theories.
Complexity in biological modeling is often contrasted with the apparent simplicity of models from the physical sciences. Indeed, many of the seminal examples of physics models are surprisingly simple and have very few parameters. Consider for example, the diffusion equation that is typically characterized by a single parameter [11]. Furthermore, the forms for many of the simple, phenomenological models of physics were guessed long before the microscopic mechanisms were understood. In contrast, the immense complexity of biological models often give rise to arguments that biology demands a new approach to mathematical modeling and that analogies drawn from physics are not likely to be useful for guiding computational biology. In many cases the justification for simple models in physics can be traced to either a small parameter or the symmetries of the underlying physical interactions. That these symmetries are not present in living systems gives credence to this perspective.
Despite the complexity of the underlying mechanisms, biological systems, like physical systems, often exhibit relativity simple collective behavior, especially when only a few QoIs are considered at a time. Adaptation, for example, is a common biological function that, as we have seen, could be modeled by a simple function with just one parameter. This situation is not unlike the diffusion equation from physics. In both cases, a simple macroscopic form can be expressed, independent of the microscopic details, with a few parameters that are easily inferable from data.
The stability of macroscopic behaviors to microscopic perturbations leads to the concept of a universality. Universality has been used with great success in physics by mapping the behavior of many different systems into a relatively small number of universality classes. Once the appropriate universality class has been identified, a simple, computationally tractable model can be used to calculate all universal physical quantities. For example, the critical exponents of many different fluids can be predicted almost exactly by the Ising model, a toy model of ferromagnetism. It does not matter that the Ising model is not a mechanistically accurate model of fluids because it is in the same universality class. There has been considerable speculation about the extent to which universality may or may not prove useful in biology or other complex systems. Here we consider one such argument that is particularly relevant in the context of the manifold boundary approximation.
One source of complexity in biology arises when attempting to predict how the simple collective behavior will be altered by microscopic perturbations, such as mutating genes or applying protein-targeting drugs, or how a desired collective behavior could be engineered from microscopic components. Indeed, this is a much more challenging question that is not easily answered by phenomenological models without mechanistic information. However, this problem is not unique to biology. In physics, for example, the Ising model does not predict the critical temperature and pressure of a fluid, only the properties at the critical point. Similarly, macroscopic, phenomenological models of material strength do not give any insights into how to engineer stronger alloys. Phenomenological models have limited predictive power for experiments that manipulate microscopic control knobs. As experimental and engineering efforts in physics, biology, and other scientific fields have advanced to the realm of the microscopic, these simple macroscopic theories need to be explicitly connected to their microscopic mechanisms. How does one systematically identify the microscopic parameter combinations that control the non-universal behavior of a system?
It is true that the types of questions advanced by both modern physics and biology demand new approaches to modeling beyond what has been “unreasonably successful” historically in the physical sciences [56]. Indeed, the challenges faced by biological and physical modeling are shared by many disciplines across the sciences. How do microscopic mechanisms govern collective behavior and how can that behavior be controlled and engineered? Simple, phenomenological models can play an important role in answering these questions since they distill the essence of the system behavior. What is often missing, however, is an explicit connection between the phenomenology and the mechanistic description. The manifold boundary approximation method is a step toward providing such a bridge in a general way. It is our hope that similar analysis can lead to a likewise comprehensive picture of other complex processes both in physics, biology, and elsewhere.
The Manifold Boundary Approximation Method (MBAM) is a model reduction scheme described in reference [26]. As the name suggests, it is based on a geometric interpretation of information theory (known as information geometry [48, 49, 57–61]) that is applicable to a wide range of model types. In this section we give a more algorithmic description and presentation specialized to the types of models common in systems biology, i.e., those that are formulated as differential equations of chemical kinetics that would be fit to data by least squares. Notably, this excludes stochastic differential equations. In principle, the MBAM formalism can be applied to SDEs, but we do not address that question here. Throughout this section, we refer to the relevant information geometric objects (manifold, metric, geodesics, etc.) and provide external references for completeness. However, the reader can ignore these technicalities if desired and implement the method as summarized here.
We assume the existence of a model of a biological system with many parameters θ that can be evaluated to make predictions. Examples of possible predictions include the concentrations of specific chemical species at specific times in response to specific stimuli. Approximations inherently disregard pieces of the model, so it is necessary to decide the objective of the model, i.e., which model behaviors the approximation should preserve. Therefore, from the many possible predictions, the modeler selects a subset that we refer to as Quantities of Interest (QoI). We denote these by
r m ( θ ) = y m ( θ ) σ m (53)
where m is an index that enumerates the QoIs, ym(θ) denotes the prediction of the model for the corresponding QoI evaluated at parameters θ, and σm represents the tolerance with which the QoI should be preserved. The QoI is analogous to a data point ym(θ) with experimental uncertainty σm.
In practice, the QoIs will often include predictions for which experimental data is available. The data will then be used to calibrate the reduced model. However, QoIs may also include predictions for which data is unavailable but for which the modeler would nevertheless like to make predictions. Alternatively, QoIs may include a very small subset of possible predictions as we have done here for the case of EGFR signaling.
The underlying idea of the MBAM is that rm(θ) can be interpreted as a vector in ℝM, where M is the number of QoIs. If the model contains N parameters, then this vector sweeps out an N-dimensional hyper-surface embedded in ℝM. This hyper-surface is known as the model manifold and denoted by M. For biological systems such as we consider here (in addition to models from many other fields), the model manifold is bounded. Furthermore, the model manifold has many cross sections that are very thin. Consequently, M often has an effective dimensionality that is much less than N. Our goal is to construct a low dimensional approximation to the model manifold by finding the boundaries of M. The procedure for doing this can be summarized as a four step algorithm.
First, from an estimate of the parameters θ0 calculate the matrix
g μ ν = ∑ m ∂ r m ∂ θ μ ∂ r m ∂ θ ν . (54)
This matrix is the Fisher Information Matrix (FIM) of the model and corresponds to the Riemannian metric on M. Calculating the eigenvalues of this matrix reveal the “sloppiness” of the corresponding parameter inference problem. The eigenvectors with small eigenvalues correspond to the parameter combinations that have negligible effect on the QoIs. We denote the direction of the smallest eigenvector by v0.
The second step is to calculate a parameterized path through parameter space θ(τ) corresponding to the geodesic originating with parameters θ0 and direction v0. This is found by numerical solving a differential equation:
d 2 d τ 2 θ μ = ∑ ν , m ( g - 1 ) μ ν ∂ r m ∂ θ ν A ( v ) m (55)
where A(v) is the directional second derivative:
A m ( v ) = ∑ μ ν d θ μ d τ d θ ν d τ ∂ 2 r m ∂ θ μ ∂ θ ν . (56)
(As an aside, in order to avoid unnecessary complications for the uninitiated, we have not used many of the standard differential geometric conventions, including the Einstein summation convention or the use of raised and lowered indices to denote contravariant and covariant vector components.) It is possible to estimate Am(v) efficiently using finite differences
A m ( v ) ≈ r ( θ + h v ) + r ( θ - h v ) - 2 r ( θ ) h 2 , (57)
where v = d θ d τ.
The solution to Eq (55) is a parameterized curve through the parameter space. Along this curve, the modeler monitors the eigenvalues of the FIM (Eq (54)). A boundary of the model manifold is identified by the smallest eigenvalue of gμν approaching zero. When the smallest eigenvalue becomes much less than the next smallest, then the corresponding direction will reveal a limiting approximation in the model. This leads to step three.
The approximation will typically correspond to one or more parameters approaching zero or infinity in a coordinated way. The goal is to identify this limit and analytically evaluate it in the model. This is done explicitly for several models in this manuscript. The result of the process is a new model with one less parameter than the previous. We denote the new vector of parameters by ϕ and the QoIs for this approximate model by y ˜ m ( ϕ ) / σ m
Finally, the values of the parameters ϕ in the approximate model are calibrated to the parameters θ0 by minimizing the sum of square distance between
min ϕ ∑ m y m ( θ 0 ) - y ˜ m ( ϕ ) σ m 2 . (58)
This four-step procedure is iterated, removing one parameter at a time, until the model becomes sufficiently simple.
A python script that can be used for calculating geodesics is available on github [62].
The procedure just described requires a few comments, particularly as it applies to biological systems. First, the MBAM requires a parameter estimate as a starting point θ0, which usually cannot be estimated accurately. Although an accurate estimate of θ0 might be elusive, it has been shown that the resulting reduced model is largely independent to these uncertainties. Indeed, one purpose of the MBAM is to remove the unconstrained parameters from the model. The reason for this is seen by considering a geometric argument given in reference [26]. Huge variations in parameter values can result when fitting to data, but these variations all lie within the same statistical confidence region, which means they map to nearby points on the model manifold. Starting from any points within this confidence region will identify the same sequence of boundaries as the true parameters.
For most systems biology models, the microscopic parameters are restricted to positive values (reaction rates, Michaelis-Menten constants, Hill coefficients, and initial concentrations). In order to guarantee positivity, we assume that these parameters have been log-transformed in the model, i.e., θ = log k, where k are the reaction rates, etc. This serves the dual purpose of non-dimensionalizing the parameters, that is important for the initial eigendirection of the FIM to point to the narrowest width of the M. MBAM is a semi-global approximation method that is enabled by a correspondence between local information (FIM) and global structure (boundaries). This correspondence is less likely to hold if the parameters are not log-transformed.
We use the term semi-global to denote something between purely local and fully global. For the case of the enzyme-catalyzed reaction in Fig 1, the MBAM approximation is a global approximation; the Michaelis-Menten model is capable of well-approximating the full range of behavior of the mass-action kinetics. However, one could imagine, a more complicated model manifold with several narrow “arms” extending from a central location (something like a star). Beginning from a point in one of the arms of the manifold, the MBAM will likely only approximate the behavior along of the principal axis of that arm. Because of this possibility, we describe MBAM as semi-global. With the exception of the enzyme-substrate reaction (Fig 1), it is unknown whether the approximations given in this paper are global or semi-global. This is due to the intrinsic difficulties in both exploring and characterizing high-dimensional spaces.
|
10.1371/journal.pgen.1003565 | Genome-scale Analysis of Escherichia coli FNR Reveals Complex Features of Transcription Factor Binding | FNR is a well-studied global regulator of anaerobiosis, which is widely conserved across bacteria. Despite the importance of FNR and anaerobiosis in microbial lifestyles, the factors that influence its function on a genome-wide scale are poorly understood. Here, we report a functional genomic analysis of FNR action. We find that FNR occupancy at many target sites is strongly influenced by nucleoid-associated proteins (NAPs) that restrict access to many FNR binding sites. At a genome-wide level, only a subset of predicted FNR binding sites were bound under anaerobic fermentative conditions and many appeared to be masked by the NAPs H-NS, IHF and Fis. Similar assays in cells lacking H-NS and its paralog StpA showed increased FNR occupancy at sites bound by H-NS in WT strains, indicating that large regions of the genome are not readily accessible for FNR binding. Genome accessibility may also explain our finding that genome-wide FNR occupancy did not correlate with the match to consensus at binding sites, suggesting that significant variation in ChIP signal was attributable to cross-linking or immunoprecipitation efficiency rather than differences in binding affinities for FNR sites. Correlation of FNR ChIP-seq peaks with transcriptomic data showed that less than half of the FNR-regulated operons could be attributed to direct FNR binding. Conversely, FNR bound some promoters without regulating expression presumably requiring changes in activity of condition-specific transcription factors. Such combinatorial regulation may allow Escherichia coli to respond rapidly to environmental changes and confer an ecological advantage in the anaerobic but nutrient-fluctuating environment of the mammalian gut.
| Regulation of gene expression by transcription factors (TFs) is key to adaptation to environmental changes. Our comprehensive, genome-scale analysis of a prototypical global TF, the anaerobic regulator FNR from Escherichia coli, leads to several novel and unanticipated insights into the influences on FNR binding genome-wide and the complex structure of bacterial regulons. We found that binding of NAPs restricts FNR binding at a subset of sites, suggesting that the bacterial genome is not freely accessible for FNR binding. Our finding that less than half of the predicted FNR binding sites were occupied in vivo further challenges the utility of using bioinformatic searches alone to predict regulon structure, reinforcing the need for experimental determination of TF binding. By correlating the occupancy data with transcriptomic data, we confirm that FNR serves as a global signal of anaerobiosis but expression of some operons in the FNR regulon require other regulators sensitive to alternative environmental stimuli. Thus, FNR binding and regulation appear to depend on both the nucleoprotein structure of the chromosome and on combinatorial binding of FNR with other regulators. Both of these phenomena are typical of TF binding in eukaryotes; our results establish that they are also features of bacterial TF binding.
| Regulation of transcription initiation by transcription factors (TFs) is a key step in controlling gene expression in all domains of life. Genome-wide studies are revealing important features of the complexity of transcription regulation in cells not always apparent from in vitro studies. In eukaryotes, both the inhibition of TF binding by chromatin structure and the combinatorial action of multiple TFs contribute to the genome-wide pattern of TF binding and function [1]–[5]. In contrast, our knowledge of transcriptional regulation by bacterial TFs stems largely from elegant in vitro experiments that have provided atomic resolution views of TF function [6]. Much less is known about how chromosome structure and combinatorial action affect bacterial TF binding and transcriptional regulation on a genome-wide scale [7]. Previous studies have suggested that, in contrast to the chromatin-restricted TF binding in eukaryotes, the Escherichia coli genome is permissive to TF binding because the occupancy pattern for some TFs correlates well with match to consensus sequence and consequent binding affinity [8]–[10]. Other studies suggest that nucleoid-associated proteins (NAPs; for example H-NS, Hu, Fis, and IHF) organize the chromosome into discrete domains and structures that may affect transcriptional regulation [7], [11]–[13], but possible global effects of NAPs on TF-binding have not been systematically tested. To investigate the roles of TF action and chromosome structure in a prototypical bacterial regulon, we studied the regulon of the anaerobic TF FNR.
FNR is widely conserved throughout the bacterial domain, where it evolved to allow facultative anaerobes to adjust to O2 deprivation [14]. Under anaerobic conditions, E. coli FNR contains one [4Fe-4S] cluster per subunit, which promotes a conformation necessary for FNR dimerization, site-specific DNA binding, and transcription regulation [15], [16]. Genome-wide transcription profiling experiments [17]–[19] established that E. coli FNR controls expression of a large number of genes under anaerobic growth conditions, in particular those genes whose products function in anaerobic energy metabolism. However, corresponding studies to establish which promoters are directly or indirectly regulated by FNR under comparable growth conditions have yet to be reported.
Studies of the regulatory regions of a few FNR controlled promoters have provided key insights into the mechanism of transcriptional regulation by FNR and the characteristics of FNR binding sites [20], [21]. From these studies we know that FNR binding sites can have only a partial match to the consensus sequence of TTGATnnnnATCAA, and be located at variable positions within promoter regions, directing whether FNR has either a positive or negative affect on transcription. At FNR repressed promoters, FNR binding site locations range from upstream of the −35 hexamer (which binds region 4.2 of RNA polymerase σ70), to overlapping the transcription start site (TSS; +1). At most FNR activated promoters, the center of the binding site is ∼41.5 nt upstream of the TSS, placing FNR in position to interact with both the σ70 and α subunits of RNA polymerase (RNAP) [21], [22]. Very few promoters are known to have FNR binding sites centered at −61.5 or greater, a position dependent typically on interactions with only the α subunit of RNAP [21]. The predominance of FNR binding sites positioned at −41.5 nt may reflect a preference for a particular activation mechanism, but it also could reflect sample bias in the limited number of activated promoters that have been studied to date. Thus, current knowledge is insufficient to allow accurate prediction of FNR binding sites genome-wide.
Many FNR regulated promoters are controlled by multiple TFs (for example CRP, NarL, NarP, and NAPs [7], [20], [21]), which can have either positive or negative effects on FNR function depending on the promoter architecture. For example, the narG promoter is activated by FNR, IHF, and the nitrate-responsive regulator, NarL [23], [24]; in contrast, the dmsA promoter is activated by FNR, but repressed by NarL [25], [26]. At the nir promoter, NarL displaces IHF to overcome a repressive effect of IHF and Fis, and thereby enhances FNR-dependent transcription [27]. Thus, in the presence of the anaerobic electron acceptor nitrate, FNR function can be either enhanced or repressed by NarL depending on the organization of TF-binding sites within the promoter region. In this way, the requirement of additional TFs for combinatorial regulation of promoters bound by FNR resembles transcriptional regulation in eukaryotes [28]. Such complex regulatory patterns cannot currently be inferred simply by identifying the locations of TF binding sites or by the strength of the FNR binding site. Direct measure of occupancy at these sites by each TF and correlation with the resulting transcripts in different growth conditions is needed to understand how complex bacterial regulatory networks coordinate gene expression. As an important first step, Grainger et al. used chromatin immunoprecipitation followed by microarray hybridization (ChIP-chip) to examine FNR occupancy using a FLAG-tagged FNR protein in E. coli cultures grown anaerobically in a rich medium [29]. Although many new FNR binding sites were identified, these data were not obtained from cells grown in the growth media used for reported transcriptomic experiments [17]–[19] and thus the datasets cannot readily be compared.
To systematically investigate FNR binding genome-wide, we performed chromatin immunoprecipitation followed by microarray hybridization (ChIP-chip) and high-throughput sequencing (ChIP-seq) for WT FNR from E. coli grown anaerobically in a glucose minimal medium (GMM). Computational and bioinformatic analyses were used to refine a FNR position weight matrix (PWM). The PWM was used to determine the relationship between ChIP-seq/ChIP-chip enrichment and match to the PWM, and to identify predicted FNR binding sites not detected by ChIP-seq. To examine the subset of high-quality predicted FNR binding sites that lacked a FNR ChIP-seq peak, we obtained and analyzed aerobic and/or anaerobic ChIP-chip data for NAPs H-NS and IHF along with analysis of previously published aerobic ChIP-seq data for the NAP Fis [30] to determine if NAP occupancy might prevent FNR binding. Further, the effect of H-NS on FNR occupancy was examined directly using ChIP-chip analysis of FNR as well as on O2 dependent changes in expression in the absence of H-NS and its paralog StpA. After identifying FNR binding sites genome-wide, we performed whole genome transcription profiling experiments using expression microarrays and high-throughput RNA sequencing (RNA-seq) to compare a WT and Δfnr strain grown in the same medium used for the DNA binding studies. The transcriptional impact of FNR binding genome-wide was investigated by correlating the occupancy data with the transcriptomic data to determine which binding events led to changes in transcription, to identify the direct and indirect regulons of FNR, and to define categories of FNR regulatory mechanisms. Finally, the aerobic and anaerobic ChIP-chip and ChIP-seq distributions of the σ70 and ß subunits of RNAP throughout the genome were analyzed to determine the role of O2 and FNR regulation on RNAP occupancy and transcription.
TF binding sites were mapped genome-wide in E. coli K-12 MG1655 using ChIP-chip and/or ChIP-seq for FNR, σ70 and ß subunits of RNAP, H-NS, and IHF under aerobic or anaerobic growth conditions, as indicated (Figure 1). In addition, we analyzed a publically available Fis data set collected under aerobic conditions [30]. The ChIP-chip distribution of the ß subunit of RNAP suggested widespread transcription under both aerobic and anaerobic conditions, as expected, whereas the O2-dependent changes in ß occupancy indicated those genes that are differentially regulated by O2. Further, binding, and thus transcription, by the σ70 housekeeping form of E. coli RNAP was observed throughout the chromosome; peak finding algorithms identified a large number of anaerobic σ70 ChIP-seq peaks (2,106) and aerobic σ70 ChIP-seq peaks (2,446) (Table S1). About 700 of the σ70 peaks showed statistically significant O2-dependent changes in occupancy (Table S2). The O2-dependent differences in RNAP occupancy suggest extensive transcriptional reprogramming in response to changes in O2, providing an excellent model system for examining genome-scale changes in transcription.
Comparison of the profiles of other DNA binding proteins indicated that the number of binding sites for NAPs genome-wide was much greater than for the TF FNR. ChIP-seq and ChIP-chip analyses identified 207 FNR peaks, 722 anaerobic H-NS enriched regions, 782 aerobic H-NS enriched regions, 1,020 anaerobic IHF enriched regions (Tables S3, S4, and S5) and published analysis of Fis identified 1,464 aerobic enriched regions [30]. The unbiased distribution of H-NS and IHF throughout the chromosome supports previous genome-wide studies of these NAPs [30]–[33]. H-NS is known to form filaments that cover multiple kb of DNA [7], [12], [13], [30], [34], [35] and we observed that half of the identified aerobic (390) and anaerobic (356) H-NS enriched regions were over 1 kb in length, referred to as extended H-NS binding regions (Table S3). Comparison of the aerobic and anaerobic H-NS binding distributions suggests H-NS occupancy is not greatly affected by O2 (Figure 1). For FNR, the number of high-confidence ChIP peaks (207) identified (Table S5) was just a few fold lower than the number of genes found to show FNR-dependent changes in expression (between 300–700) [17]–[19]. These binding site data were used to determine features of FNR binding genome-wide.
A small number of FNR peaks showed a large degree of variation in peak height across the genome. Previous studies of the repressor LexA reported that ChIP-chip peak height correlated with the match to the consensus sequence [10], suggesting that differences in site occupancy may reflect relative binding affinities to individual sites. Because FNR is a global regulator with a more degenerate binding site than LexA, we tested whether we could use this parameter to gain additional information about FNR binding-site preferences. A PWM (Figure 2 Inset) was constructed from an alignment of sequences from the ChIP-seq peaks and the scores representing the match to the PWM were determined with the algorithm PatSer (Table S5) [36]. In contrast to the studies of LexA [10], we found a poor correlation between the height of the FNR ChIP-seq peak and the match to the FNR PWM for the site predicted within each peak (Figure 3A). The same lack of correlation was also observed with FNR ChIP-chip data, indicating that this was not specific to the detection method. Additionally, there was a lack of correlation between FNR peak height and the number of known FNR binding sites. Furthermore, the majority of the FNR ChIP-seq or ChIP-chip peaks had similar heights, regardless of the score of the FNR motif present (Figure 3A).
One explanation for this latter result is that most FNR binding sites were saturated for binding in vivo. To examine this possibility directly, we performed ChIP-chip experiments over a range of cellular FNR dimer concentrations below the normal anaerobic cellular level of ∼2.5 µM [37], controlled by varying IPTG levels in a strain with fnr fused to an IPTG-inducible promoter. Peak areas for 35 selected FNR sites, representing a distribution of peak heights, were quantified for several cellular FNR dimer concentrations (∼0.45, ∼0.7, ∼1.9, and ∼2.5 µM). These plots showed a typical binding saturation curve for both novel and previously identified FNR binding sites, and revealed that all sites examined were saturated for binding at the normal cellular FNR dimer level of ∼2.5 µM (Figure 3B, Figure S1). However, because the broad distribution of peak heights between different sites was still observed, despite the fact that the sites were maximally occupied, we concluded that variation in peak height was not related to strength of FNR binding (Figure 3, Figure S1). As a control, we tested four FNR peaks that were determined to be non-specific due to enrichment in a Δfnr control ChIP-chip experiment and these peaks showed no change in peak height when FNR levels were varied (Figure S1). Thus, we conclude that differences in peak height in the ChIP-seq and ChIP-chip experiments for FNR were most likely due to differences in cross-linking efficiency or immunoprecipitation at particular genomic locations and not to differences in FNR binding affinity.
A well-known challenge in genomic studies is the use of computational tools to accurately predict DNA binding sites, particularly for global regulators like FNR that have degenerate binding sites. To investigate the usefulness of the PWM generated from our set of ChIP binding sites for predicting FNR sites genome-wide, we initially used a PatSer [36] threshold low enough that a FNR motif was identified in each FNR ChIP-seq peak. However, this threshold resulted in >10,000 possible genomic FNR binding sites. In contrast, if we used a precision-recall (PR) curve [38] to determine the optimal threshold to predict FNR binding sites (ln(p-value) of −10.75), then we obtained a more reasonable number (187) of predicted FNR binding sites (Figure 2, Table S6). Surprisingly, fewer than half of these sites (63 of 187) corresponded with a FNR ChIP-seq peak (Table S6), despite the fact that some predicted sites without a corresponding ChIP-seq peak had higher quality PatSer scores than those with a ChIP-seq peak. Although it is possible that some of the predicted sites without a ChIP-seq peak contain flanking sequence elements that disfavor FNR binding, we considered the possibility that many are functional sites but either FNR binding was masked by other DNA binding proteins or FNR cross-linking failed for other reasons.
NAPs are known to affect the binding of some TFs in E. coli [7], [12]. To ask if the NAPs H-NS, IHF, or Fis might occlude the 124 predicted FNR binding sites lacking a FNR ChIP-seq peak, we analyzed ChIP-chip data for H-NS and IHF, obtained from the same growth conditions, and publicly available ChIP-seq data for Fis [30]. Nearly all of these FNR sites (111 of 124 sites; ∼90%; silent FNR sites) were enriched in IHF, H-NS, or Fis, consistent with the idea that these NAPs occupy the silent FNR sites and thereby block FNR binding (Table S6, Figure S2). Similar occupancy was observed when the 124 predicted FNR sites were compared with H-NS and IHF enrichment from published ChIP-chip and ChIP-seq data performed under different growth conditions [30], [31], [33]. In comparison, only ∼20% (14 of 63 sites) of the FNR sites that coincided with a FNR ChIP-seq peak were enriched in a NAP ChIP signal, significantly less than NAP occupancy at FNR sites lacking a peak (p-value<0.05). In contrast, we found ∼50% of the previously identified LexA binding sites [10] were co-occupied with H-NS. We conclude that the NAPs H-NS, IHF, or Fis likely prevent FNR binding at some sites by occlusion.
We also examined whether the silent FNR sites are preferentially occluded by the extended H-NS binding regions. The extended binding regions of H-NS (>1 kb) likely represent H-NS filaments that are known to cover multiple kb of DNA and silence transcription [7], [12], [13], [30], [34]. Consistent with this notion, our results showed that the extended H-NS binding regions were negatively correlated with RNAP (ß) ChIP-chip occupancy and this silencing occurred in both the presence and absence of O2 (p-value<0.05) (Figure S3). In contrast, shorter H-NS enriched regions (<1 kb) were both positively and negatively correlated with RNAP ChIP-chip occupancy under aerobic and anaerobic growth conditions. The 46 silent FNR sites bound by H-NS were more likely to be occupied by extended H-NS binding regions (42 sites) than by short H-NS binding regions (4 sites) (p-value<0.05; example in Figure S3C), suggesting that extended H-NS binding regions may inhibit FNR binding at silent FNR sites.
To investigate the impact of H-NS binding on FNR occupancy, we characterized FNR ChIP-chip peaks in a strain deleted for both hns and stpA; stpA encodes a H-NS paralog that partially compensates for H-NS in a Δhns mutant [39], [40]. Many new FNR peaks (196) appeared in the Δhns/ΔstpA strain (Figure 1, Figure 4A–C, Table S7), and a large fraction (81%; 158 FNR peaks) of these new peaks corresponded to H-NS binding regions in the WT strains, indicating that FNR binding was unmasked in the absence of H-NS and StpA. The distribution of the FNR PWM scores of the FNR sites found within the FNR ChIP-chip peaks unmasked by the absence of H-NS and StpA was similar to that found in the WT strain (Figure 4C, Tables S6 and S7). The majority (78 of 99) of silent FNR sites lacking FNR peaks in the Δhns/ΔstpA strain were enriched for IHF and/or Fis, suggesting that these NAPs still occluded FNR binding in the absence of H-NS and StpA (Table S6). Taken together, these results establish that removal of H-NS and StpA allowed FNR to bind to sites covered by H-NS in WT strains.
Nearly all FNR peaks found in the WT strain were retained in the Δhns/ΔstpA mutant (163 of 169 peaks; Figure 4A, Table S7), but a small proportion (∼15%) showed a significant increase in peak average (average log2(IP/INPUT) value of the binding region) in the Δhns/ΔstpA strain (Figure 4D). The majority of these FNR peaks with increased peak averages were also bound by H-NS in the WT strain, suggesting that removing H-NS allowed for increased cross-linking or immunoprecipitation of FNR at these loci likely due to changes in chromosomal structure in the absence of H-NS and StpA [35]. In contrast, removing H-NS did not affect FNR occupancy or cross-linking at locations lacking H-NS ChIP signal in WT strains. We conclude that H-NS reduces or blocks FNR binding at many locations in vivo.
To determine which FNR binding events from the WT strain caused a change in gene expression, the FNR occupancy data were correlated with the 122 operons differentially expressed (DE) by FNR (Table S8). Surprisingly, less than a half of the 122 operons were correlated with a FNR ChIP-seq peak while less than a fourth of the 207 FNR ChIP-seq peaks were correlated with a FNR-dependent change in expression (Figure S4). To address this unexpected result, we systematically analyzed the regulation of all of these operons by incorporating published data and classified the operons into seven regulatory categories (Figure 5). Category 1 (Table 1) contained operons that were directly activated by FNR because they showed a FNR-dependent increase in anaerobic transcript levels and a FNR ChIP-seq peak within 500 nt of the translation start site of the first gene of an operon. Category 2 (Table 1) contained operons that were directly repressed by FNR (showed a FNR-dependent decrease in expression and had a FNR ChIP-seq peak). Categories 3–5 contained a surprisingly large number of operons (156) with a FNR ChIP-seq peak within 500 nt of the translation start site of the first gene of an operon but no FNR-dependent change in expression. Previously published studies (23 operons) and our additional collation of other relevant TF-binding sites (52 operons) suggest that at least half (75) of these sites may be directly regulated by FNR under alternative growth conditions (Table S9). For example, Category 3 (Tables 2 and 3) contained operons known or proposed to be co-regulated by FNR and another TF under growth conditions not used in our study. Category 4 (Table 4) contained operons known to be repressed by another TF under our growth conditions. Category 5 (Table S9) contained operons with other potential regulatory mechanisms. Category 6 (Table 5, Table S10) contained operons that were indirectly regulated by FNR because no FNR ChIP-seq peak was found within 500 nt of the translation start site despite showing a FNR-dependent change in expression. Finally, Category 7 (Table S11) contained operons with a FNR peak identified only in the Δhns/ΔstpA strain, which also showed potential FNR regulation in the absence of H-NS and StpA.
The 32 operons directly activated by FNR (Table 1) contain some of the best-studied FNR regulated operons. In addition to operons associated with anaerobic respiration (dmsABC, frdABCD, nrfABCDEFG, narGHJI) [41]–[43], this category included glycolytic (pykA) and fermentative enzymes (pflB and ackA), which would be expected to promote mixed acid fermentation of glucose to ethanol, acetate, formate and succinate in the absence of an added electron acceptor (Figure 6), the conditions used in this study. As expected, we also found that these promoters showed an increase in σ70 occupancy, as illustrated by representative FNR and σ70 data for FNR activation of dmsABC (Figure 7), providing a proof-of-principle for our approach. While expression of many operons in this category was known to be FNR regulated, only about half had been shown to directly bind FNR (Table 1).
FNR also directly activated operons with functions that illustrate the broader role of FNR in anaerobic metabolism: pepE, a peptidase, suggesting peptide degradation in E. coli similar to that observed in Salmonella [44]; ynjE, an enzyme involved in biosynthesis of molybdopterin, a cofactor used by anaerobic respiratory enzymes [45]; pyrD, a dihydroorotate dehydrogenase in pyrimidine biosynthesis [46]; and ynfK, a predicted dethiobiotin synthetase and paralog of BioD of the biotin synthesis pathway. The activation of the biofilm TF bssR by FNR suggests a link between biofilm formation and anaerobiosis (Table 1). FNR directly activated the carnitine-sensing TF CaiF, confirming a link between FNR and carnitine metabolism [29], [47]. In addition, the FNR-enriched region found upstream of fnrS supports FNR direct transcription activation of this small regulatory RNA [48], [49], although the fnrS sRNA was not represented in our gene expression arrays and was too small to be detected by our RNA-seq protocol (Table 1).
To determine the position of FNR binding sites relative to the TSS, we used the FNR PWM (Figure 2 Inset) to search the FNR enriched regions using a PatSer score threshold low enough to identify FNR sites from every ChIP peak [36]. A majority (89%) of the FNR ChIP-seq peaks in the FNR direct regulon contained one FNR binding site (Table 1). Of the 23 promoters directly activated by FNR with a known TSS, 19 FNR sites were centered at −41.5 (±4 nt), the known position of a Class II site, while one site was centered at −60.5 (Class I site) (Table 1), supporting previous results suggesting a bias toward FNR binding Class II sites in activated promoters.
Analysis of the 21 operons directly repressed by FNR revealed both simple and complex repression mechanisms (Table 1). The majority of the operons directly repressed by FNR showed expression patterns similar to that of ndh, encoding the aerobic NADH dehydrogenase II, which showed a FNR-dependent decrease in expression and decrease in σ70 occupancy under anaerobic growth conditions (Figure 7). These operons included nrdAB, the aerobic ribonucleotide reductase; hisLGDC, a subset of the histidine biosynthesis enzymes; fbaB, the class I fructose-1,6-bisphosphate aldolase involved in gluconeogenesis; and can, the carbonic anhydrase. FNR also repressed iraP, which encodes the anti-adaptor protein that stabilizes σS, and rmf, which encodes the stationary phase inducible ribosome modulation factor.
In contrast, a subset of operons showed complex repression similar to cydAB, with an anaerobic dependent increase in expression despite the fact that anaerobic expression increased further in a strain lacking FNR, indicating partial repression (Table 1) [50]. Nearly all of these operons are also co-regulated by ArcA (Park and Kiley, Personal Communication) suggesting that, like cydAB, FNR and ArcA co-regulation could lead to maximal expression of these genes under microaerobic conditions [50]. These operons include hdeD, gadE and hdeAB-yhiD, involved in acid stress response, and ompC and ompW, encoding outer membrane proteins. The finding that strains lacking ompC, rmf, and rpoS show decreased viability compared to single or double mutants [51] suggests that these proteins may function in a common stress response, potentially necessary under microaerobic growth conditions. Interestingly, for the 16 promoters directly repressed by FNR with a known TSS, the FNR binding sites were broadly distributed, ranging from −125.5 to overlapping the +1 (Table 1). In sum, these results indicate the surprising finding that FNR directly represses a broad set of functions, including some stress responses, expanding the role of FNR beyond simply repressing genes associated with aerobic respiration.
Finally, comparison of the transcriptomic data to changes in σ70 holo-RNAP ChIP-seq occupancy under aerobic and anaerobic growth conditions revealed that nearly all FNR-regulated operons are expressed using σ70 RNAP. Increases or decreases in σ70 enrichment under anaerobic conditions correlated well, for the most part, with the expression changes for promoters activated or repressed by FNR, respectively, as well as expression changes in anaerobic and aerobic WT cultures (Table 1, Tables S2 and S12). Three operons, which lacked σ70 enrichment, have been shown to be dependent on σE (hcp-hcr) [52], σN (hycABCDEFGHI) [53] and σS (fbaB) [54], raising the possibility that alternative σ factors transcribe a subset of the FNR direct regulon.
Comparison of our FNR data with published regulatory data suggested that many FNR regulated operons were co-activated by TFs not active during growth in GMM, specifically NarL, NarP and CRP. For example, FNR-dependent transcription of napFDAGHBC, encoding the periplasmic nitrate reductase, requires co-activation by the NO3−/NO2− sensing response regulator NarP [55]. Transcriptomic data [19] showed FNR and NarL or NarP dependent activation in the presence of NO3− and/or NO2− (Table 2) [19] for nine operons that we found associated with FNR ChIP-seq peaks but lacking a FNR-dependent change in expression in our transcriptomic experiments, suggesting co-activation by NarL or NarP when NO3− and/or NO2− is present.
Another possible co-activator of operons in this group is CRP, which is inactive under glucose fermentation conditions presumably because of decreased cAMP [56]. Although previous studies have shown that ansB is co-activated by FNR and CRP [57], we did not observe binding of FNR upstream of ansB in this study, potentially due to differences in growth conditions. Nevertheless, 12 operons within this group showed an increase in anaerobic expression in transcriptomic data obtained from WT strains grown with carbon sources other than glucose (e.g. glycerol, mannose, arabinose or xylose) compared to growth in glucose (Table 3) [19], [58] (Park and Kiley, Personal Communication). A majority (nine) contained distinct CRP and FNR binding sites, suggesting co-activation by FNR and CRP when glucose is absent and cAMP levels are increased (Table 3). Interestingly, for the other three of these operons, guaB, ptsH and uxaB, the identified FNR binding site overlapped the CRP binding site, suggesting potential competition between FNR and CRP for binding when both TFs are active (Table 3).
We propose that FNR activation of ten operons is repressed by Fur under the iron replete conditions used here, similar to the known regulation of feoABC, encoding a ferrous iron uptake transporter [59]. In addition to feoABC, nine additional operons known to be bound by Fur had a FNR ChIP-seq peak but lacked a FNR-dependent change in expression, suggesting that Fur repression masked FNR regulation of these operons (Table 4).
Expression of several of the remaining operons associated with FNR ChIP-seq peaks are known to require other TFs but were not known to be co-regulated by FNR, potentially explaining the lack of FNR-dependent regulation under our growth conditions. A subset of these FNR-regulated operons may be co-regulated by OxyR (active under oxidative stress), CadC (active at low external pH) or PhoP (active in low Mg2+ concentration) (Table S9). In a recent SELEX study [60], three BasR binding sites were identified upstream of operons containing FNR peaks but without a FNR-dependent change in expression, suggesting BasR could possibly influence FNR regulation at these three promoters (Table S9).
In some cases, promoter architecture may mask FNR regulation. A small number of operons (12) contained multiple TSSs, raising the possibility that FNR may regulate transcription from a TSS that does not increase the total transcript levels to above the cutoff used in our analyses (Table S9). Alternative σ factors, active under other growth conditions, may also play a role in regulating transcription of a subset of these operons (Table S9). Taken together, we conclude that although FNR serves as a global signal for anaerobiosis, many operons likely require the combinatorial integration of TFs sensing other environmental signals for expression.
Surprisingly, a large number of operons (70) were differentially expressed by FNR but were not associated with a FNR ChIP-seq peak, suggesting they are regulated by FNR indirectly (Category 6, Table S10). To determine whether any of these operons had a FNR site upstream that was missed by ChIP-seq, sequences 500 nt upstream of these operons were searched using the FNR PWM and the algorithm PatSer with the PR curve determined threshold (Figure 2) [36]. Only one operon, hmp, contained a predicted FNR-binding site and previous data also supported FNR binding to hmp [61]. Thus, 69 operons are indirectly regulated by FNR. The indirect regulation by FNR could be easily explained for 11 operons targeted by the small RNA FnrS, which is directly activated by FNR [48], [49]. These RNAs increased in the FNR− strain because of the lower FnrS levels (Table 5) [48], [49].
To determine whether FNR binding to sites unmasked by the absence of H-NS and StpA caused a change in expression, we assayed if any of the corresponding genes were differentially expressed by O2 only in the Δhns/ΔstpA strain. Of the 158 new FNR peaks unmasked in the Δhns/ΔstpA strain, 18 genes showed an anaerobic increase in expression (Table S11), and consistent with this, many of the promoters contained a FNR binding site at a position associated with activation (e.g. near −41.5). For example, hemolysin E (hlyE), in agreement with previous results [62], and the anaerobic NAP Dan (ttdR) [63] showed increased expression under anaerobic conditions only. This suggests a possible role of Dan in the absence of H-NS and StpA. Only two genes showed a decrease in expression in the absence of H-NS and StpA (yncD and feaR) under only anaerobic growth conditions. However, the expression of the vast majority of genes having FNR bound at unmasked sites resulted in changes under both aerobic and anaerobic growth conditions, indicating that changes in nucleoid structure that occur in the absence of H-NS and StpA could cause misregulation of transcription. For example, H-NS and Rho coordinate to regulate transcriptional termination and the absence of H-NS may cause increased transcriptional readthrough of Rho-dependent terminators [64]. Thus, it seems likely that our analysis provides an underestimate of the impact of H-NS on FNR function, since physiological conditions that alter H-NS activity are likely to have less severe effects on nucleoid structure.
By combining genome-wide FNR occupancy data from ChIP-seq and ChIP-chip experiments with transcriptomic data, we uncovered new features of bacterial transcriptional regulation and the FNR regulon. Our findings suggest that in vivo FNR occupies only a subset of predicted FNR binding-sites in the genome, and that FNR binding can be blocked by NAPs like H-NS. Furthermore, the lack of correlation between match to consensus of FNR binding sites and ChIP enrichment suggests that variations in ChIP signal result from changes in cross-linking efficiency or epitope access rather than variable occupancy. We found that the FNR regulon is malleable; the set of genes controlled by FNR can be readily tailored to changing growth conditions that may activate or inactivate other TFs, allowing flexible reprograming of transcription. This strategy would allow the regulon to expand or contract depending on available nutrients, providing a competitive advantage in the ecological niche of E. coli of the mammalian gut [65].
The finding that there was little relationship between peak height and the quality of the FNR motif differs from the results found for LexA, which showed a correlation between peak height and match to consensus [10]. Our data suggest that FNR peak height may be more related to the efficiency of cross-linking or immunoprecipitation since sites that appear to be saturated for binding displayed significantly different peak heights. Thus, at least for FNR, peak height cannot be used to assess relative differences in site occupancy between chromosomal sites. Cross-linking or immunoprecipitation of FNR may be less efficient than for LexA because the larger number of other regulators bound at FNR-regulated promoters may affect accessibility to the cross-linking agent or FNR immunoprecipitation.
FNR sites having either a strong match to consensus (for example, ydfZ – TTGATaaaaAACAA) or a weak match (for example, frdA – TCGATctcgTCAAA) were saturated for binding at FNR dimer concentrations at its cellular level (∼2.5 µM) [37]; thus, in vivo most accessible FNR sites are likely to be fully occupied. These data also revealed that FNR occupancy was not significantly different for strong and weak sites over the tested range of FNR dimer concentrations, suggesting that in vivo FNR binding is unlikely to be dictated solely by the intrinsic affinity of FNR binding sites.
Our finding that not all predicted FNR binding sites are bound by FNR in vivo offers new insight into the accessibility of the genome for binding TFs. Previous studies have predicted anywhere from 12 to 500 FNR binding sites in the E. coli genome [66]–[69], depending on the algorithm used. Of the 187 FNR binding sites predicted here, only 63 contained a corresponding FNR ChIP-seq peak in the WT strain, suggesting many high quality FNR sites are not bound. Although some of these silent sites may result from false negatives in the ChIP experiments (e.g. failure to immunoprecipitate FNR bound at some sites), only five of the 124 silent FNR sites (acnA, aldA, hyfA, hmp and iraD) showed any evidence of FNR regulation in prior studies [70]. Rather, several lines of evidence suggest that binding of NAPs or other TFs masks FNR binding at many of these sites in vivo. First, we observed that binding sites for the NAPs IHF, H-NS, and Fis were statistically overrepresented at the positions of silent FNR binding sites, suggesting these proteins occlude FNR binding. Second, we found that in the absence of H-NS and StpA, additional FNR binding sites became available for FNR binding as detected by ChIP, suggesting that NAPs influence FNR site availability in vivo. A similar effect has been observed in eukaryotes, where extensive research on TF site availability has shown that chromatin structure in vivo can block binding of TFs (e.g. Pho4, Leu3 and Rap1) to high quality DNA binding sites [1], [2], [4]. Additionally, known changes to chromosomal structure by IHF, Fis, and H-NS have been shown to inhibit DNA binding of other proteins [7], [12], [71]. Thus, if the binding profiles of NAPs change under alternative growth conditions, then the occluded FNR binding sites would likely become available for FNR binding.
Nonetheless, the fact that the 207 FNR-enriched regions from this study included 80% of the 63 regions identified by Grainger et al. (Table S5), despite the difference in the growth conditions and experimental design [29], suggests that the overlapping subset of FNR binding events may reflect a core set that is insensitive to growth conditions or binding of other TFs. Furthermore, binding events specific to each growth condition may be reflective of either changes in accessibility of FNR to binding sites due to changes in DNA-binding protein distribution or perhaps increases in activity of a second TF that binds cooperatively. Other regulators, such as CRP, a closely related member of the FNR protein family, also appear to have more binding sites available genome-wide than are occupied in vivo under tested growth conditions. Shimada et al. identified 254 CRP-cAMP binding sites using Genomic SELEX screening, which was 3–4 fold more than the number of CRP sites previously identified by ChIP-chip experiments [72], [73]; thus not all chromosomal CRP sites appear to be accessible for binding, although additional experiments would be required to explicitly examine the accessibility of CRP binding sites throughout the genome. Taken together, these results suggest that the restrictive effect of chromosomal structure could influence TF binding beyond FNR.
Environmental stimuli that change NAP distribution would also change TF binding site accessibility and affect transcription. For example, as E. coli enters the mammalian GI tract, it experiences a temperature increase from ∼25°C to 37°C, and this increase in temperature has been shown to affect transcription of a number of operons, including increased expression of anaerobic-specific operons [74], [75]. Because H-NS binding is sensitive to changes in temperature [76], [77], an explanation for these temperature-dependent transcriptional changes [74], [75] could be genome-wide decreases in H-NS binding and distribution; these changes could increase the accessibility of the binding sites for FNR and other TFs to regulate transcription. Supporting this explanation, several genes with a temperature dependent increase in expression showed FNR binding and regulation in the absence of H-NS and StpA, including hlyE, feaR, yaiV, and torZ. The activity of NAPs can also be affected by the binding of other condition specific TFs. For example, ChIP-chip and Genomic SELEX analysis of the stationary phase LysR-type TF, LeuO, suggested that binding of LeuO antagonized H-NS activity, but not necessarily H-NS binding, throughout the genome in Salmonella enterica and E. coli [78], [79].
Thus, a picture emerges from our data that binding of FNR is dependent on characteristics of the genome beyond the presence of a FNR binding site; this restrictive effect of chromosome structure by NAPs may affect binding of other TFs in bacteria. NAPs have been shown to occlude and affect binding of TFs and other DNA binding proteins, such as restriction endonucleases and DNA methylation enzymes, suggesting a general role of NAPs in regulating genome accessibility by bending, wrapping and bridging the DNA structure [7], [12], [13], [27], [42], [76], [80], [81]. Additionally, NAPs influence DNA supercoiling, which has been shown to affect binding of the TFs Fis and OmpR in S. enterica [82], [83], providing another mechanism by which NAPs can change the chromosomal structure to influence TF-DNA binding. Taken together, our results support a dynamic model of complex genome structure that affects TF binding to control gene regulation in bacteria.
Although expression of a subset of the operons in the FNR regulon appeared to require only FNR for regulation (Categories 1 and 2), our findings point to widespread cooperation between FNR and other TFs for condition-specific regulation (Categories 3 and 4). Changes in activity of these TFs would result in FNR regulation to adapt to changes in environment, such as growth in non-catabolite repressed carbon sources (CRP) [57], anaerobic respiration of nitrate (NarL and NarP) [19], and growth in iron-limiting conditions (Fur) [84]. Although this co-regulation provides insight into growth conditions that should allow FNR-dependent changes in gene expression, the synergistic regulators for many promoter regions bound by FNR are currently unknown (Category 5), but would likely be identified in future genome-scale studies using different growth conditions, particularly microaerobic growth, which has been shown to affect FNR regulation of virulence genes in the pathogen Shigella flexneri [85].
Overall, our results suggest that the regulation of a subset of FNR-dependent promoters in E. coli may depend on combinatorial regulation with other TFs, a mechanism that resembles regulation of eukaryotic promoters [8], [20], [86]. These experimental data support previous in silico regulatory models generated using published data [87]–[89], suggesting combinatorial regulation may be common in E. coli. Further, ChIP-chip and ChIP-seq analyses of other TFs in E. coli (e.g. CRP, Fis, and IHF) and Salmonella typhimurium (e.g. Sfh, a H-NS homolog), identified many TF binding sites that did not correlate with changes in gene expression in corresponding TF-specific transcriptomic experiments [30], [33], [73], [90], [91]. These results raise the possibility of potential combinatorial regulation for other TFs, although additional analysis is required to support this notion.
We found that FNR directly controls expression of five secondary regulators, most of which are also regulated by specific cofactors, suggesting that the scope of the indirect FNR regulon (Category 6) is also likely to change depending on growth conditions. Of the five regulators, three act in an apparent hierarchal manner. The small RNA FnrS, which is upregulated by FNR and is suggested to stimulate mRNA turnover, decreased the mRNA levels of multiple FnrS target genes in GMM [48], [49]. Expression of the TF CaiF was also activated by FNR, but the genes regulated by CaiF were not expressed in GMM because CaiF requires the effector carnitine to be active [92]. FNR activated BssR, a TF involved in biofilm formation. About ∼40 operons are thought to be controlled by BssR [93], but none of the five BssR-dependent operons in the FNR indirect regulon that we tested by qRT-PCR showed any change in expression in a BssR− strain (data not shown); thus, under our growth conditions, BssR appeared to be inactive.
FNR also directly repressed the expression of two TFs, including the pyruvate sensing TF PdhR which represses several operons in the absence of pyruvate [94], [95]. Although one might expect that PdhR repressed genes would increase anaerobically, many of these genes are redundantly repressed by ArcA (Park and Kiley, Personal Communication); thus the impact of PdhR may be negligible under anaerobic growth in GMM. Similarly, the TF GadE, which is active at low pH [96], was also directly repressed by FNR and accordingly the operons in the GadE regulon were not identified as part of the indirect FNR regulon in GMM.
Finally, we note the caveat that some operons that appear indirectly regulated by FNR may change expression as a result of indirect physiological and metabolic effects in a FNR− strain, which may alter the activity of other TFs, resulting in mis-regulation of operons. For example, our data show that FNR does not directly regulate arcA transcription, but previous results have suggested that ArcA activity may be affected by the metabolic changes that occur when fnr is deleted [97]. Thus, although a subset of ArcA regulatory targets (29 operons) showed potential indirect FNR regulation, such effects were likely caused by changes in the phosphorylation state of ArcA resulting from metabolic changes in a FNR- strain (Table S13) (Park and Kiley, Personal Communication).
In conclusion, our results reveal complex features of TF binding in bacteria and expand our understanding of how E. coli responds to changes in O2 and other environmental stimuli.
A subset of predicted FNR binding sites appear to be inhibited by NAPs and are available in the absence of H-NS and StpA, suggesting that the bacterial genome is not freely accessible for TF binding and that changes in TF binding site accessibility could result in changes in transcription. Finally, correlation of the occupancy data with transcriptomic data suggests that FNR serves as a global signal of anaerobiosis but the expression of a subset of operons in the FNR regulon requires other regulators sensitive to alternative environmental stimuli. This strategy is reminiscent of global regulation by CRP-cAMP [73] in that FNR, like CRP, is bound at many promoters under specific conditions without corresponding changes in mRNA levels, suggesting a common strategy whereby promoters are primed to be activated when the appropriate growth conditions are encountered.
All strains were grown in MOPS minimal medium supplemented with 0.2% glucose (GMM) [98] at 37°C and sparged with a gas mix of 95% N2 and 5% CO2 (anaerobic) or 70% N2, 5% CO2, and 25% O2 (aerobic). Cells were harvested during mid-log growth (OD600 of ∼0.3 using a Perkin Elmer Lambda 25 UV/Vis Spectrophotometer). E. coli K-12 MG1655 (F-, λ-, rph-1) and PK4811 (MG1655 ΔfnrΩSpR/SmR) [99] were used for the ChIP-chip, ChIP-seq and transcriptomic experiments unless otherwise specified. All data obtained in this study used GMM as the growth media, and although we know that not all promoters directly regulated by FNR are expressed under these conditions, this has the advantage that both mutant and parental strains exhibit the same growth rate.
For experiments that varied the in vivo concentration of FNR, a strain that contained a single, chromosomal copy of WT fnr under the control of the Ptac promoter at the λ attachment site was constructed. Following digestion of pPK823 [99] with XbaI and HindIII, the DNA fragment containing fnr was cloned into the XbaI and HindIII sites of pDHB60 (ApR) [100] to form pPK6401. Plasmid pPK6401 was transformed into DHB6521 [100] and the Ptac-fnr construct was stably integrated into the λ attachment site using the Lambda InCh system as described [100] to produce PK6410. P1vir transduction was used to move the Ptac-fnr, ApR allele into strain PK8257, which contains the FNR activated ydfZ promoter-lacZ fusion and deletion of lacY. This strain was transformed with pACYClacIQ-CAM [101] to generate PK8263. To determine the effect of FNR on the expression of the BssR regulon, a ΔbssR strain was constructed by P1vir transduction of ΔbssR::kanR from the Keio collection [102] into MG1655 to generate PK8923. To determine the role of H-NS on FNR binding, first stpA was recombined with the CmR gene, cmr, using λ red recombination and the pSIM plasmid [103]. P1vir transduction introduced the Δhns::kanR allele from the Keio collection [102] into the strain lacking stpA to generate the Δhns/ΔstpA strain.
Total RNA was isolated as previously described [104]. The concentration of the purified RNA was determined using a NanoDrop 2100, while the integrity of the RNA was analyzed using an Agilent 2100 Bioanalyzer and the RNA Nano LabChip platform (Agilent).
Total RNA (10 µg) from two biological replicates each of MG1655 (+O2 and −O2) and PK4811 was reverse transcribed using random hexamers (Sigma) and the SuperScript II Double-Stranded cDNA Synthesis Kit (Invitrogen) following the manufacturer's protocol. The cDNA (1 µg) was fluorescently labeled with Cy3-labeled 9 mers (Tri-Link Biotechnologies) with Klenow Fragment (NEB) for 2 hours at 37°C and recovered using ethanol precipitation. Labeled dsDNA (2 µg) was hybridized onto the Roche NimbleGen E. coli 4plex Expression Array Platform (4×72,000 probes, Catalog Number A6697-00-01) for ∼16 hours at 42°C in a NimbleGen Hybridization System 4 (Roche NimbleGen) following the manufacturer's protocol. The hybridized microarrays were scanned at 532 nm with a pixel size of 5 µm using a GenePix 4000B Microarray Scanner (Molecular Devices), and the PMT was adjusted until approximately 1% of the total probes were saturated for fluorescence intensity. The data were normalized using the Robust Multichip Average (RMA) algorithm in the NimbleScan software package, version 2.5 [105]. ArrayStar 3.0 (DNASTAR) was used to identify genes that showed at least a two-fold change in expression between the WT and Δfnr strains and were significantly similar among biological replicates, using a moderated t-test (p-value<0.01) [106]. Genes were organized into operons using data from EcoCyc [70]. An operon was called differentially expressed (DE) if only one gene within an operon showed a statistically significant change in expression. NimbleGen microarrays identified 214 statistically significant DE genes that were contained within 134 operons
The anaerobic MG1655 and FNR− samples from the normalized whole genome expression microarray data from Kang et al. [18] were also analyzed. Genes were determined to be DE if they had a change in expression greater than or equal to two-fold and if the genes were found to be statistically similar between biological replicates using a t-test (p-value<0.01). An operon was called DE if only one gene within an operon showed a statistically significant change in expression. This analysis identified 204 significant DE genes in 130 operons. Sixty operons were found to be DE in both the NimbleGen and Kang et al. data sets (Table S8). Of the 70 operons found DE in only the Kang et al. data set, 41 operons were just below the significance threshold in the NimbleGen data set and 11 operons resulted from activation of the flagellar regulon due to an insertion upstream of flhDC, which was absent in the isolate of MG1655 used in this study.
The Δhns/ΔstpA aerobic and anaerobic expression data were obtained from stand specific, single stranded cDNA hybridized to custom designed, high-density tiled microarrays containing 378,000 probes from alternate strands, spaced every ∼12 bp through the genome as described previously [107] except Cy3 was used instead of Cy5. Microarray hybridization and scanning were performed as described above except that the PMT was adjusted until the median background value was ∼100. All probe data were normalized using RMA in the NimbleScan software package, version 2.5 [105]. Gene probe values found to be significantly different between two biological replicates using a Benjamini & Hochberg corrected t-test (p-value<0.05) were eliminated from further analysis. Genes were called DE if the median log2 values were different by more than two-fold and if the genes were significantly different using an ANOVA test (p-value<0.05).
To enrich for mRNA from total RNA, the 23S and 16S rRNA were removed using the Ambion MICROBExpress kit (Ambion) following manufacturer's guidelines, except the total RNA was incubated with the rRNA oligonucleotides for one hour instead of 15 minutes. The rRNA depleted RNA samples isolated from two biological replicates of MG1655 and its FNR− derivative were processed by the Joint Genome Institute (JGI) for RNA-seq library creation and sequencing. The RNAs were chemically fragmented using RNA Fragmentation Reagents (Ambion) to the size range of 200–250 bp using 1× fragmentation solution for 5 minutes at 70°C (Ambion). Double stranded cDNA was generated using the SuperScript Double-Stranded cDNA Synthesis Kit (Invitrogen) following the manufacturer's protocol. The Illumina Paired End Sample Prep kit was used for Illumina RNA-seq library creation using the manufacturer's instructions. Briefly, the fragmented cDNA was end repaired, ligated to Illumina specific adapters and amplified with 10 cycles of PCR using the TruSeq SR Cluster Kit (v2). Single-end 36 bp reads were generated by sequencing on the Illumina Genome Analyzer IIx, using the TruSeq SBS Kit (v5) following the manufacturer's protocol. Resulting reads were aligned to the published E. coli K-12 MG1655 genome (U00096.2) using the software package SOAP, version 2.20 [108], allowing no more than two mismatches. Reads aligning to repeated elements in the genome (for example rRNA) were removed from analysis. For reads that had no mapping locations for the first 36 bp, the 3–30 bp subsequences were used in the subsequent mapping to the reference genome. Reads that had unique mapping locations and did not match annotated rRNA genes were used for further analysis. For each gene, the tag density was estimated as the number of aligned sequencing tags divided by gene size in kb and normalized using quantile normalization. The tag density data were analyzed for statistically significant differential expression using baySeq, version 2.6 [109] with a FDR of 0.01, and genes were organized into operons using data from EcoCyc [70]. An operon was called DE if only one gene within an operon showed a statistically significant change in expression. The RNA-seq analysis identified 133 statistically significant DE operons (197 genes). Altogether, microarray and RNA-seq experiments identified 258 operons DE by FNR and slightly fewer than half of these operons (122) were found in at least two of the transcriptomic experiments (Figure S5, Table S8).
ChIP assays were performed as previously described [110], except that the glycine, the formaldehyde and the sodium phosphate mix were sparged with argon gas for 20 minutes before use to maintain anaerobic conditions when required. Samples were immunoprecipitated using polyclonal antibodies raised against FNR, IHF or H-NS, which had been individually absorbed against mutant strains lacking the appropriate protein. In the case of FNR, affinity purified antibodies were used in some experiments, purified using the method previously described [111]. For RNA Polymerase, a σ70 monoclonal antibody from NeoClone (W0004) or a RNA Polymerase ß monoclonal antibody from NeoClone (W0002) were used for immunoprecipitation. For FNR, neither lengthening the cross-linking time nor increasing or decreasing the amount of FNR antibody used in the ChIP protocol showed significant changes in the FNR ChIP-chip peak heights or number of peaks identified. For ChIP-chip, FNR (three samples), FNR− (one sample), β (two samples), H-NS (two samples) and IHF (two samples) were fluorescently-labeled using Cy3 (INPUT) and Cy5 (IP) and hybridized for ∼16 hours at 42°C in a NimbleGen Hybridization System 4 (Roche NimbleGen) to custom designed, high-density tiled microarrays containing 378,000 probes from alternate strands, spaced every ∼12 bp through the genome. The hybridized microarrays were scanned at 532 nm (Cy3) and 635 nm (Cy5) with a pixel size of 5 µm using a GenePix 4000B Microarray Scanner (Molecular Devices), and the PMT was adjusted until approximately 1% of the total probes were saturated for fluorescence intensity of each dye used. The NimbleScan software package, version 2.5 (Roche NimbleGen) was used to extract the scanned data. ChIP-chip data were normalized within each microarray using quantile normalization (“normalize.quantiles” in the R package VSN, version 3.26.0) [112] to correct for dye-dependent intensity differences as previously described [113]. Biological replicates were normalized between microarrays using quantile normalization as previously described [113], and the normalized log2 ratio values (IP over INPUT) were averaged. There was a strong correlation between enriched regions of ChIP-chip biological replicates (R = 0.7). ChIP-chip peaks for FNR, H-NS and IHF were identified in each data set by the peak finding algorithm CMARRT, version 1.3 (FDR of 0.01) [114] and proportional Z-tests were used to determine significant differences between proportional data.
For ChIP-seq experiments, 10 ng of immunoprecipitated and purified DNA fragments from the FNR (two biological replicates) and σ70 samples (two biological replicates from both aerobic and anaerobic growth conditions), along with 10 ng of input control, were submitted to the University of Wisconsin-Madison DNA Sequencing Facility (FNR samples and one σ70 sample) or the Joint Genome Institute (one σ70 sample) for ChIP-seq library preparation. Samples were sheared to 200–500 nt during the IP process to facilitate library preparation. All libraries were generated using reagents from the Illumina Paired End Sample Preparation Kit (Illumina) and the Illumina protocol “Preparing Samples for ChIP Sequencing of DNA” (Illumina part # 11257047 RevA) as per the manufacturer's instructions, except products of the ligation reaction were purified by gel electrophoresis using 2% SizeSelect agarose gels (Invitrogen) targeting either 275 bp fragments (σ70 libraries) or 400 bp fragments (FNR libraries). After library construction and amplification, quality and quantity were assessed using an Agilent DNA 1000 series chip assay (Agilent) and QuantIT PicoGreen dsDNA Kit (Invitrogen), respectively, and libraries were standardized to 10 µM. Cluster generation was performed using a cBot Single Read Cluster Generation Kit (v4) and placed on the Illumina cBot. A single-end read, 36 bp run was performed, using standard SBS kits (v4) and SCS 2.6 on an Illumina Genome Analyzer IIx. Basecalling was performed using the standard Illumina Pipeline, version 1.6. Sequence reads were aligned to the published E. coli K-12 MG1655 genome (U00096.2) using the software packages SOAP, version 2.20, [108] and ELAND (within the Illumina Genome Analyzer Pipeline Software, version 1.6), allowing at most two mismatches. Sequence reads with sequences that did not align to the genome, aligned to multiple locations on the genome, or contained more than two mismatches were discarded from further analysis (<10% of reads). For visualization the raw tag density at each position was calculated using QuEST, version 1.2 [115], and normalized as tag density per million uniquely mapped reads. The read density was determined for each base in the genome for the IP and INPUT samples for FNR and σ70 samples. For FNR, peaks were identified using three peak finding algorithms: CisGenome, version 1.2, NCIS, version 1.0.1, and MOSAiCS, version 1.6.0 [116]–[118] (FDR for all of 0.05), while σ70 peaks were identified using NCIS, version 1.0.1 (FDR of 0.05). Further discussion of these algorithms is in Text S1. Differences between aerobic and anaerobic σ70 ChIP-seq occupancy were determined using a one-sided, paired t-test (p-value<0.01) comparing 100 bp surrounding the center of each peak. To normalize between +O2 and −O2 samples, the read counts for the enriched regions (peaks) for each sample were shifted by the negative median read count value of the background (un-enriched) signal. The p-values were adjusted using the Bonferroni method to correct for multiple testing. There was a strong correlation between ChIP-seq biological replicates (R = 0.8) as well as between ChIP-chip and ChIP-seq data (Figure S6). All data were visualized in the MochiView browser [119].
Additional ChIP-chip -O2 data sets were performed for WT FNR and a Δfnr [99] control. The 15 FNR peaks identified only in ChIP-chip had low IP/INPUT ratios and were eliminated since ChIP-seq is known to have increased signal to noise relative to ChIP-chip [120]. The Δfnr -O2 ChIP-chip data identified 71 peaks that corresponded to peaks in the FNR -O2 ChIP-seq data, indicating they were not FNR specific, and were removed from the FNR ChIP-seq dataset (Table S5).
To construct the FNR PWM, the sequence of a region of ∼100 bp around the nucleotide with the largest tag density within each of the FNR ChIP-seq peaks (the summit of each peak) found by all three peak finding algorithms was analyzed. MEME was used to identify over-represented sequences [121] and the Delila software package was used to construct the PWMs [122]. To search all ChIP-seq peaks for the presence of the FNR PWM, a region of 200 bp around the summit of each FNR ChIP-seq peak was searched with the FNR PWM using PatSer, version 3e [36], and the top four matches to the FNR PWM, as determined by PatSer PWM score, were recorded at each ChIP-seq peak. The standard deviation of the PatSer scores for the four FNR predicted binding sites at each ChIP-seq peak was determined and used as a threshold to determine the number of predicted binding sites at each peak. If the PatSer predicted FNR binding site at a peak with the highest PatSer score was more than one standard deviation greater than the PatSer predicted FNR binding site with the second best PatSer score, that peak was identified as having only one predicted FNR binding site. For FNR peaks (∼11%) with the two best PatSer predicted FNR binding site scores less than one standard deviation apart, a Grubbs test for outliers was used a single time to identify outliers within the four PatSer predicted FNR binding sites at a peak (α of 0.15, critical Z of 1.04). If a PatSer predicted FNR binding site at a FNR peak was identified as an outlier, it was removed from analysis and the standard deviation was re-calculated using the remaining three PatSer binding site scores at that peak. The remaining PatSer predicted FNR binding sites at the FNR peak were then re-examined as described above. After removing outlier PatSer predicted FNR binding sites, a peak was determined to contain two predicted FNR binding sites if the two best predicted FNR binding sites at that peak had PatSer scores less than one standard deviation apart.
The precision-recall curve was constructed using the FNR PWM and searching throughout the genome using PatSer, version 3e [36]. Precision was defined as True Positives (locations with a FNR ChIP-seq peak and a predicted FNR binding site) divided by True Positives plus False Positives (locations with a predicted FNR binding site but no FNR ChIP-seq peak). Recall was defined as True Positives divided by True Positives plus False Negatives (locations with a FNR ChIP-seq peak but no FNR predicted binding site). A high precision value means all predicted binding sites are true positives, but there is a high false negative rate. A high recall value means all true positives have been captured, but there is a high false positive rate.
The strain with fnr under the control of Ptac (PK8263) was used to study changes in [FNR] on ChIP-chip peak height. Cultures were grown anaerobically overnight in MOPS+0.2% glucose and were subcultured to a starting OD600 of ∼0.01 in MOPS+0.2% glucose plus Cm20 and various [IPTG] (4 µM IPTG, 8 µM IPTG, and 16 µM IPTG). After this initial step, growth, ChIP-chip experiments (two biological replicates of 4 and 8 µM IPTG and three biological replicates of 16 µM IPTG were used) and initial analysis were identical to the procedures described above. Estimates of FNR concentration were determined by quantitative Western blot as previously described [37]. A novel method of normalization was developed to compare peak areas between IPTG concentrations for 35 peaks that showed a large distribution in peak heights and 4 peaks that were classified as false positives by enrichment in the Δfnr ChIP-chip sample. The peak finding algorithm CMARRT identified peaks in the WT FNR ChIP-chip sample, and this peak region was trimmed to include the center 50% of the peak region. This trimmed region was used for each [IPTG] sample for consistency. For each of the 39 peaks examined, the probe values in a region of ∼3000 bp beyond the peak boundary (∼1500 bp upstream and downstream of the peak boundary) was selected for analysis from each sample. Within the ∼3000 bp region, the probes beyond the peak boundary were considered background for each sample. The median of the background (un-enriched) probes was calculated and the log2 IP/INPUT probe values for the entire peak region (enriched and un-enriched) were shifted by the negative median value of the background probes. The peak average (average of log2 IP/INPUT values) and standard deviation was determined for 39 peak regions to compare between samples at each [IPTG] and WT ChIP-chip samples. A one-sided, paired t-test was performed between all conditions (p-value<0.05) to determine statistically significant changes in average peak values.
Growth, ChIP-chip experiments, normalization and peak calling was performed as described above. To normalize between WT and Δhns/ΔstpA samples, the enriched regions (peaks) for each sample were shifted by the negative median log2 IP/INPUT value of the background (un-enriched) probes. The peak averages (average of log2 IP/INPUT values) were determined for each condition (WT and Δhns/ΔstpA) at each FNR peak found in either strain background. A one-sided, paired t-test with Bonferroni correction was performed between the two conditions (p-value<0.05) to determine the statistically significant change in peak averages. For peaks found in both WT and Δhns/ΔstpA, peaks were identified as significantly higher in Δhns/ΔstpA using a one-sided, paired t-test with Bonferroni correction performed between the two conditions (p-value<0.05) and if the FNR peak average in the Δhns/ΔstpA strain was greater than the standard deviation found for WT peak average.
The ChIP-chip and ChIP-seq data can be visualized on GBrowse at the following address: “http://heptamer.tamu.edu/cgi-bin/gb2/gbrowse/MG1655/”. All genome-wide data from this publication have been deposited in NCBI's Gene Expression Omnibus (GSE41195) (Table S14) [123].
|
10.1371/journal.pntd.0000486 | Bioluminescent Imaging of Trypanosoma brucei Shows Preferential Testis Dissemination Which May Hamper Drug Efficacy in Sleeping Sickness | Monitoring Trypanosoma spread using real-time imaging in vivo provides a fast method to evaluate parasite distribution especially in immunoprivileged locations. Here, we generated monomorphic and pleomorphic recombinant Trypanosoma brucei expressing the Renilla luciferase. In vitro luciferase activity measurements confirmed the uptake of the coelenterazine substrate by live parasites and light emission. We further validated the use of Renilla luciferase-tagged trypanosomes for real-time bioluminescent in vivo analysis. Interestingly, a preferential testis tropism was observed with both the monomorphic and pleomorphic recombinants. This is of importance when considering trypanocidal drug development, since parasites might be protected from many drugs by the blood-testis barrier. This hypothesis was supported by our final study of the efficacy of treatment with trypanocidal drugs in T. brucei-infected mice. We showed that parasites located in the testis, as compared to those located in the abdominal cavity, were not readily cleared by the drugs.
| Human African trypanosomiasis or sleeping sickness, caused by two subspecies of Trypanosoma brucei, is endemic in Subsaharan Africa. There is no vaccine and the currently used drugs are toxic and can cause severe side effects and even death. At present, we do not know how and when parasites can leave the blood and penetrate into organs (especially the brain). Such knowledge will be very helpful to develop and validate new drugs that can clear the parasite from both the blood and the tissues. In this study, we developed a novel technique allowing us to track the parasites in a live animal by the use of light signals. By following the luminescent parasites in the mouse we showed that, interestingly, the organisms migrate very early in infection to the testes. Here, they may be protected from the immune system and from drugs. Indeed when treating the mice with a sub-optimal dose of medicine, the parasites in this location were not cleared and subsequently could cause a reinvasion into the blood of the host.
| Human and animal African trypanosomoses are important protozoan infections endemic in Africa, Latin America and Asia, caused by several species such as Trypanosoma brucei, T. evansi, T. equiperdum, T. congolense and T. vivax. Different species, strains within the species and clones within various strains show different tissue tropism that may further vary within hosts [1]. Currently no vaccines against human and animal trypanosomoses are available; and a limited range of drugs exist to treat these diseases. Moreover, most of the drugs used in second stage sleeping sickness show a high toxicity while in animal trypanosomosis drug resistance becomes more and more problematic [2]. Our current knowledge on tissue tropism, mechanisms by which trypanosomes invade and spread into tissues, the temporal course of invasion and the drug accessibility to trypanosomes in tissues, is incomplete. To complement classical anatomopathological examinations, real-time biophotonic imaging seems straight forward. Bioluminescence in vivo imaging allows longitudinal monitoring of an infection in the same animal, a desirable alternative to analyzing a number of animals at many time points during the course of the infection. To date, most bioluminescence models have been generated to monitor pathogenic bacterial infections, such as Salmonella, and bacterial meningitis [3],[4]. Among pathogenic protozoa only Plasmodium berghei, Leishmania amazonensis and Toxoplasma gondii have been engineered to express the firefly luciferase and used in bioluminescence imaging [5]–[7]. To our knowledge, no bioluminescent model for trypanosomes has been developed. Here we report the generation of recombinant Renilla luciferase expressing parasites, and the validation of the use of a real time biophotonic detection of parasites to study the dissemination of African trypanosomes in mice in vivo and the efficiency of treatment with trypanocidal drugs in vitro and in vivo.
The Rluc gene was PCR-amplified from pGL4.70 (Promega) and cloned into the pHD309 plasmid [8] using the In-Fusion PCR cloning kit (ClonTech). Plasmids were screened via HindIII/BamHI double restriction-digestion, sequenced, and those with the correct insert in frame were selected and propagated in E. coli. Ten µg of the Rluc-pHD309 plasmid was linearized using NotI (10 U, 3 hours at 37 C). T. brucei bloodstream forms (Lister 427 host cell line 90-13 and AnTat 1.1E) were cultured at 37°C, 5% CO2 in IMDM medium (Gibco) supplemented with 10% (v/v) heat-inactivated fetal calf serum, 36 mM sodium bicarbonate, 136 µg.ml−1 hypoxanthine, 39 µg.ml−1 thymidine, 110 µg.ml−1 NaPyruvate, 28 µg.ml−1 bathocuproine, 0.25 mM β-mercaptoethanol, 2 mM L-cystein and 62.5 µg.ml−1 kanamycin [9]. A pellet of 2×107 parasites was resuspended in 400 µl warm cytomix (2 mM EGTA, 120 mM KCL, 0.15 mM CaCl2, 10 mM K2HPO4/KH2PO4 pH 7.6, 25 mM Hepes, 0.5%Glucose, 1 mM Hypoxanthine, 100 µg/ml BSA, 5 mM MgCL2) [10] and transferred into a 4 mm cuvette; 10 µg of linearized DNA construct was added and left for one minute at 37°C. Subsequently the mixture was pulsed once in a Gene Pulse Xcell square wave electroporator at 1250 V, 25 Ohm, 50 µF and transfected cells were added to 12 ml of preheated IMDM, plated in a 24-well plate (24-time 500 µl) and incubated at 37°C for 24 h. Next, 500 µl of preheated IMDM containing 10 µg/ml hygromycin were added to obtain a final selection concentration of 5 µg/ml. Positive clones were evident at 6 days post transfection.
Transfected Trypanosoma brucei brucei AnTat 1.1E bloodstream form trypanosomes were grown for 3 days in mice. Ten microlitres of blood were spread over a microscope slide and fixed with acetone for 15 minutes. The fixed cells were incubated at room temperature with primary and secondary antibodies for 1 h and 30 min, respectively, and washed two times for 5 min with PBS after each of the incubations. The primary antibody, monoclonal mouse anti-Renilla luciferase (Millipore), was diluted 1∶2 in PBS. The secondary antibody, fluorescein isothiocyanate (FITC)-conjugated goat anti-mouse (Jackson), was diluted 1∶100 in a solution of 0.1 mg.ml−1 Evans Blue and 1 µg.ml−1 DAPI in PBS. Cells were analyzed on an Olympus BX-41 UV microscope, and images were captured by a Colorview II camera (Soft Imaging Systems) and Cell_D software (Soft Imaging Systems) was used for analysis.
The Renilla Luciferase Assay System (Promega) was used to measure in vitro luciferase activity. Non-transformed T. brucei Lister 427 and T. brucei 427–Rluc-pHD309 clones were grown up to a total of 1×107 parasites (10 ml of 1×106 cells/ml), spun down at 1500 g for 10 minutes, resuspended in 20 µl IMDM medium and subsequently added to 100 µl of Renilla Luciferase Assay (1 µl of 100× Renilla Luciferase Assay substrate dissolved into 100 µl of Renilla Luciferase Assay Buffer). The level of Renilla luciferase activity (RLU) from 1×106 samples was monitored at different time points after substrate addition in a luminometer. To monitor the signal of lysed cells, the same amount of cells was lysed and measured in the same system according to manufacturers instructions.
Mice were anaesthetized with 2.3% isoflurane. At different days after infection, mice were injected intraperitoneally with 100 µL of coelenterazine (2 µg/µl dissolved in methanol) (Synchem) diluted with 90 µL PBS pH 7, anaesthetized with isoflurane and light emission in photons/second/cm2/steradian (p/sec/cm2/sr) was recorded in an IVIS Imaging System 100 (Xenogen LifeSciences) and Living Image® 2.20.1 software (Xenogen) for 180 seconds. Measurements started 3–5 minutes after substrate injection to allow the spread of the coelenterazine.
Mice at 25 days after infection were deeply anaesthetized with isoflurane, sacrificed and testes were dissected. To examine presence of trypanosomes within and outside blood vessels in the testis, sections were cut, mounted, fixed and immunostained with anti-AnTat 1.1 VSG (1∶5.000) and goat polyclonal anti-glucose transporter 1 (1∶40; GLUT-1, Santa Cruz Biotechnology, Santa Cruz, CA, USA) as described previously [11]. Glut 1 usually used to stain cerebral blood vessels has also been shown to be expressed by testicular endothelial cells [12]. Sections were examined and analysed using a Nikon fluorescence microscope. Photomicrographs were taken with a Zeiss AxioCam digital camera.
WST-1(tetrazolium salt) salts are cleaved to formazan by cellular oxidoreductases. The augmentation in enzyme activity lead to an increase in the amount of formazan dye formed. The viable cells were quantified by the formazan dye produced by metabolically active cells. To measure the drug sensitivity of cordycepin and other drugs, 25000 parasites were cultured in 100 µl in vitro in a 96 well culture plate with serial drug dilutions. Viability of the parasites was measured by adding 10 µl of WST reagent and further incubation for 2 hours. Readings were taken by a multiwell scanning spectrophotometer at excitation wavelength of 450 nm.
In order to measure single cell proliferation trypanosomes at different days of culture CFSE labeling was performed as described for mammalian cells [13]. To each milliliter of cell suspension, 2 µl of carboxyfluorescein diacetate, succinimidyl ester or CFDA-SE (CFSE) 5 mM stock solution was added and immediately mixed to ensure uniform staining, resulting in a final concentration of 10 µM CFSE. The cells were incubated 15 min at 37 C and the cells were quenched by adding 5 volumes of culture medium. Cells were analysed by flow cytometry with logarithmic detection of green fluorescence.
All mice were housed in filter-top cages and maintained in SPF barrier facilities in individual ventilated cages at the Karolinska Institute, Stockholm, or at the Institute for Tropical Medicine Antwerp. Animal ethics approval for the infection of live animals with recombinant trypanosomes was obtained from the respective Animal Ethical Committees of the Karolinska Institute (Sweden) and the Institute of Tropical Medicine, Antwerp (Belgium).
For stable transfection of Renilla luciferase (Rluc) into the β-tubulin region of the bloodstream form of monomorphic T b. brucei Lister 427 and pleomorphic T. b. brucei AnTat 1.1, the Rluc gene was PCR-amplified and cloned into the pHD309 plasmid [8]. Plasmids were screened via HindIII/BamHI double restriction-digestion, sequenced, and those with the correct insert in frame were selected and propagated in E. coli. For transfection, 2×107 parasites were electroporated with NotI linearized DNA construct in a BioRad Gene Pulse Xcell square wave electroporator. Two independent transfections were performed and three clones from each population were selected for further luminescence experiments
The kinetics of luciferase activity of T. brucei Lister 427 and AnTat 1∶1 clones showed a fast reaction and prolonged response during time. Both live and lysed cells showed a high relative luciferase units (RLU) activity, although the RLU signals for lysed cells were about 5 to 10 times higher (Figure 1A and data not shown). A linear relationship between concentration of live parasites and the RLU could be observed (Figure 1B). The Renilla Luciferase Assay System (Promega) was used to measure in vitro luciferase activity of live and lysed parasites.
Subsequently, we verified whether the RLU signal measured from live cells was due to substrate uptake and not residual activity of free luciferase released from damaged or live cells during manipulation. First, cells were spun down, the supernatants collected and the cell pellet resuspended and luciferase activity measured in supernatants and cell pellets. Over 70% of the original light emission was generated by the cell pellet whereas a negligible signal was detected in the supernatant (data not shown). As a second control, FACS analysis performed on 106 non-lysed parasites incubated with propidium iodide, a marker for non-viable cells, showed no incorporation of the dye, indicating that over 99% of the cells were intact after manipulations (data not shown). Taken together, these results suggest that the luciferase substrate coelenterazine penetrates or is taken up by live trypanosomes and that the detected luciferase is not secreted or released by the trypanosomes.
Renilla luciferase was expected to locate in the cytoplasm. Indeed, luciferase was immunostained throughout the whole cell with a slightly higher concentration around the flagellum (Figure 1C).
We then investigated if coelenterazine is toxic for parasites since this could hamper the follow up of the infection in vivo. Parasites were grown in vitro in the presence of different concentrations of coelenterazine and parasite growth was measured during 72 hours. Coelerenterazine at concentrations required for in vivo usage did not inhibit parasite proliferation (Figure 1D).
T. brucei undergoes a life cycle stage differentiation from a long slender to a short stumpy form [14]). We analysed whether slender and stumpy forms of T. brucei express luciferase activity and cleave WST-1. Parasites were alive in our culture condition until day 4–5 of culture. Exponential growth of parasites was observed during the first 48 h in cultures whereas at day 3–4 similar parasite concentrations were observed. Confirming previous data [14], the lack of increase in parasite density in the cultures was due to an arrest in proliferation rather than to an increased parasite death, as visualized by the dilution of CFSE by parasites during the days 1 and 2 but not 3 and 4 of culture (Figure 1 E). Long slender had lower WST-1 cleavage ability per cell than stumpy forms (Figure 1 G), whereas stumpy forms showed negligible luciferase activity compared to long slender forms (Figure 1 F).
The outcome of infection with the monomorphic T. brucei Lister 427 in female mice was then studied. Mice infected i.p. with 10, 100 or 1000 parasites were inoculated i.p. with 20 µg/kg coelenterazine 2–4 minutes before light measurement. All mice died 5 days after infection with 1000 T. b. brucei, while when inoculated with 10 or 100 parasites mice survived up to day 11 after infection (Figure 2A). A fraction of mice survived that inoculation. With exception of one animal, none of the surviving animals showed detectable parasitemia. Light emission, usually located in the peritoneal cavity, was observed in all mice showing positive parasitemia (Figure 2A).
The inoculation of BALB/c male mice with 2.104 pleomorphic luciferase tagged T. b. brucei AnTat 1.1E resulted in a prolonged survival, similar to that observed after infection with the isogenic non-recombinant parasites (data not shown). Mice showed signs of morbidity circa 3 weeks after infection but no increased light emission or parasitemia. The intensity of light emission was not always associated with parasitemia levels. Interestingly, a preferential localization of parasites in the testis was detected in several animals infected with T. b. brucei AnTat 1.1E (Figure 2B), an observation that was reproduced when infecting male BALB/c mice with T. brucei Lister 427 (data not shown). When performing bioluminescence experiments in female mice, no apparent sequestration to the sexual organs (in casu the ovaries) was observed (data not shown). We studied if the pressure exerted by the adaptive immune responses determined preferential localization in the testis. For this purpose we infected B- and T cell-deficient RAG1−/− mice with 100 T. b. brucei Lister 427 recombinant parasites. RAG1−/− mice also showed testis localization of T. b. brucei after infection indicating that testis localization is probably due to parasite tropism for testis or enhanced parasite growth in this organ (Figure 3A).
The immunostaining of trypanosomes in the testis of mice 25 days after infection with T. b. brucei AnTat 1.1E confirmed the information provided by the bioluminescent technique. Trypanosomes were observed within and outside blood vessels, in the interstitial stroma between seminiferous tubules (Figure 3B).
In the experiments described above, biophotonic emission could be mainly detected in the abdominal cavity, and less frequently in the thorax and head of infected animals. Whether such localization was due to a preferential dissemination of the parasite in the abdomen and pelvis or to a non-homogenous distribution of coelenterazine in vivo was investigated. To analyze these possibilities mice infected with T. b. brucei AnTat 1.1E were sacrificed and light production measured in organs after incubation with the substrate ex-vivo. Light was detected in the brain, spleen, lung and testis and to a lesser extent in the liver of infected mice. No light emission was detected in uninfected control animals (Figure 4A). Thus, a non-homogeneous distribution of coelenterazine after inoculation probably accounted for the light production pattern observed in vivo. Hence, we compared the light production after intraperitoneal (ip) and intravenous (iv) inoculation of coelenterazine into mice infected with T. b. brucei AnTat 1.1E. While an abdominal localization of light emission was detected in mice inoculated i.p. (Figure 4B) with coelenterazine, the iv inoculation of the substrate resulted in a thorax and cranial localization, suggesting an incomplete body distribution of the substrate by either route (Figure 4C).
Whether recombinant parasites could be used for testing the efficiency of trypanocidal compounds in vitro was then studied. Light detection and parasite viability at different time points after incubation with the trypanocidal adenosine analogue cordycepin showed similar kinetics (Figure 5A, B). Parasites were also incubated with different concentrations of cordycepin and both, luciferase activity and parasite viability were equally diminished at similar concentrations of cordycepin (Figure 5C).
Subsequently, the luciferase-labeled parasites were used to validate the biophotonic method for testing of trypanocidal compounds in vivo. Treatment with 7 doses of cordycepin and the adenosine deaminase inhibitor deoxycoformycin cures experimental infections with T. b. brucei [15]. A sub-curative treatment with cordycepin and deoxycoformycin in RAG1−/− mice infected with luciferase tagged strains resulted in waning of biophotonic emission (Figure 5D). Several days after treatment, light production and parasitemia were detectable. Some mice showed light production in testis suggesting that T. brucei are protected by the testis-blood barrier from suboptimal doses of trypanocidal drugs (Figure 5D). In contrast, neither parasitemia or light emission were detected in luciferase-tagged T. b. brucei infected BALB/c mice treated daily for 7 days with cordycepin and deoxycoformycin starting 5 days after infection (Figure 6).
In this paper, we demonstrate for the first time the feasibility of detecting live trypanosomes through real-time in vivo and ex vivo luminescence imaging. We opted to use Renilla luciferase rather than firefly luciferase since previous studies in procyclic trypanosomes (insect stage) showed that the firefly luciferase accumulated in glycosomes [16]. This may impede the growth of bloodstream mammalian forms due to major changes in the energy metabolism and thereby hamper in vivo bioluminescence studies (George Cross, personal communication). There may be two reasons why the Renilla luciferase worked so well in vivo (i) the substrate coelenterazine is less polar than d-luciferin, and might pass through the cell membranes more readily and (ii) the C-terminus of Renilla luciferase (VLKNEQ) does not appear to have a peroxisomal targeting sequence, whereas firefly has the classic GGKSKL. Hence, we showed that Renilla luciferase was located in the cytoplasm where the substrate accumulates and does not hamper energy metabolism in the glycosome. A dose of 20 µg coelenterazine was used as a substrate, as described previously for the monitoring of metastasis in mice using bioluminescence [17]. Of importance is the stage-dependent luciferase activity, being significantly lower in stumpy than in slender forms. The transition of slender into stumpy bloodstream forms includes cell cycle arrest and a decrease in protein synthesis, probably due to decrease polysome formation [18] that could account for lower luciferase activity in this life stage. On the contrary, the stumpy forms showed increased WST-1 reduction probably attributed to the increased levels of oxidoreductases in this form compared to long slender forms [19].
Analogous to other models [20], live T. b. brucei produced light after addition of substrate and distinct temporal differences in light production were revealed following intravenous or intraperitoneal delivery. Bhaumik and Gambhir [21] stated in their discussion that biodistribution of coelenterazine and the potential toxicity of repetitively using coelenterazine in living mice should be further investigated. They hypothesized that is was likely that coelenterazine will be accessible to many tissues because of its diffusible nature. We found that repeated injection of coelenterazine did not show toxic effects on the mice. However, the distribution of coelenterazine in vivo seemed not homogenous and depended largely on the way of administering the substrate. This is in accordance with recent findings [22] which showed that intranasal administration of luciferin rather than intraperitoneal injection increased the sensitivity of detecting nasal and pulmonary airway infections by a 30-fold. Hence the route of substrate administration should be considered in the interpretation of the real time images. According to the tissues of interest, either intraperitoneal, intravenous, or a dual injection, should be considered. Another possibility would be to increase the dose of substrate to verify if the local tissue concentration of coelenterazine is sufficient to give a detectable signal. According to the toxicity assays (Figure 1D) it would be possible to increase the dose by a 10-fold.
The lack of toxicity of coelenterazine for parasites and the host at the doses used in vivo, and the lack of light emission by killed recombinant parasites support the strength of luciferase-tagged parasites to study in vivo parasite dissemination as well as drug compound screening, both in vitro and in vivo. The Rluc-pHD309 plasmid integrates at the conserved β-tubulin of the Trypanosoma species, hence other T. brucei strains and taxa can easily be transfected with the Renilla luciferase marker resulting in new models to monitor drug sensitivity and the spread of parasites in a murine model.
A very interesting finding was the abundance of parasites in the testis. T. b. brucei parasites could be observed extravascularly in testis but not in the seminiferous duct, suggesting sexual transmission is unlikely. Accordingly, we observed that no female immunodeficient mice became infected when mated with T. brucei- infected BALB/c mice. Of interest, the natural transmission of Trypanosoma equiperdum closely related to T. brucei, occurs during copulation [23]. The distribution of trypanosomes in the testicular tissue is in accordance with a previous study showing that trypanosomes were present in the intertubular tissues, but never crossed the basal lamina of seminiferous tubules [24]. In that study, necrosis of cells in the seminiferous tubule and a mononuclear infiltration in the interstitium was noted. It might be that in T. equiperdum infections in equines this may contribute to disease transmission. Biophotonic real time detection of parasite dissemination will be useful to study T. equiperdum models to examine tissue tropisms and transmission routes.
We should note that the current model uses intraperitoneal injection which may somehow bias the observed dissemination of the parasites. In future models it may be interesting to perform subcutaneous infections which mimic tsetse delivery.
The possibility that parasites have a preferential tropism for testes can also be of importance when considering drug development, since parasites might be protected from many drugs by the blood-testis barrier. In line with this, parasites were detected in testes upon reactivation of the infection in mice treated with sub-curative doses of cordycepin and deoxycoformycin.
It could be further speculated that the proximity of parasites to Leydig cells located in the interstitial tissues might affect the endocrine balance, contributing both to the pathology of disease. In line with this idea, testosterone levels, testicular responsiveness to exogenous gonadotropin and number of testicular LH receptors were reduced in T. b. brucei infected rats indicating gonadal imbalance [25]. Decreased concentrations of testosterone were detected in patients with human African trypanosomiasis [26]. In accordance, we observed that no offspring was generated after mating 4 male mice 20 days after infection with T. brucei AnTat1.1E with 2 uninfected females each for 10 days. All females remained uninfected, suggesting male sterility and absence of sexual transmission of the parasites. The preference of parasites for the testes does not appear to be a result of immune pressure since it also occurred in RAG1−/− mice, lacking B and T cells.
In conclusion, this bioluminescent model opens new avenues to examine the dissemination of parasites of different Trypanosoma species into different organs, and the in vivo monitoring of drug efficiency.
|
10.1371/journal.pcbi.1002080 | How Haptic Size Sensations Improve Distance Perception | Determining distances to objects is one of the most ubiquitous perceptual tasks in everyday life. Nevertheless, it is challenging because the information from a single image confounds object size and distance. Though our brains frequently judge distances accurately, the underlying computations employed by the brain are not well understood. Our work illuminates these computions by formulating a family of probabilistic models that encompass a variety of distinct hypotheses about distance and size perception. We compare these models' predictions to a set of human distance judgments in an interception experiment and use Bayesian analysis tools to quantitatively select the best hypothesis on the basis of its explanatory power and robustness over experimental data. The central question is: whether, and how, human distance perception incorporates size cues to improve accuracy. Our conclusions are: 1) humans incorporate haptic object size sensations for distance perception, 2) the incorporation of haptic sensations is suboptimal given their reliability, 3) humans use environmentally accurate size and distance priors, 4) distance judgments are produced by perceptual “posterior sampling”. In addition, we compared our model's estimated sensory and motor noise parameters with previously reported measurements in the perceptual literature and found good correspondence between them. Taken together, these results represent a major step forward in establishing the computational underpinnings of human distance perception and the role of size information.
| Perceiving the distance to an object can be difficult because a monocular visual image is influenced by the object's distance and size, so the object's image size alone cannot uniquely determine the distance. However, because object distance is so important in everyday life, our brains have developed various strategies to overcome this difficulty and enable accurate perceptual distance estimates. A key strategy the brain employs is to use touched size sensations, as well as background information regarding the object's size, to rule out incorrect size/distance combinations; our work studies the brain's computations that underpin this strategy. We modified a sophisticated model that prescribes how humans should estimate object distance to encompass a broad set of hypotheses about how humans do estimate distance in actuality. We then used data from a distance perception experiment to select which modified model best accounts for human performance. Our analysis reveals how people use touch sensations and how they bias their distance judgments to conform with true object statistics in the enviroment. Our results provide a comprehensive account of human distance perception and the role of size information, which significantly improves cognitive scientists' understanding of this fundamental, important, and ubiquitous behavior.
| The perception of distances by monocular vision is fundamentally ambiguous: an object that is small and near may create the same image as an object that is large and far (Figure 1A). More precisely, the monocular image size of the object (, visual angle) does not uniquely specify the physical distance (), because and the object's physical size (, diameter) are confounded, . Subjectively we are not usually aware of this visual ambiguity because we perceive object distances unambiguously across a variety of conditions – this work examines how humans perform distance disambiguation by studying whether and how haptic size information is applied to these judgments. Despite previous evidence that adults [1] and infants [2] use object size information, like familiar size, to disambiguate (Figure 1B) the otherwise ambiguous visual information, debate exists [3], summarized by [2]. Recently, Battaglia et al. [4] reported that the brain merges image and haptic sensations in a principled fashion to unambiguously infer distance. Incorporating haptic size information is particularly interesting because it requires sophisticated causal knowledge of the relationship between distance, size, and the multisensory sensations available to the brain to overcome size/distance ambiguity.
Bayesian models provide the exact machinery needed to capture the size-distance perceptual ambiguity, the knowledge required to interpret noisy sensations, and how noisy sensations should be merged with prior knowledge to draw statistically sound perceptual estimates of object distances. This work uses Bayesian models to explicate, test, and confirm/deny a variety of hypotheses about the role of size information in human distance perception. Our results provide a significantly more comprehensive, quantitative account of the underlying computational processes responsible for incorporating size information into distance perception than any previous report.
We formulated a family of Bayesian perception/action models, whose model structure and parameters encoded different assumptions about observer's internal knowledge and computations. We analyzed Battaglia et al.'s [4] data within this context, and used statistical model-selection methods to infer the most probable model and associated parameters for explaining their data.
By committing to a full probabilistic model of observers' sensation, perception, and decision-making processes, we leveraged Battaglia et al.'s [4] data to uncover properties of: 1) the image and haptic sensory noise, 2) the observer's prior knowledge about size and distance, their causal relationship with the sensations, and how they are applied during perceptual processing, and 3) the decision-making strategy by which observers' perceptual inferences yielded psychophysical measurements. Important elements obscured from Battaglia et al.'s [4] original analyses were revealed: the present findings answer four key questions about how size influences human distance perception (described in Model section). Using a full observer model allows us to transcend simplistic debates about whether humans are “optimal vs. sub-optimal” by providing a more textured account of perceptual phenomena that quantifies the sensory quality, what internal knowledge is involved, how they are merged and exploited, and how decisions result. This allows vague questions like “Is perception Bayesian?” to be reformulated into more precise ones like “To what degree does the brain encode uncertainty and apply structured knowledge to perceptual inference?”
Our family of candidate observer models treat the world, observer, and observer's responses as one coherent interrelated physical system, which are represented in the models' structures and parameters using formal probabilistic notation. The fundamental assumptions are that world properties ( and ) generate pieces of sensory evidence, or cues, (, and the haptic size information ), and the observer's perceptual process uses probabilistic (i.e. sensitive to various sources of noise and uncertainty) inference to compute the posterior distribution over the distance given sensory cues, and (Figure 1). The literature [5]–[8] reports many similarities between behavior prescribed by optimal Bayesian inference models, and humans' use of sensory cues, prior knowledge, and decision-making for perceptual inference. The perceptual task used by [4] is well-suited to Bayesian modeling because of important effects of uncertainty and especially the use of auxiliary information (in this case, ) for disambiguating hidden causes (i.e. ). In fact, disambiguation of hidden causes using indirectly-related data is a key, beneficial feature of Bayesian inference, termed “explaining-away” [9]; we hypothesize that human distance perception in the presence of auxiliary size cues is consistent with probabilistic explaining-away.
Battaglia et al.'s [4] experimental task asked participants to intercept a moving ball, and treated their interception distances as perceptual distance judgments. Specifically, participants intercepted the ball as it moved at some distance, after a brief exposure to the ball that in some cases offered the ability touch the ball and feel its physical size and in other cases did not provide explicit size sensations. Our candidate observer models also make distance judgments using the sensory input available to human participants, so a direct comparison between human and model behaviors is possible.
We derived all our candidate models from a base, ideal observer model (IO) that contains internal knowledge about the distributions of sensory noise that corrupt the sensations and , has knowledge about the prior distributions over and , the relationship between , , and , and the relationship between and (Figure 1, lower-right insets, black arrows). In Bayesian parlance these pieces of knowledge fall under the rubric of generative knowledge, or background information about the data's generative process that can aid in inferring the underlying causes. The IO estimates by computing and selecting the that maximizes it (“maximum a posteriori”, MAP, decision rule). This computation requires merging image-size and haptic cues, as well as prior distance and size knowledge, in a manner Bayes' rule prescribes to yield optimal information about (Figure 1's caption illustrates the inference process). We formulated this IO, as well as the other candidate observer models by enumerating all combinations of the following hypothetical questions: 1) Does the observer use the haptic size cue?, 2) Does the observer know the haptic cue's reliability, and integrate the cue appropriately?, 3) Does the observer know the image-size cue's reliability, and integrate the cue appropriately?, 4) Does the observer perform MAP estimation, or rather estimate the distance by averaging a limited number of samples drawn from the posterior? The models were designed to allow standard model-selection methods to decide which hypothetical candidate model, and associated parameters, were best-supported by the experimental human data. Thus we were able to select the most accurate hypothesis, among the field we pre-specified, as the best explanation for how human distance processing uses size information. Moreover, we compared the resultant parameter estimates with measurements reported by other studies, and found they conform with previous findings regarding perception's computational dynamics, which provides independent verification of our conclusions' validity.
Our results indicate humans incorporate haptic size information for distance perception, consistent with Bayesian explaining-away. We also found that all but one participant underestimated the haptic cue's reliability (specifically, they overestimated its sensory noise variance) and integrated the haptic information to a lesser degree than the IO prescribed, similar to the human underuse of auditory information for spatial localization reported by [10]. We found that participants' priors over size and distance were comparable to the experiment's actual random size and distance parameter distributions, implying participants applied knowledge of probable stimulus parameters in their perceptual processing (possibly learned or assumed during the experiment). Last, the sample-averaging estimation model, as opposed to the MAP-estimator, best-accounted for participants' distance judgments, a finding consistent with a growing body of results from perceptual studies that suggest perceptual judgments result from posterior sampling processes [11]–[13].
The observer models have three components: 1) the sensation model describes how the distal stimulus determines the proximal stimulus, 2) the perception model describes how the distal stimulus is inferred from the proximal stimulus, 3) the decision-making model describes how the inferred distal representation guides action.
The scene properties relevant for object distance perception are the object's physical distance and physical size; the relevant sensory cues they generate are visual angle and felt (“haptic”) size. As noted in the Introduction, visual angle is proportional to the ratio of size and distance; so, taking the of each of these variables transforms this relationship into a linear sum (below). Our sensation model uses this -transformed representation for two reasons: 1) Weber-Fechner phenomena support a noise model in which the standard deviation linearly scales with signal magnitude (which can be accomplished with independent noise in log-coordinates), and 2) this -linear approximation is analytically tractable, as we will show. So we assume a linear Gaussian model, meaning the scene properties are a priori Gaussian distributed, and the sensory and motor noise are additive, zero-mean Gaussian, and the sensory generative process is linear, in the log domain.
Log-distance, log-size, log-visual angle, and log-haptic size are represented as: , , , , respectively. The relationship between , , and (by “small angle approximation” to ) is:and between and is:where and represent image-size and haptic sensory noise with standard deviations (SDs) and , respectively. The notation indicates that the parameter represents a property of the scene; this is distinct from the observer's knowledge about the scene, defined in the next section with no tilde.
It follows that the distribution of sensory cues conditioned on the scene properties are:(1)(2) We assume observers' internal prior probabilities over and are:
Battaglia [14] derives model observers for perceptual inference in linear Gaussian contexts under a variety of assumptions – we co-opt the “explaining-away” derivations (Sec. 3.4 in [14]) for the current size/distance perception context. All model observers are assumed to use their knowledge of the world, i.e. the sensory noise ( and ) and prior distributions (, , , and ), to compute beliefs about . These beliefs are represented as the posterior distribution, (which is Gaussian):(3)where,(4)
For those familiar with “standard” cue combination, Eqs. 3 and 4 are similar to the “optimal cue combination” formulae in [15], and in fact by looking closely at the Bayes' net in the lower right of Fig. 1B, one can see that the subgraph composed of variables , , and represents the standard two-cue “cue combination” situation. However, our present situation is distinct from [15] because we focus on data fusion in conditions where one cue () is only indirectly related to the desired property () by its ability to disambiguate another cue (). The intuition for the weights in Eq. 4 is as follows. Because provides information about to improve inference of , the numerator of assigns sensory cue more influence when prior knowledge of and are weaker (higher and ). Similarly, 's numerator dictates that is more influential when information about is weaker (higher ). Interpreting is less straightfoward, but essentially holds that when information about is poor, because both the prior over and sensory cue are weak (higher and ), then is more exclusively influential for inferring , whereas if either prior knowledge about or sensory cue are strong, and that information jointly guide inference of . Last, 's numerator assigns stronger influence to prior knowledge of only when the sensory cues and prior knowledge of are weak.
Human observers who do use for distance perception are modeled above by Eq. 4. The hypothesis that observers do not use , either because is unavailable or because they are not capable, is formulated:(5)where,(6)
Eq. 5 is algebraically equivalent to taking in the formulation in Eq. 4. Whether humans do (Eq. 3) or do not (Eq. 5) use to make distance judgments is the first of our hypothesis questions (see Table 1). Also, whether humans know the true sensory noise magnitude i.e. whether they use vs. , and/or vs. , are the second and third of our hypothesis questions (Table 1).
The model observer uses beliefs about to select a position at which to intercept the moving ball. We assume that participants attempt to minimize the difference between their judged distance and the true distance, which for Gaussian distributions may equivalently correspond to minimizing a MAP, mean-squared, or symmetric Heaviside loss functions. However accessing their perceptually-inferred information about is not necessarily trivial: we consider that they may select the maximum probability , i.e. (or )), as their judgment of distance, or instead draw a number, , of independent samples from or and compute their sample mean as a judgment, is the fourth (and last) of our hypothesis questions (Table 1). These distinct models may imply different neural representations for posterior beliefs about distance, which we address in the Discussion.
Additionally our models all include an element of motor noise, the small degree of error between judged and the experimentally-measured , due to motor imprecision when performing an interception. For consistency with known parameters of motor control, we selected an additive, Gaussian motor noise term , that was added to the distance judgment to form .
We combine the sensation, perception, and decision-making models described above to define a set of coherent model observers that input sensations, combine them with internal knowledge to form beliefs about distance, and form decisions that are output as interception responses in the experimental task.
By varying the models' structure and parameters we encoded the four hypothesis questions in the Introduction (subsequently referred to as “Q I, II, II, IV”) to form the candidate observer models (Table 1):
In total 12 distinct candidate models spanned the possible combinations of the four questions (the reason the total is 12, instead of 16, is because for candidate models that do not include the use of haptic information [Q I], the question of whether the observer knows the haptic cue noise magnitude or not [Q II] is inconsequential and those models are redundant).
First, we describe how the model observers predict responses in the experimental interception task and illustrate responses produced by each model. Second, we describe how the model's parameters were inferred given each participants' response data. Third, we show how we computed the human data likelihood under each model and how we quantitatively compare them to determine which model provides the best account of the human data.
The central result of our study is quantitative selection of the model that best explains the data, which we determine by comparing the models' DIC scores, to answer the four hypothetical questions posed above. Figure 5A shows raw DIC scores, and Figure 5B shows the difference between the best model's DIC (indicated by circle on x-axis) and the other models' DIC scores. We defined DIC significance as described in the previous subsection: models whose DIC differed by greater than 10 were deemed “sigificantly” different (dashed horizontal line and * in Figure 5B) and greater than 15 deemed “highly significantly different” (solid horizontal line and ** in Figure 5B); this is a conservative modification of the criteria mentioned in [17].
We found that all participants incorporate haptic size information to make their distance judgments (Q I). Also, we found 5 of 6 participants misestimated their haptic size noise and thus incorporated the haptic information less than optimally prescribed, while one participant applied the haptic cue in proportion to its reliability (Q II); the following section addresses the nature of the misestimation. All participants incorporated the visual image-size cue optimally, in accordance with its noise magnitude (Q III). All participants used a sample-averaging strategy over MAP decision-making (Q IV). With respect to Q IV, the DIC scores were always worse for the MAP model versus the sample-averaging model, by an average DIC difference of 129 (Figure 5), so we exclusively focus on the sample-averaging models (odd numbers) for the remaining discussion. Figure 5 depicts each participant's DIC scores for each sample-averaging model, the left graph shows the absolute DIC values and the right graph shows the differences between best model DICs and the other models' DICs. Participant 6 was an author. Participant 3's DIC differences between Model 7 and Models 5 and 11 was not significant under our conservative criteria, however Model 7 was still better by DICs of and , respectively, which is considered marginally significant under typical uses of AIC/DIC [17].
Participant 5, the only participant whose DIC favored the hypothesis that the haptic noise magnitude was correctly known (Q II), had the worst DIC scores across participants, as well as substantially different parameter value estimates from the other participants (see next paragraphs). Upon closer inspection of participant 5's data, it was qualitatively the noisiest: in Battaglia et al.'s [4] simple regression analysis of this data their statistical analysis determined participant 5's data was so significantly different from the other participants' that it ought to be excluded as an outlier. The reason we included it in the current analysis was to determine whether there was still some patterns the previous analysis had not detected. Though the parameters still yield meaningful values, because of the major differences between raw DIC scores, the DIC-favored model, the parameter value estimates, and the general noisiness of the response data, we strongly suspect this participant either was not focusing on performing this task, was randomly selecting answers on a large fraction of the trials, and in general should be distinguished in further analysis due to these aberrations: so, we report participant 5's parameter estimates separately from the other “inlier” participants.
Figure 6 shows Participant 1's model-predicted compared against the actual values, for the best model, 7, as well as several that differ by one assumption (Table 1). The spread in the dots is due to sensory noise and the random posterior sampling process, how neatly the actual data falls within the ranges predicted by a particular model (black error bars) is indicative of the model's explanatory quality. Notice the pattern of more varied no-haptic vs. haptic in Models 7, 11, 5, a direct prediction of the sampling models over MAP. Though MAP decisions incur more bias in the no-haptic condition, they actually have less trial-to-trial variance than in the haptic condition. This is due to the fact that the prior does not vary between trials, while the more informative haptic cue does.
A possible concern is that participants learned to use the haptic cue during the course of the experiment, and that the weak DIC scores of Models 1–4 in comparison to Models 5–12 actually reflect the effects of associative learning rather than knowledge the participants brought into the experiment. We evaluated this possibility by performing the same DIC analysis on data from only the first day to test whether Models 5–12 were still favored over their 1–4 counterparts. The results unequivocally confirm the results on the data from the final 3 days above: for every participant, the DIC analysis across the models shows that the no-haptic models (1–4) have worse DIC scores than their haptic model counterparts (5–12). The best no-haptic models' DICs are below the best haptic models' DICs by margins of {, , , , , } for Participants 1 through 6, respectively. In fact, removing the sampling models, even the no-haptic models (1 and 3) with the best DIC scores still have worse scores than the haptic models (5, 7, 9, and 11) with the worst DIC scores. This firmly supports conclusion that the haptic cue is used even on the first day of trials.
Though it might seem that given the 6 to 10 “free” parameters in our general observer model, we could “fit” any data, we are actually inferring the best parameters and using the posterior's expected values rather than the most probable a posteriori parameters. Moreover DIC acknowledges the possibility for overfitting and counters it by penalizing overfits through the complexity term, thus affirming that the chosen model's structure and parameters are accurate and robust explanations of the humans' judgments. Moreover because we encoded different hypotheses within the models we could clearly distinguish those hypotheses best-supported by the data. Lastly, despite the possibility that we could fit a variety of data, the remainder of this section shows that the individual inferred parameter values are consistent with known perceptual parameters measured in other studies.
A secondary result of this work, beyond providing answers to the 4 hypothetical questions, is that the inferred parameter values () our analysis yielded can be meaningfully interpreted. Though there is no guarantee that the inferred parameters are unique, they offer an indication of what the analysis finds probable. All reported parameters are MCMC expections, from which we compute meansSEs across participants and report the values in log coordinates. First, we present the SDs in terms of Weber fractions for the sensory noise, with discrimination thresholds corresponding to .
The image-size noise SD, , and assumed noise SD, , were coupled in the best-fit models (7 and 11) for all participants. Their values correspond to Weber fractions of to (meanSE of ) for the inlier participants, and for participant 5. This is comparable to the Weber fractions of measured in humans by [18] for parallel line separation discrimination, and by [19] for line length discrimination. Because our task did not involve interval-wise discrimination of pairs of stimuli, but rather absolute perception, it is to be expected that our noise magnitudes will be slightly higher.
The haptic noise SDs, , and assumed haptic noise SD, , were uncoupled in the inlier participants' best-fit model (7), and coupled for participant 5's best fit model. The inlier participants' haptic noise SDs correspond to Weber fractions of between and (meanSE of ), and for participant 5. A Weber fraction of was measured in humans by [20] for haptic size discrimination of objects between and mm in width using a similar haptic stimulus presentation apparatus, but with two fingers gripping the object rather than one finger probing the size. Because two fingers are likely to provide a more precise size measurement and because their participants performed interval discriminations of pairs of objects, our somewhat elevated Weber fraction are reasonable values. The inlier participants overestimated their haptic noise SDs, with their assumptions corresponding to Weber fractions of to (meanSE of ). The consequences of overestimating haptic noise are that the observers do not achieve the level of disambiguation possible by fully incorporating the haptic cue, and apply prior knowledge about the ball's size and distance relatively more heavily (Figure 3).
Our analysis provided information about the observer models' prior knowledge, and found it strikingly similar to the sample statistics of the experimental stimuli's distances and sizes, with slightly higher SDs (remember the stimuli were uniformly distributed in the mm domain). The meanSE estimated prior distance mean and SD parameters, and , across all participants were and log-mm, respectively; the experimental distance mean and SD were and log-mm, respectively. The mean estimated prior size mean and SD parameters, and , across participants were and log-mm, respectively; the experimental size mean and SD were and log-mm, respectively. This indicates participants learned the range of possible stimuli presented in the experiment and applied that knowledge toward improving their judgments, to the effect of lowering the posterior variance (Figure 3). To further investigate the source of participants' prior knowledge, we ran our full analysis on only the first day of participants' trials, to measure what difference between inferred parameters exist between early and later in the experiment. We found that participants' first-day priors for and were and log-mm, respectively; and, participants' first-day priors for and were and log-mm, respectively. So, the prior means did not shift significantly (in terms of SE interval overlap), but the prior SD values did. It appears that participants rapidly learned the prior means, which are more easily estimable from experience and also may be assumed to some extent (the true prior distance mean is at the center of the virtual workbench, and the balls' sizes were directly observable in haptic condition trials). However, participants appeared to use more diffuse prior and parameters early in the experiment, which is consistent with making weaker prior assumptions about the range of distance/size variation (top row of Figure 3).
Our analysis provided estimates of participants' motor noise SD, whose meanSE across participants was log-mm, which amounts to a SD of mm at a reach distance of mm, and mm at reach distance of mm, the extremal distances presented in the experiment. A value of log-mm was reported [21] under similar reaching conditions.
The sample-averaging models generally outperformed the MAP estimate models (Q IV) with respect to DIC scores. The inlier participants had values between and (meanSE of ), and participant 5's estimate was . Of course in our model must be integer-valued, but these real valued estimates are robust means across our MCMC analysis samples. An alternative interpretation of is that it is an exponent applied to the posterior distribution, from which one sample is then drawn after renormalizing. For a Gaussian distribution, because , drawing sample, , from yields a Gaussian-distributed: . And, drawing samples, , from the unexponentiated and averaging yields a sample mean with the same distribution:
Between the DIC analysis and the validity of the inferred parameters, we conclude that model 7 is both structurally and parametrically accurate. This strongly supports model 7 and its encoded hypotheses as a coherent computational account of the underlying processes responsible for size-aided distance perception.
We conclude that humans can use haptic size cues to disambiguate and improve distance perception, but that the degree to which they incorporate haptic size information is lower than the ideal observer prescribes. We also conclude that the distance responses are best explained as a process of drawing several samples from the posterior distribution over distance given sensations, and averaging them to form a distance estimate. This behavior is broadly consistent with a Bayesian perceptual inference model in which mistaken generative knowledge about haptic cues is used, and beliefs about distance are accessed by drawing samples from an internal posterior distribution.
The brain's use of sensory cues for disambiguating others has been reported in a variety of perceptual domains, and broadly falls under the category “perceptual constancy”. Constancy effects, like the present distance constancy, involve situations in which an observer cannot unambiguously estimate a scene property due to confounding influences from other “nuisance” properties, and so leverages “auxiliary” cues (in this study, haptic size) to rule out inconsistent possibilities. Auxiliary disambiguation effects, like constancy, have other names in the literature, like “cue promotion” [22], “simultaneous contrast” [23], and “taking-into-account” [24]. Many studies have reported “size constancy”, distance cues disambiguating object size perception [25]–[32], so it is not entirely surprising that size cues can conversely disambiguate distance perception.
Humans underestimating non-visual cue reliabilities and thus integrating them less strongly has been measured before by [10], [32]. There are several potential reasons for this phenomenon, one idea that has recently garnered support [33]–[36] is that sensory cues are used in accordance with their causal relationships to the unobserved scene properties: when the brain believes cues are unrelated to the desired scene property, it down-weights or outright ignores them. In the present study, this would mean the brain is unwilling to fully apply the haptic size cues because they might originate from a source independent of the ball, for instance imagine the hand touched a ball behind a photograph of a different ball; of course, such miscorrespondences are uncommon in nature, but examples like “prism adaptation” demonstrate the brain can accommodate and recalibrate in such situations. Another possibility is non-visual cues to spatial properties may be experienced far less frequently in life, and had fewer opportunities on which to be calibrated, so they are mistrusted.
Our finding that all our observers' responses are best modeled as sampling the posterior is consistent with recent studies and ideas about the representation and computation of probability in the brain. Using posterior sampling to generate responses in a choice task should manifest as probability matching of the options, a common finding in many behavioral tasks, including a perceptual audio-visual cue-combination task [13]. Sampling has also been used to provide a novel explanation for perceptual switching to multistable displays [11], [37]. Moreover, sampling provides an interpretation of neural activity in population codes and makes difficult probabilistic computations simple to neurally implement (see review by [38]).
Although Bayesian decisions are usually modeled as maximizing the posterior, maximization is not the best decision rule in all instances. MAP's optimality depends on both the task and the veridicality of the decision maker's posterior distribution. MAP assumes the decision maker's goal is to maximize the number of correct responses and that the posterior is based on the correct generative model for the data. When the posterior is not correct, basing responses on sampling provides exploration that can be used to improve the decision maker's policy. This idea has been extensively explored within reinforcement learning, where exploration is frequently implemented using a softmax decision strategy [39] where choices are stochastically sampled from an exponentiated distribution over the values of a set of discrete options. This idea can be generalized to the case of continuous decision variables. The value of an estimate is based on the reward function for the task. In our decision task, participants were “correct” whenever their choices fell within a narrow region relative to their posterior distribution. Approximating the experimental reward function as a delta function, the optimal strategy is to maximize the posterior. However, if we need to improve our estimate of the posterior, then it is important to estimate the error. Sampling from the posterior gives a set of values that can be used to compute any performance statistic, making it a reasonable strategy when an observer is needs information needed to learn - i.e. to assess and improve performance.
Though our models posit observers draw samples directly from the posterior and averaging, any decision rule that is sensitive to the posterior variance may produce similar predictions – for instance, it is possible that participants internally exponentiate the posterior and draw exactly one sample (detailed in Results). This means that for greater exponents, the posterior is more greatly sharpened; as the exponent approaches infinity, the posterior approaches a delta function located at the MAP estimate (after re-normalizing). This is a general strategy used in many machine learning domains to transition neatly between posteriors, MAP estimates, and “watered down” versions of the posterior. However we find this account unappealing because it implies that drawing more than one samples is less attractive to the observer's underlying perceptual mechanics than performing posterior exponentiation. Also, though our models assume posterior sample-averaging is a source for behavioral response variance (Figure 4), another possibility is that observers have uncertainty in the parameter values that characterize their generative knowledge itself, and actually draw samples of generative parameters instead of using deterministic parameter estimates. For instance, when combining haptic cues they may sample from an internal distribution over haptic reliability (). This could be a strategy for learning when the brain is uncertain about internal generative model parameters; because the observer receives feedback, and presumably wishes to calibrate the internal perceptual model, varying behavior by using different samples of internal model parameters avoids redundant feedback associated with similar behavioral responses to similar input stimuli.
Using a full probabilistic model of observers' sensation, perception, and decision-making processes provide us with answers to the four key questions we posed in the Model section. This study's analysis of data reported by [4] resulted in a much more comprehensive account of the computations responsible for distance and size perception. By formally characterizing a set of principled computational perception hypotheses, and choosing the best theoretical account of the measured phenomenology using Bayesian model selection tools, we demonstrated the power, robustness, and flexibility of this coherent framework for studying human cognition, and obtained deeper understanding of distance perception.
|
10.1371/journal.ppat.1005854 | Interaction of the Human Papillomavirus E6 Oncoprotein with Sorting Nexin 27 Modulates Endocytic Cargo Transport Pathways | A subset of high-risk Human Papillomaviruses (HPVs) are the causative agents of a large number of human cancers, of which cervical is the most common. Two viral oncoproteins, E6 and E7, contribute directly towards the development and maintenance of malignancy. A characteristic feature of the E6 oncoproteins from cancer-causing HPV types is the presence of a PDZ binding motif (PBM) at its C-terminus, which confers interaction with cellular proteins harbouring PDZ domains. Here we show that this motif allows E6 interaction with Sorting Nexin 27 (SNX27), an essential component of endosomal recycling pathways. This interaction is highly conserved across E6 proteins from multiple high-risk HPV types and is mediated by a classical PBM-PDZ interaction but unlike many E6 targets, SNX27 is not targeted for degradation by E6. Rather, in HPV-18 positive cell lines the association of SNX27 with components of the retromer complex and the endocytic transport machinery is altered in an E6 PBM-dependent manner. Analysis of a SNX27 cargo, the glucose transporter GLUT1, reveals an E6-dependent maintenance of GLUT1 expression and alteration in its association with components of the endocytic transport machinery. Furthermore, knockdown of E6 in HPV-18 positive cervical cancer cells phenocopies the loss of SNX27, both in terms of GLUT1 expression levels and its vesicular localization, with a concomitant marked reduction in glucose uptake, whilst loss of SNX27 results in slower cell proliferation in low nutrient conditions. These results demonstrate that E6 interaction with SNX27 can alter the recycling of cargo molecules, one consequence of which is modulation of nutrient availability in HPV transformed tumour cells.
| A unique feature of the high risk Human Papillomavirus (HPV) E6 oncoproteins is the presence of a PDZ binding motif (PBM) on its extreme C-terminus. This motif confers on E6 an ability to interact with a number of cellular proteins which possess PDZ domains, and this activity of E6 is important during the viral life cycle and contributes towards HPV-induced malignancy. In this study we describe a novel activity of high risk HPV E6 oncoproteins involving the direct regulation of endocytic transport pathways. This activity is dependent upon the E6 PBM and involves interaction with the endocytic cargo sorting machinery via sorting nexin 27 (SNX27). One of the consequences of this interaction is a redistribution of SNX27 with respect to components of the retromer complex and this in turn affects the composition of the endocytic transport machinery. This impacts directly upon rates of cargo recycling and in the case of HPV transformed cells, contributes towards maintaining high levels of glucose uptake. This study therefore describes a new function for the E6 oncoproteins and sheds light on how HPVs can modulate endocytic transport pathways.
| Human Papillomaviruses (HPVs) are the causative agents of a large number of human malignancies, chief among which is cervical cancer, with over 500,000 reported cases worldwide annually [1,2]. There are currently more than 150 known types of HPVs, but not all of them are etiological agents of carcinomas. The cancer-causing HPVs are classified as “high-risk” types and these include HPV-16 and HPV-18, among others [3]. A hallmark of HPV induced-malignancy is the continued expression of the viral oncoproteins E6 and E7 throughout the course of tumour development [4,5]. Inhibiting the expression of either oncoprotein in cells derived from cervical tumours results in cell growth arrest and induction of apoptosis, demonstrating a continued requirement for E6 and E7 in the maintenance of the transformed phenotype [6]. Both viral oncoproteins act cooperatively, where E7 reprograms the infected cell to enter S phase by targeting, in part, the pRb family members, thus allowing the E2F family of transcription factors to transactivate various cell cycle genes [7–9]. The E6 oncoprotein complements the action of E7 by curbing the cell’s pro-apoptotic response to unscheduled DNA replication and targets pro-apoptotic proteins such as p53 [10] and Bak [11] for proteasome-mediated degradation via the action of the E6AP ubiquitin ligase [12]. However the ability of both E6 and E7 to contribute to cancer development depends upon a large number of other important interactions. In the case of the high-risk E6 oncoproteins a typical example is interaction with cellular PDZ (PSD-95/DLG/ZO-1) domain containing proteins.
A unique characteristic of the cancer-causing E6 oncoproteins is the presence of a PDZ binding motif (PBM) on their carboxy termini [13]. An intact E6 PBM is important for the ability of E6 to cooperate with E7 in the generation of tumours in transgenic mouse models, and also has transforming potential in some tissue culture models [14–16]. In the context of the whole viral genome, loss of E6 PBM function results in a defective replicative life cycle, with reduced levels of viral DNA amplification and, ultimately, loss of the viral episomes [17,18]. A large number of cellular PDZ domain-containing targets of E6 have been reported, with some of the best-characterised being a group of proteins involved in the regulation of cell polarity [19]. These include the Discs Large and Scribble proteins, which are key regulators of cell polarity and potential tumour suppressor proteins [20,21]. In addition, MAGI-1 would also appear to be a relevant target of E6, affecting the integrity of tight junctions in HPV-positive cells [22]. Recent high throughput screens have identified many further potential PDZ domain-containing targets of E6, with Sorting Nexin 27 (SNX27) being one such intriguing candidate [23,24].
A critical element in epithelial organisation is the regulation of polarity, and whilst key elements of the pathway, such as Scrib and Dlg, are well defined, it is clear that endocytic transport pathways also play important roles in epithelial organisation and polarity control [25–27]. Within these pathways, the retromer plays an important part in the transport of proteins from endosomes to the trans-Golgi network [28]. The retromer is an oligomeric protein complex formed by a Vacuolar protein sorting (Vps) subcomplex, consisting of Vps26, Vps29 and Vps35, and a heterodimer of sorting nexins (SNXs). SNXs are proteins characterized by the presence of a phosphoinositide binding PX domain, which targets them to phosphatidylinositol-3-monophosphate-rich membranes of the endosomes [29]. Many members of the SNX family also contain a number of protein-protein interaction domains, through which they interact with their respective cargoes and ensure appropriate trafficking [29–32]. Of particular interest from the HPV E6 point of view is SNX27. It contains a C-terminal Ras Association/FERM like domain, which has been shown to associate with the Ras GTPase [33], and SNX27 is the only SNX that contains a PDZ domain [30,32,34]. Recent proteomic studies have indicated that HPV-16 and HPV-18 E6 are potential interacting partners of SNX27 [23,24]. The SNX27 PDZ domain mediates the interaction with various PBM-containing cargoes and is thought to regulate the trafficking of these proteins through the endosomal pathway [35–38]. SNX27-mediated recycling of PBM-containing cargoes from endosomes to the plasma membrane is dependent on its interaction with the retromer, and recent studies have shown that the SNX27-retromer link is crucial for the retrieval and recycling of various transmembrane proteins. These include proteins required for maintenance of cellular homeostasis and growth, among which is GLUT-1, which is essential for glucose uptake [39,40]. We therefore initiated a series of studies to investigate whether E6 can associate with SNX27 and, if so, to ask what are the potential implications of this interaction in the recycling function of the SNX27-retromer complex in the context of GLUT-1 as a cargo.
Proteomic screens for cellular interacting partners of HPV E6 identified the PDZ domain containing protein SNX27 as a potential novel target of the high risk HPV E6 proteins [23,24]. To verify if this is indeed a bona-fide interaction, a series of in vitro binding assays were performed. The E6 proteins derived from the high-risk HPV-16, 18, 31, 33, 51 and 58 were expressed as GST fusion proteins and purified. These were then incubated with in vitro translated and radiolabelled SNX27 for 2 hours at 4°C. MAGI-1, which has been shown previously to be a strongly bound PDZ domain-containing target of HPV-16 and HPV-18 E6 was used as a positive control [41]. After thorough washing of the beads, the level of binding was determined by SDS-PAGE and autoradiography. The results in Fig 1A show that SNX27 is indeed a strong interaction partner of all the high-risk E6 types tested.
The E6 proteins from low-risk HPV types lack PBMs, therefore we wished to determine whether the ability of E6 to interact with SNX27 was restricted to high-risk types. To do this, pull-down assays were performed using purified GST-HPV18 E6 and GST-HPV11 E6 fusion proteins with in vitro translated and radiolabeled SNX27. The results in Fig 1B demonstrate very strong association between HPV-18 E6 and SNX27, with much weaker, albeit still detectable, association with HPV-11 E6. To determine whether the interaction between the high-risk E6 proteins and SNX27 required the E6 PBM, the assay was repeated using wild type HPV-18 E6 and the HPV-18 E6 T156E mutation, which has been shown previously to abolish PDZ recognition [42]. The results in Fig 1C demonstrate that the T156E mutation dramatically reduces the ability of E6 to interact with SNX27, although some residual interaction is still detectable. Taken together these results demonstrate that HPV-18 E6 interacts with SNX27 primarily through its PBM, although other residues within E6 are also most likely to be involved, and this is reflected in a weak degree of association between HPV-11 E6 and SNX27.
Since SNX27 is a PDZ domain-containing protein, we wanted to determine if the interaction between SNX27 and E6 is through PBM-PDZ recognition. To do this, various C-terminal mutants of HPV-18 E6 [42] were translated and radiolabelled in vitro and used in pull down assays with purified GST-SNX27. The results in Fig 2A show that the mutation V158A, which destroys an essential part of the HPV-18 E6 PBM [42], greatly reduces the interaction with SNX27. Furthermore, the E155A mutation also severely compromises E6 recognition of SNX27, whilst the other mutants had either minimal effects or increased the level of association. This differential contribution of specific residues within the E6 PBM is in agreement with how E6 recognises other PDZ domain-containing substrates [43]. Taken together, these results demonstrate that HPV-18 E6 recognises SNX27 via sequences within the E6 PBM although, in agreement with the T156E mutation, other residues are also likely to be involved.
To confirm that the PDZ domain of SNX27 is necessary for E6 binding, PDZ deletion mutants of SNX27 were used in a series of pull down assays with GST-HPV18 E6 fusion proteins. As can be seen from Fig 2B, wild type SNX27 binds to HPV-18 E6 very strongly, but this interaction is greatly reduced with the HPV-18 E6ΔPBM mutant. The deletion of the retromer binding region of SNX27, spanning amino acids 67–77 [44] in the outer extended loop of the SNX27 PDZ domain, does not affect the interaction with HPV-18 E6. Interestingly, deletion of amino acids 97–107 and 113–121 in the core of the SNX27 PDZ domain markedly reduces the interaction with wild type HPV-18 E6, but does not abolish it. When the SNX27 mutants are used in binding assays with the E6ΔPBM mutant, interaction is found at a level similar to that seen with the wild type HPV-18 E6. These results demonstrate that the major site of recognition between HPV-18 E6 and SNX27 involves a classic PBM-PDZ interaction, although a second weaker interaction site also exists.
Many E6 interacting partners are targeted for ubiquitin-dependent proteasomal degradation. However, in an extensive series of in vitro and in vivo assays, we found no evidence that SNX27 was a degradation substrate for either HPV-16 or HPV-18 E6. Previous studies have reported E6 to be largely expressed in the nucleus, whilst SNX27 exists primarily within the early endocytic compartments in the cytoplasmic and membrane compartments of the cell (35). We were therefore first interested in determining whether HPV-18 E6 and SNX27 were present within similar cellular compartments. To do this we performed cell fractionations of HPV-18 positive HeLa cells, in the presence and absence of the proteasome inhibitor MG132 to rescue any pools of the two proteins that might be subject to proteasome-mediated degradation. In this analysis we also made use of a DOX-inducible shRNA to ablate SNX27 expression, to ascertain whether SNX27 might have any effect on the pattern of E6 expression. The results obtained are shown in Fig 3. As expected, the bulk of SNX27 is found within the membrane fractions of the cell (35). Most interestingly, the bulk of HPV-18 E6 in HeLa cells is also localised within the membrane compartment, suggesting a potentially similar location to SNX27, although loss of SNX27 has no major effect on the subcellular distribution of E6. In addition, the membrane-bound pool of E6 appears to be unaffected by proteasome inhibition, whilst there is a clear evidence of proteasome mediated regulation of E6 levels in the cytoplasmic, nuclear and cytoskeletal fractions.
Having found that E6 and SNX27 exist within the membrane fraction of HeLa cells, we were then interested in determining whether E6 could in any way modulate the subcellular localisation of SNX27. To do this, we analysed the distribution of SNX27 in HPV-18 positive HeLa cells, in the presence and absence of HPV-18 E6. The cells were transfected with control siRNA against Luciferase and siRNA against HPV-18 E6/E7 and E6AP as an alternative means of reducing E6 levels of expression [45]. After 72h, the cells were fixed and immunofluorescence analyses performed to detect SNX27, and also p53 as a control for the silencing of E6/E7 expression. We also analysed the distribution of the retromer component Vps35. The results in Fig 4 show a diffused speckled pattern of SNX27 distribution in control transfected cells, in agreement with previous studies (35), but a marked perinuclear accumulation, with increased levels of co-localisation with Vps35 following silencing of E6/E7 or E6AP, where a strong nuclear p53 staining can also be observed.
Taken together, these results suggest that E6 contributes to the maintenance of a dispersed pattern of Vps35 and SNX27 expression in HeLa cells, and that its loss results in a significant increase in Vps35/SNX27 co-localization at sites close to the nuclear membrane.
To determine whether this change in the distribution of SNX27 is related to the ability of E6 to interact with SNX27, we analysed the distribution of SNX27 in the immortalised keratinocyte cell line NIKS, expressing either wild type HPV-16 E6 or a HPV-16 E6ΔPBM mutant. The cells were fixed and stained for SNX27 and Vps35 as described above. The results obtained in Fig 5 show significant co-localisation of Vps35 and SNX27 in the control NIKS not expressing E6, whilst this becomes much more dispersed in the wild type HPV-16 E6 expressing cells with limited areas of co-localisation. In contrast, the pattern of SNX27 and Vps35 distribution in the HPV-16 E6ΔPBM cells, shows increased levels of co-localisation of SNX27 and Vps35, reflecting more the pattern seen in the control NIKS. These results indicate that changes in the subcellular distribution of SNX27 and Vps35 in the presence of HPV E6 are in part dependent upon on intact E6 PBM.
The above results indicate that E6 expression can alter the subcellular distribution of SNX27, and can affect the degree of its co-localisation with components of the retromer. We next wanted to determine whether E6 could modulate the degree of interaction between SNX27 and Vps35, and, furthermore, whether E6 could also be found in the SNX27/Vps35 complex in vivo. To do this, co-immunoprecipitation assays were performed where endogenous E6 from HeLa cells was immunoprecipitated using an anti-HPV-18 E6 monoclonal antibody. The immune complexes were then adsorbed on Protein A Sepharose beads and analyzed by SDS-PAGE followed by Western blotting for E6, SNX27 and Vps35. As can be seen from Fig 6A (left hand panel), both SNX27 and the Vps35 component of the retromer co-immunoprecipitate with E6, indicating that endogenous E6 is found in close proximity with the retromer. Likewise, HPV-18 E6 also co-immunoprecipitates with SNX27 when anti-SNX27 antibodies are used, and the specificity of this interaction is further demonstrated by loss of signal when the cells are transfected with siRNA against E6/E7 (Fig 6B). In order to determine whether E6 can affect association between SNX27 and Vps35, we performed an SNX27 immunoprecipitation from HeLa cells, which had been previously transfected with control siRNA luciferase or siRNA E6 to ablate E6 expression. The results in Fig 6A (right hand panel) show no major difference in the levels of SNX27/Vps35 interaction in the presence or absence of HPV-18 E6.
Previous studies have shown that the glucose transporter GLUT-1 requires both SNX27 and Vps35 for its efficient recycling back to the cell surface, with loss of SNX27 resulting in a marked decrease in the levels of GLUT-1 expression owing to its enhanced lysozomal degradation [40]. Since we observed that loss of E6 leads to alterations in the distribution of SNX27 and Vps35, we wanted to determine whether this might be reflected in changes in expression of a well-defined SNX27 cargo. To investigate this, we performed a series of assays to directly monitor the levels of expression of GLUT-1 in HeLa cells following ablation of either HPV-18 E6 or SNX27 expression. Cells were transfected with siRNA against Luciferase or siRNA against E6, or were treated with DOX to induce the SNX27 shRNA for 72h. After this time the cells were harvested and the levels of GLUT-1 expression analysed by western blotting. As a control for E6 ablation we monitored p53, and SNX27 was also analysed to ensure efficacy of the inducible shRNA. The results in Fig 7 demonstrate that GLUT-1 migrates as multiple bands, which is in agreement with previous studies. Most importantly however, there is a clear reduction in the levels of expression of the higher molecular weight, functional forms of the protein [46, 47], following loss of E6 or SNX27. These results demonstrate that removal of E6 in HeLa cells has a very similar effect upon GLUT-1 expression levels as removal of SNX27.
The above results demonstrate that GLUT-1 levels are reduced in HeLa cells lacking either E6 or SNX27. We were therefore interested in determining whether E6 had any influence on the composition of the endocytic compartments and their associated cargoes. To do this HeLa cells were transfected with siRNA against E6 or Luciferase for 72h and cell extracts subsequently fractionated on 5–25% OptiPrep gradients as described previously [48]. The fractions were then collected and analyzed by Western blotting. As can be seen in Fig 8, there are marked changes both in the distribution of cargoes and in the endocytic compartments following E6 ablation. In control transfected cells, the early endosome marker Rab4 appears to follow very closely the distribution of both SNX27 and Vps35. However, upon loss of E6 there is a clear shift in the distribution of both Rab4 and SNX27 within the gradient, whilst Vps35 appears to be relatively unchanged. Likewise, there is a corresponding marked shift in the distribution of the SNX27 cargo GLUT-1, and there is also an overall reduction in the levels of expression of the higher molecular weight forms of GLUT-1, similar to that seen in Fig 7. Interestingly there is no significant alteration in the mobility of the late endosomal marker LAMP2. These results demonstrate that in HeLa cells loss of E6 has quite a profound effect on the distribution of SNX27 within the cargo transporting apparatus, in support of the notion that E6 can modulate SNX27 function.
The above data suggest a significant shift in the distribution of several endocytic compartments and of at least one associated cargo, GLUT-1, following the ablation of E6 expression in HeLa cells. We were therefore interested in determining whether this alteration in the pattern of GLUT-1 expression could be visualised by immunofluorescence analysis. Cells were transfected with siRNA against Luciferase or E6 and, after 72h, analysed by immunofluorescence for GLUT-1 and Vps35 localisation, with p53 serving as a positive control for E6 ablation. As can be seen from Fig 9B, most of the GLUT-1 appears to be localized on the cell periphery in the control cells, while the Vps35 remains distributed throughout the cytoplasm as seen before. In contrast, upon E6 ablation, the majority of GLUT-1 is located in intracellular vesicle-like structures with a pool showing significant co-localization with Vps35. Similar results are also obtained in SiHa cells following ablation of E6AP expression, which we have shown previously to result in a loss of E6 [45]. In addition, analysis of GLUT-1 distribution within NIKS cells (Fig 9A) also shows a clear E6 PBM-dependent effect upon the pattern of GLUT-1 expression, although the specific distribution varies between the different cell types analysed. Taken together, these results indicate that the loss of E6 leads to an increased cytoplasmic expression of GLUT-1, a subset of which accumulates in close proximity to a component of the retromer complex.
The results in Fig 8 indicated a change in the localisation of the early endosomal marker Rab4. We therefore proceeded to investigate whether this was also reflected in a change in its subcellular distribution following ablation of E6 expression. The results in Fig 9C indeed show a significant alteration in the pattern of Rab4 distribution in HeLa cells following ablation of either E6 or E6AP expression. Once again there appears to be increased accumulation in close proximity to the Vps35 component of the retromer.
The above data demonstrate that E6 can modulate the pattern of SNX27 association with the endosomal transport machinery in a PBM-dependent manner. A consequence of this appears to be reflected in changes to GLUT-1 association with endosomal recycling complexes, with loss of E6 resulting in lower levels of GLUT-1 expression and altered intracellular distribution. Therefore we next wanted to determine whether this alteration in GLUT-1 transport was also reflected in an alteration of its function as a glucose transporter. To do this, HeLa cells were treated with siRNA against E6, or Luciferase as control, for 72h and glucose uptake was then measured using the analog 2-deoxyglucose (2-DG). As can be seen from Fig 10A, there is a significant decrease in the amount of glucose that is taken up by HeLa cells when E6 is depleted, compared with the Luciferase control. These data show that the depletion of GLUT-1upon E6 ablation is reflected in a reduction in glucose uptake by these cells, and reflects reports of similar results seeing loss of GLUT-1 expression following SNX27 knockdown [40]. This demonstrates a direct functional consequence for HPV-18 E6 modulation of the SNX27-retromer cargo transport pathway.
The above results indicate that SNX27 plays an important role in glucose uptake, an activity that is modulated by E6. They also suggest that under conditions of low glucose availability, loss of SNX27 might result in reduced rates of cell proliferation. To investigate this possibility, control HeLa cells and shRNA SNX27 HeLa cells were plated in low and high glucose, both in the absence and presence of DOX to induce the SNX27 shRNA. The cells were then counted over a period of days and the results obtained are shown in Fig 10B. As can be seen, loss of SNX27 in conditions of low glucose has a marked inhibitory effect upon continued cell proliferation, whilst high glucose availability largely mitigates these effects. These results confirm that SNX27 contributes directly towards nutrient uptake in HPV-18 positive HeLa cells.
In this study we have identified a novel activity of the high risk HPV E6 oncoproteins, linking them to the modulation of endosomal transport pathways. This appears to be mediated, at least in part, through a direct interaction between the high risk HPV E6 oncoproteins and the cellular SNX27, a protein that controls cargo fate determination in endocytic recycling. One consequence of this interaction is modulation of the endocytic transport of the glucose transporter GLUT-1, which subsequently affects the amount of glucose uptake in HPV-positive tumour cells. These results suggest that E6 can directly affect the nutrient balance in HPV infected cells, through modulation of endocytic recycling pathways to maintain sufficient nutrients for cell survival during the HPV life cycle and in progression to malignancy.
Previous proteomic analyses had suggested that HPV-18 E6 could potentially interact with different components of the endocytic sorting machinery, including SNX27 and the retromer components Vps35 and Vps26, indicating that E6 might be in close proximity to the retromer complex [23,24]. To investigate this in more detail we first performed a series of in vitro interaction assays and found that multiple high risk HPV E6 oncoproteins all share a similar ability to interact with SNX27. Mutational analyses demonstrated that the principal mode of recognition was through classic PBM-PDZ recognition, although there were some subtle differences from other reported E6 interactions with PDZ domain-containing substrates. In particular, the number of residues in the region of the E6 PBM that were critically required for SNX27 PDZ recognition appeared to be fewer than those required for MAGI or Dlg recognition. In addition, ablation of the E6 PBM or the core region of the SNX27 PDZ domain still allowed a low level of interaction, suggesting the existence of additional means by which E6 can interact with SNX27. In support of this, a weak but consistent interaction was also observed between HPV-11 E6 and SNX27, suggesting that modulation of endocytic transport might also be a feature of low risk HPV types, and this is currently a subject for further investigation.
In agreement with the proteomic analyses (23), we also found that HPV-18 E6 could co-immunoprecipitate both SNX27 and the Vps35 component of the retromer from cells, confirming a close association of E6 with the retromer complex. Indeed, cell fractionation studies demonstrated that the bulk of endogenously expressed E6 in HeLa cells is actually present within the membrane fraction of the cell, which is very similar to the localisation seen for SNX27. Throughout, we found no compelling evidence to suggest that E6 can influence the turnover of SNX27 through proteasome-dependent pathways. However, it was interesting to note that the membrane-bound forms of E6 were not subject to proteasome-mediated degradation, as there was no change in the levels of E6 expression in this fraction of the cell following proteasome inhibition. This contrasts with the situation with the cytoplasmic, nuclear and cytoskeletal pools, where all these forms of E6 were rescued following proteasome inhibition. It should also be mentioned that when E6 is overexpressed using exogenous transfection of expression plasmids, we invariably detect the bulk of E6 within the cytoskeletal compartment of the cell, emphasising the need to study the endogenously expressed E6 for these types of analyses.
Having defined SNX27 as a novel interacting partner of HPV-18 E6, we then proceeded to investigate the potential consequences of this interaction. Clearly, removal of E6 from HPV-18 positive HeLa cells has quite a marked effect on the subcellular distribution of SNX27, with significant accumulation in regions proximal to the nucleus and an apparent increase in the co-localisation with Vps35. To confirm that these alterations in the pattern of SNX27 expression were PBM-dependent, we also analysed the distribution of SNX27 in control NIKS, and in NIKS expressing HPV-16 E6 or HPV-16E6 ΔPBM. Again, there was significant co-localisation between SNX27 and Vps35 in the control NIKS and HPV-16 E6ΔPBM lines, but this was greatly decreased in the wild type E6-expressing cells. Taken together these studies suggest that the E6 association with SNX27 can modulate its subcellular distribution. To further investigate this we then analysed the distribution of SNX27 in different endosomal/lysosomal fractions using OptiPrep gradients to fractionate the subcellular membrane compartments in HeLa cells. Again, loss of E6 resulted in a marked shift in the distribution of SNX27 within the gradients. SNX27 remained in close proximity to the early endosomal marker Rab4, but shifted with respect to some of the Vps35 component of the retromer. Albeit less clearly, the GLUT-1 SNX27 cargo also appeared to shift in the gradient, suggesting that loss of E6 could also affect the distribution of an SNX27 cargo. These results were independently verified by immunofluorescence analyses, which also demonstrated an alteration in the co-localisation of Rab4 with Vps35, together with an increase in cytoplasmic GLUT-1 in cells where E6 expression had been ablated.
Having defined a role for the E6 PBM in modulating components of the endosomal transport machinery, we then proceeded to assess the effects on the function of GLUT-1. In both total cell extracts and in the gradient analyses, loss of E6 resulted in a marked decrease in the levels of expression of the functional higher molecular weight forms of GLUT-1. Similar effects were also observed following loss of SNX27 expression. Functionally, this loss of E6 resulted in a reduction in the levels of glucose uptake, suggesting that a direct consequence of E6 modulation of SNX27 function is an increase in the rates of glucose uptake, thereby favouring a more efficient use of available nutrients for continued cell proliferation and cell survival. In order to confirm that SNX27 does indeed play a significant biological role in HeLa cells we also measured cell growth following ablation of SNX27 expression under different growth conditions, and clearly found a reduction in cell proliferation when cells were grown in a low glucose medium. Taken together these results demonstrate that the association between E6 and SNX27 can modulate the endocytic transport of the GLUT-1 glucose transporter, one consequence of which is maintenance of cell proliferation under low nutrient conditions.
Several major questions arise from these studies. The first relates to the precise mechanism by which E6 can modulate SNX27 function. The interaction is largely PBM-PDZ mediated, although not exclusively. It seems unlikely that E6 and GLUT-1 could occupy the SNX27 PDZ pocket simultaneously; however recent studies have shown that the GLUT-1 affinity for the SNX27 PDZ domain is significantly higher than that of E6 [49]. This suggests that E6 association might be transitory, and be replaced in the PDZ pocket by high affinity cargoes. It remains to be determined whether the weaker association of E6 with SNX27 might also then contribute to a modulation of the sorting process, or whether the E6 PBM helps recruit SNX27 to certain endocytic compartments, favouring faster recycling. It is also possible that other, more weakly bound, cargoes of SNX27 might actually be out-competed for binding SNX27 by E6, thereby inhibiting their recycling.
Recent studies have shown an important role for the retromer in the function of SNX27, and certainly E6 appears to exist in a complex with SNX27 bound to Vps35, without apparently affecting the biochemical levels of either SNX27 or Vps35 [40,44]. Nonetheless depletion of E6 does seem to affect the degree of co-localisation of SNX27 with Vps35, both in immunofluorescence experiments and in gradient fractionations. In one of these assays, this modulation is shown to be PBM dependent. This suggests that E6 can modulate the endocytic compartment to which SNX27 is recruited, suggesting that the time spent in close proximity with Vps35 is reduced in the presence of E6. Dissecting the molecular basis for this will be an important avenue for future investigation.
Taken together, these studies indicate the existence of a novel activity for the HPV E6 oncoproteins, linking them directly to the modulation of endosomal transport pathways, and suggesting a completely novel way of modulating the cellular homeostasis, both during viral infection and in the development of malignancy.
HeLa (ATCC) and HeLa S4 shSNX27 (38) and SiHa (ATCC) cells were maintained in Dulbecco’s Modified Eagles’s Medium (DMEM) supplemented with 10% Fetal Calf Serum (Life Technology), penicillin-streptomycin (100 U/ml) and glutamine (300 μg/ml). Cells were cultured at 37°C with 10% CO2. The NIKS (Normal Immortalised Keratinocytes [50]) control, NIKS 16 E6 and NIKS 16 E6ΔPBM [51] cells were maintained in F medium (0.66 mM Ca2+) composed of 3 parts Ham's F12 medium to 1 part Dulbecco's modified Eagle's medium and supplemented with the following components: 5% fetal bovine serum (FBS), adenine (24 μg/ml), cholera toxin (8.4 ng/ml), epidermal growth factor (10 ng/ml), hydrocortisone (2.4 μg/ml), and insulin (5 μg/ml).
HeLa cells were transfected with siRNA using Lipofectamine RNAiMAX transfection reagent (Invitrogen). HeLa S4 shSNX27 cells were treated with 0.2 mg/ml Doxycycline to induce the shRNA. The following siRNAs were used: HPV-18 E6 and HPV-18 E6/E7 were custom synthesised by Dharmacon whilst the E6AP and SNX27 siRNAs were a Dharmacon Smart Pools.
GST fusion proteins were generated from pGEX2T plasmids expressing HPV-11 E6, 16 E6, 18 E6, 31 E6, 33 E6, 51 E6 and 58 E6 proteins, SNX27 [37], HPV-18 E6ΔPBM and HPV-18 E6 T156E as described previously [52]. The SNX27 deletion mutants (Δ67–77, Δ97–110 and Δ113–121) were prepared using primers against the specified regions using the pCI Neo Myc tagged SNX27 as template. The HPV-18 E6 C-terminal mutants were used for the in vitro binding assays as described previously [41]. All expression constructs were transformed into E.coli strain DH5α.
Purified GST fusion proteins were incubated with in vitro translated and radiolabelled proteins as indicated for 2 hours at room temperature. Proteins were translated in vitro using a Promega TNT Rabbit Reticulocyte Lystate kit and radiolabelled with [S35] Cysteine or [S35] Methionine (Perkin Elmer). Equal amounts of translated proteins were mixed with the GST fusion proteins immobilized on glutathione agarose beads and incubated on a rotating wheel. The beads were then washed thrice with PBS containing 1% Triton X-100 and analyzed by SDS-PAGE followed by autoradiography.
For endogenous protein co-immunoprecipitation assays, HeLa cells were seeded in 10 cm dishes and either left untreated or treated with siRNA against Luciferase or E6 as indicated for 72 hours. Cell lysates were prepared either in Lysis Buffer (20mM Tris pH 7.5, 150mM NaCl, 1mM EDTA, 1mM EGTA, 1% Triton X-100) containing protease inhibitors (Calbiochem Protease Cocktail 1); or they were extracted using the ProteoExtract Cell Fractionation Kit (Calbiochem). Then extracts were incubated with either anti-HA antibody (Roche), anti 18E6 antibody (ArborVita) or anti-SNX27 antibody (Abcam) as indicated overnight on a rotating wheel at 4°C. The immune complexes were captured using Protein A Sepharose beads and analyzed by SDS-PAGE followed by Western Blotting using anti-SNX27 antibody (Abcam), anti-Vps35 antibody (Abcam) and anti-18E6 antibody (ArborVita).
For cell fractionation studies, HeLa S4 shSNX27 cells were seeded on 6cm dishes at a density of approximately 1.2 x 105 and treated with Doxycycline or DMSO as indicated for 72 hours. The cells were then treated with MG132 or DMSO as indicated for 3 hours. Cells were collected by trypsinization and fractionated into cytoplasmic, membrane, nuclear and cytoskeletal fractions using the ProteoExtract Cell Fractionation Kit (Calbiochem) according to the manufacturer’s instructions and fractions were analyzed by SDS-PAGE followed by Western Blotting using anti-SNX27 antibody (Abcam), anti-18E6 antibody (ArborVita AVC#399), anti-α tubulin antibody (Sigma Aldrich), anti-p84 antibody (Abcam), anti-Transferrin Receptor antibody (Santa Cruz) or anti-Vimentin antibody (Santa Cruz) as indicated.
Cells were seeded on glass coverslips at a density of approximately 1.2 x 105 cells and transfected with siRNA against Luciferase, E6 or E6/E7 as indicated for 72 hours. The cells were fixed using 4% Paraformaldehyde and permeabilized using PBS containing 0.1% Triton X-100. Immunostaining was performed by incubating the coverslips in PBS containing antibodies against SNX27 (Abcam), Vps35 (Abcam), GLUT-1 (Abcam), p53 (Santa Cruz) or Rab4 (Abcam) as indicated overnight in a humidified chamber at 4°C. The coverslips were then washed thrice with PBS and incubated with the respective fluorophore conjugated secondary antibodies as indicated for 1 hour in a humidified chamber at 37°C. The coverslips were then washed thrice with PBS and twice with distilled water and mounted onto glass slides. The images were captured using the LSM510 META Confocal Microscope (Carl Zeiss) and co-localisations were quantified using the Velocity Software and Pearson’s Correlation Coefficient (PCC) calculated for each set of images, where values closer to 1 indicate closer degrees of colocalisation, whilst a value of zero would indicate no co-localisation.
HeLa cells were seeded in 10cm dishes and transfected with siRNA against Luciferase or E6 for 72 hours. Optiprep gradients (5%-25%) were prepared as described previously [44] and equilibrated at room temperature for 3 hours. Cell extracts were prepared in the homogenization buffer containing protease inhibitors as described previously [44], and the lysates were syringed to ensure breakdown of the cells. The lysates were centrifuged at 3000 x g for 5 minutes at 4°C and the post nuclear extracts (supernatants) were loaded onto the equilibrated gradients. The gradients were centrifuged at 32,000 rpm for 18 hours at 4°C and fractions were collected using a peristaltic pump. The collected fractions were mixed with chilled acetone and incubated at -20°C overnight to precipitate the proteins. The precipitates were centrifuged at 14,000 rpm for 10 minutes at 4°C and washed once with chilled absolute ethanol. The precipitates were air dried and dissolved in 2x Laemmeli’s buffer. The fractions were loaded onto 12% SDS polyacrylamide gels and the endocytic profiles were analyzed by Western Blotting using anti-GLUT-1 antibody (Abcam), anti SNX27 antibody (Abcam), anti-p53 antibody (Santa Cruz), anti- Vps35 antibody (Abcam), anti- Rab4 antibody (Abcam) and anti- LAMP2 antibody (Abcam) as indicated.
HeLa cells were seeded at a density of approximately 1.2 x 105 cells in 6cm dishes and transfected with siRNA against Luciferase or E6 as indicated using the Lipofectamine RNAiMAX transfection reagent (Invitrogen) for 72 hours. Glucose uptake was measured using the glucose analog 2-Deoxyglucose (2-DG) which can be taken up by the cells but is not metabolized. The uptake of 2-DG was measured colorimetrically using the Glucose Uptake Assay Kit (Abcam) as per the manufacturer’s instructions. The glucose uptake is measured as picomoles/μl and the measurements from three independent assays were used to generate the graph and standard deviations.
Control HeLa cells and HeLa cells (S4) containing DOX inducible SNX27 shRNA were plated in high (4.5g/l) or low glucose (1g/l) in the presence or absence of DOX to induce SNX27 knockdown. Cell numbers were then counted over a period of 4 days.
|
10.1371/journal.ppat.1004884 | Paradoxical Immune Responses in Non-HIV Cryptococcal Meningitis | The fungus Cryptococcus is a major cause of meningoencephalitis in HIV-infected as well as HIV-uninfected individuals with mortalities in developed countries of 20% and 30%, respectively. In HIV-related disease, defects in T-cell immunity are paramount, whereas there is little understanding of mechanisms of susceptibility in non-HIV related disease, especially that occurring in previously healthy adults. The present description is the first detailed immunological study of non-HIV-infected patients including those with severe central nervous system (s-CNS) disease to 1) identify mechanisms of susceptibility as well as 2) understand mechanisms underlying severe disease. Despite the expectation that, as in HIV, T-cell immunity would be deficient in such patients, cerebrospinal fluid (CSF) immunophenotyping, T-cell activation studies, soluble cytokine mapping and tissue cellular phenotyping demonstrated that patients with s-CNS disease had effective microbiological control, but displayed strong intrathecal expansion and activation of cells of both the innate and adaptive immunity including HLA-DR+ CD4+ and CD8+ cells and NK cells. These expanded CSF T cells were enriched for cryptococcal-antigen specific CD4+ cells and expressed high levels of IFN-γ as well as a lack of elevated CSF levels of typical T-cell specific Th2 cytokines -- IL-4 and IL-13. This inflammatory response was accompanied by elevated levels of CSF NFL, a marker of axonal damage, consistent with ongoing neurological damage. However, while tissue macrophage recruitment to the site of infection was intact, polarization studies of brain biopsy and autopsy specimens demonstrated an M2 macrophage polarization and poor phagocytosis of fungal cells. These studies thus expand the paradigm for cryptococcal disease susceptibility to include a prominent role for macrophage activation defects and suggest a spectrum of disease whereby severe neurological disease is characterized by immune-mediated host cell damage.
| Cryptococcus is an important cause of fungal meningitis with significant mortality globally. Susceptibility to the fungus in humans has been related to T-lymphocyte defects in HIV-infected individuals, but little is known about possible immune defects in non HIV-infected patients including previously healthy individuals. This latter group also has some of the worst response rates to therapy with almost a third dying in the United States, despite available therapy. Here we conducted the first detailed immunological analysis of non-HIV apparently immunocompetent individuals with active cryptococcal disease. In contrast to HIV-infected individuals, these studies identified a highly activated antigen-presenting dendritic cell population within CSF, accompanied by a highly active T-lymphocyte population with potentially damaging inflammatory cytokine responses. Furthermore, elevated levels of CSF neurofilament light chains (NFL), a marker of axonal damage in severe central nervous system infections suggest a dysfunctional role to this acute inflammatory state. Paradoxically, CSF macrophage proportions were reduced in patients with severe disease and biopsy and autopsy samples identified alternatively activated tissue macrophage populations that failed to appropriately phagocytose fungal cells. Our study thus provides new insights into the susceptibility to human cryptococcal disease and identifies a paradoxically active T-lymphocyte response that may be amenable to adjunctive immunomodulation to improve treatment outcomes in this high-mortality disease.
| Cryptococcus neoformans is an important cause of fatal meningoencephalitis in both those immunosuppressed from transplant conditioning or HIV/AIDS, as well as in previously healthy individuals. While AIDS-related cases represent the bulk of disease burden worldwide [1] with mortality approaching 60% in the developing world [2,3] and 20% in the developed world [4], non-HIV related cryptococcosis is a significant source of mortality and morbidity in the developed world, accounting for approximately a third of cases [5], with up to 30% mortality despite optimal therapy [4,6]. These mortality figures are derived from unselected cohorts in routine clinical settings and not clinical trials. In HIV-related disease where fungal burdens are high and cellular immunity low, recent approaches have sought to improve microbiological clearance from the CSF, an important prognostic marker [7]. These strategies have combined fungicidal drugs [8] or adjunctive cytokines such as interferon-γ (IFN-γ) [9,10]. The latter approach seeks to boost Th1-polarizing immunity, an immunological marker of survival during initial therapy [11]. In non-HIV-related disease, CSF fungal loads and effective microbiological clearance have similarly been associated with favorable outcomes [12]. However, little data is available regarding the immune milieu of these patients that could guide treatment, especially in severe or refractory cases. This has led to varying approaches for severe disease, including the use of immune intensifying regimens such as adjunctive IFN-γ [13], or immune-suppressive therapies such as steroids in the case of infections with C. gattii in previously healthy individuals [14]. Indeed, having too much or too little of a host immune response to infection may cause significant pathology [15,16]. Steroids have also been used in HIV-related cryptococcal meningoencephalitis with or without immune reconstitution syndrome (cIRIS), an aggravated paradoxical immune response to the pathogen in the setting of microbiologic control after restoration of T-cell activity from anti-retroviral therapy [17–19]. cIRIS is characterized by increases in immunoregulatory natural killer cells and IFN-γ-dependent factors such as monocyte attractant CXC motif chemokine 10 [20]. A clinical syndrome similar to cIRIS has also been reported in transplant recipients after reductions in immunoconditioning [21]. cIRIS-like syndromes are particularly problematic in cryptococcal meningitis, as the closed compartment of the brain within the skull allows little expansion with inflammation or cerebral edema. However, it is unclear whether the immunological response in cIRIS-like syndromes is similar in other cryptococcal patients.
Thus, given the controversies in management of severe cases of cryptococcal meningitis and an unexpected response to steroids in 2 sentinel cases under our care, more detailed, host-specific immune characterizations were sought to suggest a basis for rational treatment design in this complex disease. To this end, we conducted the first detailed immunological analysis of soluble and cellular responses to Cryptococcus in both blood and CSF in a consecutive cohort of non-HIV, non-transplant individuals with no comorbidities or iatrogenic immunosuppression who developed severe central nervous system disease (s-CNS). Despite minimal doses of steroids for maintenance of mental status that may have declined due to cerebral edema, immunophenotyping of s-CNS patients during therapy demonstrated a CNS compartmentalized, marked increase in activated T cell and NK cell populations, accompanied by elevated soluble IFN-γ, interferon-generating cytokines IL-18, and interferon-induced chemokines CXCL10, as compared to a cohort having non-CNS disease (non-CNS) or healthy donors (HD). Ex vivo T-cell activation with cryptococcal antigen-loaded dendritic cells (DCs) also demonstrated a compartmentalized Th1 response with increased CD4+ and CD8+ IFN-γproduction, thus providing a mechanism for the clinical deterioration and response to steroids. However, the present studies also demonstrated reduced monocyte and myeloid dendritic cells in the CSF and an M2 bias in tissue-infiltrating macrophages on examination of biopsy and autopsy specimens from patients with s-CNS disease, suggesting downstream defects in myeloid activation. These data thus identify a paradoxical Th1-biased CSF immune response in severe CNS cryptococcal disease after microbiological control, suggesting further study to investigate the role for adjunctive T-cell immunosuppressive therapy in this high mortality disease.
While the severity of the clinical conditions precluded randomization to steroid treatment, clinical responses to steroids (prednisone 1 mg/kg/d) were observed in two patients who developed deteriorating mental status despite effective fungicidal therapy and negative CSF fungal cultures. This provided a rationale for detailed immune studies, subsequently performed on 17 consecutive patients with CNS disease (Table 1) defined by positive cryptococcal cultures or antigens at the time of diagnosis. The subgroup excluded patients with previous immunosuppression or known risk factors for cryptococcal disease such as idiopathic CD4 lymphopenia [22].
The first steroid-responsive patient was a 58 year-old male who presented with dizziness, hearing loss and a left-sided facial droop. CSF cultures grew C. neoformans, identified on L-canavanine glycine bromothymol blue (CGB) media. The patient was treated with 6 weeks of intravenous amphotericin B lipid complex and flucytosine followed by fluconazole, but required placement of a ventriculoperitoneal shunt for deteriorating mental status and persistently elevated intracranial pressures. The patient was then transferred to the NIH clinical center, having a GCS of 13. Brain MRI scanning showed increased signal intensity (bright signal) on enhanced fluid attenuation inversion recovery (FLAIR) imaging within the sulci, especially the right central sulcus (Fig 1A). This patient was treated on day 1 with liposomal amphotericin B and flucytosine without improvement over a 2 month period with 10 day GCS mean scores ranging from 11–13, prompting a trial of recombinant interferon-γ 1b (red arrow in Fig 1A), but the patient developed hypersomnolence with significant worsening of mean GCS over the ensuing 10 days (pre: 12.2 +/-0.3 (SEM), N = 23; post: 10.6 +/- 0.3, N = 19; p = 0.001). As a result, recombinant interferon-γ 1b was discontinued and corticosteroids (prednisone 1 mg/kg/d) were added, after a 1 week ‘washout’ of IFN-γ, to the standing regimen. After this addition, the patient showed progressive improvement in mental status over the next 40 day period (10 day GCS, pre: 10.6 +/-0.3 (SEM), N = 19; 40 day post: 13.7 +/- 0.4, N = 21; p < 0.001) and normalization of GCS in 60 days (Fig 1B). Unfortunately, this patient developed residual bilateral cranial nerve VII and VIII palsies prior to steroid therapy, resulting in corneal scarring with consequential blindness and deafness, and gait disturbance and urinary retention from cryptococcal cauda equina syndrome.
The second patient was a 46 year-old female from Washington State who presented with intermittent fevers, worsening headaches, and altered mental status. MRI of the brain with gadolinium demonstrated multiple enhancing mass lesions in the left basal ganglia with the largest measuring approximately 3.5 x 2.8 centimeters. Adjacent smaller but more intensely enhancing lesions were seen in the left basal ganglia as well as throughout the brain, including multiple smaller lesions in the right posterior frontal lobe, left insula, right thalamus, left temporal lobe and left cerebellum (Fig 1B). The left basal ganglia lesions were associated with surrounding edema, best appreciated on FLAIR images as bright signal around the lesions, with secondary mass effect and slight impingement on the left lateral ventricle. Chest CT revealed an 8-cm right mid-lung mass, which was biopsied and grew C. gattii, genotype AFLP6A/VGIIa [23], consistent with the predominant outbreak strain found in the Pacific Northwest [24] and serum testing subsequently identified a functional anti-GMCSF autoantibody measured as previously described [25]. She was begun on liposomal amphotericin B, flucytosine and prednisone 50mg/d with improvement in mental status and MRI imaging at 24 days, which demonstrated decreased edema (decreased bright signal on FLAIR) in association with the left basal ganglia lesions and reduction in left lateral ventricular obstruction (Fig 1B, Day 24 and S1 Fig). Steroids were discontinued on day 24 to minimize possible immune suppression at the outside hospital and the patient’s status again deteriorated by day 34, prompting intensive care unit admission. Of note, the patient continued treatment with amphotericin B, remaining with negative CSF fungal cultures. Repeat MRI imaging showed recurrence of edema surrounding the left basal ganglia lesions with increased mass effect on the left lateral ventricle and slight midline shift (Fig 1B, Day 34). After consultation with the NIH, dexamethasone equivalent to prednisone 50mg/d was again added to the amphotericin B regimen with improvement in mental status and decreased enhancement and ventricular obstruction on brain MRI over the following month (Fig 1B, Day 66, S1 Fig). The patient was transferred to the NIH on Day 85 and an attempt was made to reduce the prednisone dose to 25 mg, but mild mental status deterioration and increased lesion size and edema on brain MRI (S1 Fig, Day 99) prompted an increase in steroid dose after her lumbar puncture and immunophenotyping with mental status improvement to a GCS back to normal (10 day GCS, pre: 14.3 +/-0.2 (SEM), N = 43; 10 day post: 15.0 +/- 0.0, N = 28; p < 0.01). Subsequently, amphotericin B was replaced by fluconazole and steroids were slowly tapered over the following 8 months as the cryptococcal antigen declined and was accompanied by normalization of activities. Fortunately, this patient made marked recovery with ability to resume her activities of daily living but with some residual memory deficits.
Subjects were participants in an observational cohort examining the host genetics and immunology of cryptococcal disease in previously healthy, non-HIV infected adults. Written informed consent was obtained and approved by the research ethics committee of the NIAID institutional review board. Seventeen consecutively recruited patients with culture or biopsy-proven CNS cryptococcal disease were studied. Severe disease was present in all 17 and was defined as patients demonstrating significant deteriorations in mental status (Glasgow Coma Score, GCS < 15) despite standard fungicidal therapies with amphotericin B-based regimens and 12 of 17 required ventriculoperitoneal shunting for persistently elevated intracranial pressures. The s-CNS patients had demonstrated initial responses to antifungal therapy as evidenced by improvement in clinical, CSF, and radiologic parameters but subsequently worsened along these parameters in the absence of microbiologic growth or another infectious process after induction antifungals. As shown in Table 1, the 17 patients with s-CNS had median age of 54 years, with normal CD4+ (359-1565/μl) and CD8+ (178-853/μl) T lymphocyte counts. Patients were studied while undergoing active treatment with anti-fungal regimens and had negative CSF cultures at the time of analysis and remained negative throughout treatment. Fourteen required adjunctive prednisone (1 mg/kg/d or less) for control of cerebral edema, which was either begun after immunophenotyping or was minimized prior to study as symptoms allowed with a median dose of 2.5 mg/d. Of note, all patients were screened for GMCSF autoantibodies, but only two were positive (both s-CNS with C. gattii). Comparator groups consisted of HD as well as a non-CNS infected cryptococcal cohort that submitted to a lumbar puncture for research purposes.
Twelve color immunophenotyping panels quantified absolute numbers and proportions of 12 major subtypes of immune cells in the blood and CSF [26] in the 17 s-CNS patients and were compared to 6 patients with previous fungal exposure leading to non-CNS disease as well as 11 HDs. These studies showed that s-CNS patients (Figs 2 and 3, right panels) had dramatic CSF elevations in absolute numbers of all components of the innate and adaptive immune responses as compared to non-CNS infected patients and HDs. Specifically, we observed significantly increased absolute counts of HLA-DR activated CD4+ and CD8+ T lymphocytes and B lymphocytes in the CSF of s-CNS compared with non-CNS and HD (Fig 2). We also identified significantly increased cytotoxic (CD56dim) and immunoregulatory (CD56bright) NK cells, monocytes, and myeloid dendritic cells (MyDCs) in the CSF of s-CNS cases compared with non-CNS cases and HD. Interestingly, there was a mild proportional decrease—albeit non-significant—in monocytes and MyDCs in the CSF of s-CNS cases compared with non-CNS cases and HD (Fig 3). Elevations in absolute granulocyte counts, attributed to steroid therapy in some patients, were the other CSF abnormality that differentiated meningitis patients from non-CNS patients and HD controls who did not receive steroids (Fig 3). However, no blood immunophenotyping differences were noted among the groups except for increased absolute counts and proportions of plasmacytoid dendritic cells (PlDCs) in the blood of non-CNS patients compared to s-CNS. Importantly, few CSF or blood immunophenotyping abnormalities were observed in patients with non-CNS disease compared to HDs, the latter likely due to the well-controlled nature of the disease and length of therapy.
In addition, CSF NFL levels were assayed, which are highly stable, neuron sensitive and specific biomarkers of axonal damage that fill the axon core and are released during neuronal damage [27–29]. Interestingly, we found a more than 10-fold, statistically significant increase in CSF NFL among s-CNS compared with non-CNS and with HD (S3 Fig). This suggests that the increased inflammatory response in s-CNS disease is not a benign or protective response, but is accompanied by elevated markers of neuronal damage.
Because we observed that cells of the adaptive immune system were elevated in the CSF of s-CNS patients to a similar or even greater degree than cells of innate immunity, we tested if these cells recognized cryptococcal antigens in a subgroup of 8 s-CNS patients, and if so, whether they secreted IFN-γ, which is thought to confer protective immunity. Five patients with non-CNS disease served as an appropriate comparator in this case, because they had been exposed to the cryptococcal fungus and therefore would be expected to have memory and effector T cell responses specific for cryptococcal antigens. Since it is known that only memory/effector T cells routinely cross the blood-brain barrier to perform immunosurveillance functions [30], intrathecal enrichment for cryptococcal-specific T cells have to be demonstrated against individuals with pre-existing cryptococcal exposure to be biologically meaningful.
Because the immunodominant epitopes of Cryptococcus have not been well defined and the s-CNS and non-CNS groups would require matching for major histocompatibility complex antigens (MHC) for peptide-based assays, we instead utilized a novel methodology based on the co-culture of autologous antigen (Ag)-loaded mature dendritic cells (mDCs) and T cells followed by flow cytometry-based confirmation of antigen-specificity as previously described [31]. In this assay, autologous DC’s phagocytose complex antigens (such as a cryptococcal homogenate) and select those epitopes that have the strongest binding to a patient’s MHC molecules, thus mimicking in vivo conditions as much as possible. Furthermore, this assay allows simultaneous read-outs of both CD4+ and CD8+ T cell responses, which is not possible in peptide-based assays. In addition to cryptococcal cell homogenate, we used a more defined mannoprotein (MP) preparation, which has been used previously in HIV-related cryptococcal infections to assess T-cell immune responses [11,32,33].
Using this approach, significant proliferative and secretory responses of CD4+ and CD8+ cells were exhibited in response to DCs loaded with either cryptococcal homogenate or purified cryptococcal MP. Higher proliferation of blood CD4+ T lymphocytes co-cultured with Ag-loaded DCs (Fig 4A, top left panel) was observed in the non-CNS patients compared with the s-CNS patients. In contrast, higher proliferation of CSF CD8+ T lymphocytes (Fig 4A, bottom right panel) was observed in the s-CNS patients after co-culture, suggesting a compartmentalization of the proliferative responses. Proliferation of CSF CD4+ cells was also observed in s-CNS patients compared to the unstimulated cells from the same patients, but was not statistically greater than the non-CNS patients in this small cohort of patients. In addition, increased IFN-γ secretion (Fig 4B, right panel) but not TNF-α was observed in both CD4+ and CD8+ T cells from CSF of s-CNS patients compared to those from non-CNS patients. These data suggest an ongoing compartmentalized antigen-specific Th1 response in the CSF.
Interestingly, responses to homogenate were greater than to that of the purified MP. MP was made from an acapsular strain (cap67) having serotype D background and serotype D is unusual in the U.S., whereas the homogenate was derived from H99, a serotype A strain—a clinical strain that is much more prevalent in the U.S.. In addition, MP is a purified component of the fungus, which undergoes mannosylation [34,35], but does not contain cell wall carbohydrates, lipids, or nucleic acids found in homogenates that could boost responses due to toll-like receptor (TLR) and c-lectin receptor (CLR) stimulation [36,37].
To further assess CSF functional responses, expression of representative soluble cytokine and chemokines was determined among patients at the time of the cellular studies described in Fig 3. CSF from s-CNS patients demonstrated significantly increased intrathecal production of soluble IFN-γ, as well as interferon-induced IL-18, and chemokines CXCL10, as compared to the non-CNS patients or HD (Fig 5a–5c). Although IL-6 induces c-reactive protein (CRP) in the liver via the JAK/STAT pathways [38], no correlation was noted between serum CRP and CSF IL-6 and there was no difference between serum CRP of CNS vs. non-CNS cases (Table 1). The pro-inflammatory IL-6 (Fig 5g) similarly demonstrated increased expression in CSF of s-CNS patients as was the cytokine, IL-10 (Fig 5e). Interestingly, TNF-α levels were not significantly elevated (Fig 5d), consistent with the observed lack of increased TNF-α cellular production by T-cells incubated with ex vivo antigen-stimulated DCs described in the previous section and/or a functionally insufficient immune response from monocytes. S-CNS patients also produced significantly elevated levels of the chemokines CCL2 (monocyte chemoattractant protein-3 [MCP-1], CCL3 (macrophage inflammatory protein-1α [MIP-1α]), CCL7 (monocyte chemoattractant protein-3 [MCP-3]), and CCL19 (macrophage inflammatory protein-3β [MIP-3β]), compared to HD (Fig 5i–5l). Since these chemokine ligands are chemotactic for monocytes and T-lymphocytes, but CXCL10 is chemotactic only for CXCR3+ T-lymphocytes, one can infer the relative infiltration of monocytes compared to T-cells in the CSF based on the ratio of CCL3/CXCL10, for example. Statistically significant decreased CCL2/CXCL10 and CCL3/CXCL10 among s-CNS compared with HD and CCL19/CXCL10 among s-CNS compared with non-CNS and with HD corroborated decreased circulating monocyte CSF populations vis à vis T lymphocyte trafficking to the CSF (S2 Fig). In addition, non-CNS patients had significantly more detectable intrathecal IL-4 than s-CNS and HD (Fig 5f). These data again suggest a robust Th1-polarized response with evidence of active inflammatory activation within the intrathecal compartment.
Analysis of the intrathecal compartment by CSF sampling may be limited due to differences between CSF circulating and tissue-penetrant cells. Thus, brain biopsies obtained from two of the s-CNS patients as part of their diagnostic workup were examined using CD68, CD163, CD3, CD4 and CD8 staining. These biopsies were done prior to steroid therapy. Both biopsies demonstrated a predominance of CD68 positive monocyte/macrophages as well as the presence of T-cell infiltration in the region of the biopsy (Fig 6), suggesting that reductions in CSF circulating monocytes may be due to tissue infiltration of this cell population in these patients. Interestingly, while macrophages underwent tissue infiltration, iNOS and CD200R1 staining in one of the biopsy specimens that contained sufficient tissue for more detailed analysis suggested the macrophages underwent a non-protective alternative activation and less than 1% of fungal cells were observed to be co-localized with macrophages, suggesting poor phagocytic function during cryptococcal brain infection of this patient (Fig 6B).
Because of the limited availability of brain tissue in the live s-CNS cohort, an analysis of NIH clinical center records (1957–2009) was conducted to identify non-AIDS infected patients who died of cryptococcal meningitis by autopsy (Table 2). All patients were treated with amphotericin-based regimens but two had had Hodgkin’s disease. Review of the autopsy report allowed exclusion of patients with known comorbidities such as sarcoidosis, HIV or other co-infections and inclusion of patients who died with cerebral herniation. During his neurological decline, one patient also developed bronchopneumonia from aspiration. Characteristics of the five s-CNS patients who died of cryptococcal meningoencephalitis are listed in Table 2. In these patients, immunohistochemistry (IHC) showed extensive leptomeningeal and Virchow-Robin space infiltration of CD68 and CD163 macrophages as well as the presence of CD3, CD4, and CD8 T lymphocytes (Fig 7).
This suggests that the apparent decrease of myeloid cells in the CSF obtained in the present cohort may have been due to relative increases in monocyte/DC infiltration of the brain parenchyma. This would appear to be an effective monocyte response. However, further analysis of the autopsy specimens using fluorescent probes of tissue sections (Fig 8) indicated that the majority of the patients (4/5) demonstrated an M2 non-protective bias (CD200R1+) to their CD68 positive monocyte/macrophages and calcofluor white fungal staining (Fig 9) demonstrated poor uptake of fungal cells (< 1%) by these macrophage populations, regardless of the monocyte activation status, consistent with the result in the biopsy specimen.
The present study sought to comprehensively investigate innate and antigen-specific intrathecal immunity of HIV-uninfected individuals with clinically severe, cryptococcal meningoencephalitis to guide treatment strategies of a disease group with high mortality. These studies were prompted by several potential conflicting issues in human cryptococcal disease including 1) a presumed T-cell defect in previously healthy patients based on paradigms developed with HIV-associated infections, 2) the unknown contributions of microbiological control vs. dysfunctional/paradoxical immune responses in CNS disease and 3) the resultant ambiguities as to whether microbiological control should be intensified in severe cases with antifungals or augmentative cytokines such as IFN-γ or adjunctive immune suppression with steroids or other agents. A dramatic clinical and radiological response in two patients to steroids further provided a rationale for these studies, focusing on subsequent, consecutive patients recruited to the NIH clinical center with CNS cryptococcosis. Likely due to referral bias, all 17 were severely affected with altered mental status, suggesting a label of s-CNS disease to indicated the severity of the cohort. The clinical process appears to be quite long lasting, with clinical severity and MRI brain findings following the clearance of cryptococcal antigen over many months or a year as described in S1 Fig.
Interestingly, despite an expectation of low T-cell function, this cohort of s-CNS disease, displayed intrathecal expansion and activation of cells of both innate and adaptive immunity including HLA-DR+ CD4+ and CD8+ cells and NK cells. These expanded CSF T cells were enriched for cryptococcal-antigen specific CD4+ and CD8+ T cells that are critical in protection from the fungus [39], secreting the protective cytokine, IFN-γ [40] as well as a lack of highly elevated CSF levels of typical T-cell specific Th2 cytokines IL-4 and IL-13 (IL13 was barely significant)[41], or typical Th17 cytokines IL-17 or GM-CSF (also not shown). Animal studies have suggested that IL-17 is not required for classical activation of macrophages in cryptococcal infections [42] and may not be as important as Th1-type responses [43]. Because patients with s-CNS had more than 1000-fold increased numbers of CSF T cells, in comparison to systemic Cryptococcus disease controls (i.e., non-CNS) and HDs, lack of prototypic Th2 and Th17 cytokines in the CSF of these patients is highly informative and in our opinion represents strong evidence for Th1 skewing of infiltrating T cells. This Th1-biased response was unexpected as such a response is thought to be protective against C. neoformans in both mouse studies [44] and human experience where T-cell defects from HIV/AIDS [11] or during solid organ transplant conditioning [45], predispose to cryptococcal infections. Interestingly, Jarvis et al noted different antigen-specific functional T cell phenotypes in HIV-infected patients with cryptococcal meningitis and another intracellular pathogen—tuberculosis [11]. These represent important findings and suggests that the clinical worsening in Patient 1 (Fig 1) after recombinant IFN-γ therapy may have been due to exacerbation of this paradoxical inflammatory state and suggests caution in using this suggested approach [9,10] in non-HIV-related s-CNS cryptococcal disease. Indeed, we found that this inflammatory response was accompanied by a more than 10-fold elevation in CSF NFL, a biomarker of axonal damage in diverse neurological disease [27,28] and suggests that the paradoxical inflammatory response results in ongoing neurological damage. Thus, the present study represents a critical application to human disease whereby too much of an immune response can be detrimental by virtue of host-induced damage [15,16].
Interestingly, we observed this aggravated immune response in patients infected with either C. neoformans or C. gattii, the latter of which is reported to cause increased inflammatory responses in mice [46] and human patients [14]. This paradoxical immune response in non-HIV-infected individuals bears resemblance to that of cIRIS in HIV/AIDS that occurs after initiation of anti-retroviral therapy (ART). In cIRIS, patients who initially show a clinical and microbiological response to anti-fungal therapy develop, after initiation of ART, clinical worsening of meningitis symptoms including deteriorations in mental status or increased lymphadenopathy, cutaneous lesions or pulmonary nodules, despite continued negative fungal cultures [47]. Immunologic studies of HIV-related cIRIS identified within the blood compartment, a similar Th1-driven response consisting of enrichment of CD4+ and CD8+ T cells coexpressing activation markers CCR5 and CXCR3 with detection of IFN-γ-dependent chemokines such as CXCL10 [48]. Within the CSF compartment, cIRIS is associated with elevations of IFN-γ, TNF-α, G-CSF, VEGF, and eotaxin (CCL11) [49,50]. The paradoxical response is thought to be due to a reversal of HIV-associated immune defects in the presence of continued fungal antigen tissue burden [51].
However, a predominant difference between cIRIS and the intrathecal inflammatory response of s-CNS patients in the present study may lie in the M2 skewing of CNS-tissue infiltrating macrophages in the latter group which is a non-protective response in animal models of cryptococcosis [52]. This M2 phenotype was suggested by the predominance of CD68+, CD200R1+ monocytes found in our autopsy and biopsy marker studies and was also corroborated by a lack of increased CSF levels of IL-12p40 and TNF-α as well as elevated CSF levels of IL-10 in the s-CNS patients. This M2 skewing was present regardless of the presence of steroids and was present early in the course of the disease as demonstrated in the diagnostic biopsy #1. It is unclear whether the predominant intrathecal cellular source of these differentially-expressed cytokines are T cells or monocytes/macrophages, which are traditionally robust sources of TNF-α and IL-10 [53]. Given the Th1 bias of T cells in the present study, the tissue infiltrating M2 cells are a potential source of IL-10 and the M2 bias likely reduced production of fungal protective TNF-α [54,55]. This latter defect may also explain the critical lack of fevers in these patients, which is a major factor leading to diagnostic delays in this patient population.
A second main difference of the non-HIV response from cIRIS is that no immunostimulatory or immune restoring interventions were required to produce the paradoxical immune response in the HIV-uninfected individuals. The fact that clinical deterioration in the present series occurred, usually after a month of fungicidal therapy and with negative CSF cultures suggests that liberation of fungal antigens during therapy may have induced intrathecal activation of antigen-specific T cells. Indeed, intact C. neoformans cells prior to therapy are protected from immune recognition by an immune suppressive polysaccharide capsule [56], an extracellular laccase [57] and components of the fungal cell wall [58], lysis of which results in release of intracellular fungal antigens.
Because the presented functional studies are extremely laborious, require large amounts of fresh biological samples and the targeted patient population is rare and includes very sick subjects, a natural limitation of our study is its small sample size and limited sample numbers per patient. In addition, the variability in treatment duration and steroid use for cerebral edema may make heterogeneity a source of concern, though the significant findings despite this are noteworthy. The present study was thus an important first step in the description of the immune response in this population, but did not attempt to use specific immune responses to stratify outcomes or follow the evolution of the disease, although elevated Th1 responses were identified over a prolonged time frame as long as symptoms and antigen persisted. The limitation of the patient population to those previously healthy helped to add a measure of uniformity to the population, but may limit the application to more immunosuppressed patients such as those undergoing transplant conditioning. In addition, referral bias likely led to a higher disease severity that may not be applicable all previously healthy patients with CNS disease. However, the expected high mortality of this group makes its study highly compelling for suggesting studies to improve therapy. In addition, it is important to note that the results here represent patients that deteriorate during effective fungicidal therapy and should not be applied to clinical issues on initial presentation when microbial control may be more important.
Despite the robust Th1 bias and inflammatory response, the immune response appeared to be ineffective to prevent or clear the fungus. Recent data demonstrating the relative rarity of non-HIV cryptococcal disease [5] despite high rates of exposure [59] also support a defective host response. Data acquired in the present study all point towards a defective innate immunity, and, more specifically, defects of the myeloid lineage. We identified poor proportional monocyte recruitment to the circulating pool of the CSF cells and a predominant non-protective M2 response in most of the autopsy specimens, as well as an inability to phagocytose the fungus in situ. This is consistent with our recent data demonstrating the presence of a STAT5 phosphorylation-blocking GM-CSF autoantibody in a subset of patients with cryptococcal disease [25,60] and suggest that future targeted studies should focus on the mechanistic assessment of cells of myeloid lineage, including evaluation of phagocytosis [61], oxidative burst [62]and antigen processing [63] important to cryptococcal defense. In conclusion, the present study suggests that an ineffective macrophage response in the setting of intact T-cell activation has the potential to cause increased susceptibility of the patient to the fungus and severe CNS disease after initiation of effective anti-fungal therapy. We infer that there may be either intrinsic or extrinsic defects at the immune synapse between Th1 cells and antigen presenting monocytes or further downstream that lead to a paradoxical M2 polarization. These studies also provide pathophysiological support for studies of adjunctive immunosuppressive trials in severe non-HIV related cryptococcal disease to control paradoxical immune responses after sustained microbiological control in a CNS fungal disease with high mortality.
The National Institute of Allergy and Infectious Diseases (NIAID) Institutional Review Board (IRB) approved this study under NIAID Protocol #93-I-0106. The National Institute of Neurological Disorders and Stroke (NINDS) approved the study under NINDS Protocol #09-N-0032. All subjects provided written informed consent directly or via their durable power of attorney, after obtaining National Institutes of Health (NIH) Bioethics consultation.
We studied 17 consecutive patients with cryptococcal meningitis who were all found to exhibit characteristics of severe CNS cryptococcal disease (s-CNS) from an ongoing recruitment cohort of 80 patients with cryptococcal disease in previously healthy individuals without known immunocompromising conditions (HIV 1/2, HTLV-1, and Hepatitis B and C seronegative individuals, idiopathic CD4 lymphopenia (ICL), those without antecedent known malignancy, liver/renal failure, sarcoidosis, systemic autoimmune disorders, and recipients of transplants or immunosuppressive medications such as corticosteroids, calcineurin inhibitors, etc…) S-CNS was defined as persistent or deteriorating mental status changes (Glasgow Coma Score, GCS < 15) despite adequate amphotericin B-based antifungal therapy (≥6 weeks) and all had repeated negative CSF cultures after 6 weeks of therapy. Comparator groups were selected as patients within the larger cohort with non-CNS disease including 5 with pulmonary and 1 with isolated spinal bone disease and 11 healthy donors (HD). Diagnosis was made by CSF India ink stain and/or culture, or a positive mucicarmine stain and/or culture from a biopsy site or a positive cryptococcal antigen testing (Meridian Cryptococcal Latex Agglutination System; Meridian Bioscience), according to previous case definitions [64]. We performed an extensive chart review of 17 s-CNS cases and 6 non-CNS cases and tabulated pertinent clinical information related to the timeline of their infection in Table 1. Mental status was evaluated as Glasgow Coma Score (GCS) by clinical staff of the NIH clinical center during 24 h monitoring, using a standardized protocol and expressed as a 10 day average +/- SEM of all measurements. (http://www.commondataelements.ninds.nih.gov/Stroke.aspx#tab=Data_Standards). Evaluations were performed 3–4 times per day between the hours of 8 am and 11 pm.
Immunophenotyping of peripheral blood cells was performed within 60 min of ex vivo collection on anticoagulated blood after osmotic lysis of erythrocytes. CSF samples were placed on ice immediately after collection. Within 15 min of collection, the CSF (usually 20 ml) was spun and cell pellets were resuspended in 400 ml ice-cold X-Vivo media (Lonza, Alpharetta, GA). The Fc receptors of cells were blocked by using 2% immune serum globulin to prevent non-specific activation. A 12-color immunophenotyping panel to distinguish 12 leukocyte cell types was adapted using the markers in S1 Table from that previously described [26].
120 ml of heparinized blood samples were collected eight days before the scheduled lumbar puncture (LP). Isolation of PBMCs and generation of DCs with GM-CSF, IL-4 and 5% human serum were performed as described previously [31]. On day six of culture, immature DCs were co-incubated overnight with the following antigens (Ag’s) and subsequently stimulated for 48 hours: mannoprotein (MP), prepared as described [32], and cryptococcal heat-killed suspension (Crypto; both 5 μg/ml) with maturation factors, TNF-α, IL-1B, IL-6 and PGE2. Cryptococcal heat killed suspension was prepared by glass-bead fractured C. neoformans H99 strain ATCC 208821 ( = CBS8710) that had been heated at 90°C for 30 min and the suspension allowed to settle overnight at room temperature. The suspended material was recovered and quantified by protein analysis by using a BioRad assay according the manufacturer’s directions (BioRad, Hercules, CA). Optimal Ag-concentrations were determined in preliminary experiments by titration (i.e. inducing T cell proliferation without evidence of toxicity).
At the day of LP (day eight of culture) peripheral T cells were purified from frozen PBMCs by negative selection (Miltenyi Biotech, San Diego, CA). Peripheral and CSF T cells were then cultured in round bottom 96 well plates with autologous Ag-loaded mDCs (3x103 mDCs: 3x103 T cells) in a total volume of 100 μl X-Vivo media (Lonza) at 37°C, 5% CO2. After seven days of proliferation, T cells were re-stimulated overnight for the intracellular cytokine staining (ICCS) with twice as much newly differentiated, identically loaded autologous mDCs as used for the first co-culture in presence of Brefeldin A and monensin (eBioscience). Ag-specificity was detected by flow cytometric analysis of cytokine-producing CD4+ and CD8+ T cells.
T cells were analyzed for the surface markers CD3 (UCHT1), CD4 (RPA-T4), and CD8 (SK1), and intracellular expression of GM-CSF (GM2F3), IL-17 (eBio64DEC17), TNF-α (MAb11) and IFN-γ (B27; BD Biosciences, San Jose, CA and eBioscience). Data were analyzed with BD FACS Diva 6.1 (BD Biosciences).
Absolute numbers of cytokine-producing cells were determined by normalization with APC beads. For inter-patient comparison we standardized the number of cytokine positive events (i.e. IFN-γ+, TNF-α+, and double positive events) to 1000 beads each.
Mean fluorescence intensities (MFIs) of cytokine-secreting CD4+ and CD8+ T cells were determined by subtraction of the respective isotype controls. Negative values were adjusted to zero.
After lumber puncture, CSF samples were centrifuged at 300 g and stored at -80C until analysis. CSF from unstimulated CSF flow cytometry experiments on s-CNS, non-CNS, and HD above were analyzed using a sandwich-enzyme linked immunosorbent assay (Luminex, Bio-Rad). After including standards, we assayed for intrathecal evidence of IL-4, IL-6, IL-8/CXCL8, IL-10, IL-12p40, IL-13, IL-17, IL-18, IP-10/CXCL10, IFN-α2, IFN-γ, TNF-α, MCP-1/CCL2, MIP-1α/CCL3, MCP-3/CCL7, MIP-3β/CCL19, M-CSF, and GM-CSF. Detected amounts were expressed in log10 picogram/milliliter (pg/ml). CSF levels of NFL were measured according to the manufacturer’s instructions using a sandwich ELISA method (Uman Diagnostics AB; Umea Sweden).
Brain immunohistochemistry (IHC) on five patients who died from s-CNS cryptococcal disease (and two biopsy specimens from live patients with s-CNS disease) were collated. The inclusion criteria for autopsies were based on findings one week prior to death: (1.) cryptococcal meningoencephalitis with or without extraneural infection (CSF India ink or culture positive and WBC > 50 and opening pressure > 200 mm); (2.) Glasgow Coma Score < 8 for > 48 hours; (3.) the presence of Babinski or other upper motor neuron signs or seizure activity; and (4.) exclusion of competing causes of death. Chromogenic staining using horse radish peroxidase-conjugated secondary antibody for T lymphocyte markers (CD3 [F7.2.38 clone], CD4, and CD8) and macrophage markers (CD68 [KP1 clone] and CD163) was performed (Dako North America, Carpenteria, CA) and representative images captured using Leica Software at 10x and 20x magnification. Polarization of macrophages was performed on autopsy and biopsy specimens using double immunofluorescence staining according the manufacturer’s directions (Abcam). M1 was characterized by CD68/iNOS expression and M2 was identified by CD68/CD200R1 co-staining as described [65]. Calcofluor white was used to stain cryptococcal cells and DAPI (4',6-diamidino-2-phenylindole) for nuclear staining. To detect cryptococcal DNA in the paraffin-embedded specimens, a DNA extraction was performed by using the blackPREP FFPE DNA Kit (Analytik Jena, Jena, Germany) with an additional lyticase treatment. Real-time PCRs were performed on the extracts to detect cryptococcal DNA, and specifically C. gattii, as described previously [66,67].
Ten day GCS averages were determined by averaging scores over successive 10 day intervals (N > 20 determinations), beginning on the first day of admission among each of two illustrated s-CNS patients. We used repeated measures ANOVA to account for the inter-correlated GCS values to determine statistical significant changes 10 days before and 10 days after corticosteroid therapy. For the CSF cytokine/chemokine analysis among s-CNS, non-CNS, and HD, the non-parametric Kruskal-Wallis test was utilized, controlling for multiple comparisons between each group using the Dunn procedure [68]. For NFL chain analysis, since we were interested in assessing only neuronal damage using s-CNS as the comparator, we performed a non-parametric Mann-U Whitney test adjusting for comparison between s-CNS vs. non-CNS and s-CNS vs. HD using a Bonferroni correction. CSF cytokine, chemokine, and NFL analysis was performed using PRISM (version 6.0, Graphpad Software, La Jolla, CA).
For immunophenotyping assays, one-way ANOVA (analysis of variance) was performed to explore the association of explanatory variable (diagnosis: s-CNS, non-CNS, and HDs) with 56 response variables which are the absolute numbers or proportions of adaptive and innate immune cells collected from CSF and blood staining samples. Since most of these response variables did not follow a normal distribution, Box-Cox transformation was applied (log- transformation for most of the variables) prior to ANOVA. Tukey’s method was used for pair-wise multiple comparisons. P-values were adjusted for multiple testing. A two sample t-test was used to test the difference between Ag-specific T cell proliferation and between MFIs in the non-CNS and s-CNS cohorts. Peripheral and intrathecal T cells were analyzed separately for CD4+ and CD8+ T lymphocytes. Zero MFI values were treated as missing values. Log transformation was applied to all outcome measures. These statistical analyses were performed using SAS version 9.3 (SAS Institute, Cary, NC) with p ≤0.05 considered a significant level.
|
10.1371/journal.pcbi.1005760 | A coupled-oscillator model of olfactory bulb gamma oscillations | The olfactory bulb transforms not only the information content of the primary sensory representation, but also its underlying coding metric. High-variance, slow-timescale primary odor representations are transformed by bulbar circuitry into secondary representations based on principal neuron spike patterns that are tightly regulated in time. This emergent fast timescale for signaling is reflected in gamma-band local field potentials, presumably serving to efficiently integrate olfactory sensory information into the temporally regulated information networks of the central nervous system. To understand this transformation and its integration with interareal coordination mechanisms requires that we understand its fundamental dynamical principles. Using a biophysically explicit, multiscale model of olfactory bulb circuitry, we here demonstrate that an inhibition-coupled intrinsic oscillator framework, pyramidal resonance interneuron network gamma (PRING), best captures the diversity of physiological properties exhibited by the olfactory bulb. Most importantly, these properties include global zero-phase synchronization in the gamma band, the phase-restriction of informative spikes in principal neurons with respect to this common clock, and the robustness of this synchronous oscillatory regime to multiple challenging conditions observed in the biological system. These conditions include substantial heterogeneities in afferent activation levels and excitatory synaptic weights, high levels of uncorrelated background activity among principal neurons, and spike frequencies in both principal neurons and interneurons that are irregular in time and much lower than the gamma frequency. This coupled cellular oscillator architecture permits stable and replicable ensemble responses to diverse sensory stimuli under various external conditions as well as to changes in network parameters arising from learning-dependent synaptic plasticity.
| The mammalian olfactory bulb responds to odor stimulation by generating fast oscillations in its electrical field potential. Such oscillations are indications that a substantial number of principal neurons in the olfactory bulb are coordinating their activities in time, which often means that their action potentials are synchronized, or partly synchronized, such that the pattern of small differences in their spike times contains olfactory sensory information. We are interested in the mechanisms by which olfactory bulb circuitry can transform sensory information from the temporally unsophisticated spike rates of primary sensory neurons into this sophisticated cortical format. We present a biophysically explicit, multiscale dynamical model of the olfactory bulb network that generates these oscillations. The elements of this model are designed to adhere to experimental findings from individual neurons, membrane currents, and synapses as well as the functional network. Together, these elements generate gamma oscillations exhibiting the full range of properties of those in the biological system. We show that these dynamics arise from an inhibition-coupled oscillator framework, a type of dynamical system with some established mathematical properties. This finding enables us to understand how the olfactory system translates sensory information for distribution in the central nervous system, and how different areas of the brain can mechanistically coordinate with one another so as to regulate the flow of sensory information to appropriate target structures.
| The mammalian main olfactory bulb (OB) plays a central role in processing and relaying olfactory information from the primary sensory epithelium to subcortical and cortical areas [1]. This processing transforms the information content of the primary representation, but also has been proposed to transform the underlying physical metric by which this information is encoded, from rate-coded population activity organized on a respiration timescale to a spike timing-based representation aligned to a faster timescale that is determined by the intrinsic dynamics of cortical neural ensembles [2]. Odor stimulus-evoked activation of the OB generates fast, gamma-band (30–80 Hz) local field potential (LFP) oscillations that are thought to be largely synchronous across the extent of the OB [3]. Such oscillations reflect the tightly constrained synchronization of a large neural assembly, which in the OB (and its arthropod analogues) has long been believed to play some role in the encoding and processing of olfactory information [4–14].
It is generally accepted that OB gamma oscillations are intrinsic, and mediated by a fast negative feedback loop formed between principal output neurons (mitral and projecting tufted cells; MCs) and a class of inhibitory GABAergic interneurons (granule cells; GCs), interacting via dendrodendritic synapses in the external plexiform layer (EPL) of the OB (Fig 1A; [15–25]). However, the underlying mechanisms generating these oscillations remain elusive. Several different dynamical architectures have been proposed or assumed to mediate OB gamma oscillogenesis. First, a pyramidal-interneuron network gamma (PING) mechanism is often assumed [22, 26], inspired by the anatomical predominance of the excitatory-inhibitory reciprocal synapses that constitute the EPL network. Early theoretical modeling of OB network dynamics also was based on this anatomical architecture [27]. However, PING networks do not incorporate cellular resonance properties, such as the intrinsic subthreshold oscillations (STOs) of MCs [28]. Second, an interneuron network gamma (ING) mechanism has been theoretically proposed [29, 30]; however, this mechanism relies on inhibitory interactions among granule cells, which were intimated by early EEG work [16] and by the discovery of GABAergic synaptic inputs onto granule cells [31] but since have been ruled out. The PING and ING architectures have been reviewed by [32]. Third, OB network oscillations have been proposed to be driven directly by the intrinsic subthreshold dynamics of MCs [33]. This model highlighted the dynamical capacities of intrinsic MC subthreshold oscillations (STOs; [28]) and resolved some limitations of the PING architecture regarding observed OB dynamics (e.g., it permitted stable gamma oscillation frequency in the presence of fluctuating afferent drive). However, this model required substantially higher-frequency MC STOs than have been experimentally described, and also was not clearly compatible with the sparse spiking behavior of GCs [34]. Fourth, a hybrid network based on inhibition-coupled intrinsic cellular oscillators has been proposed [35], in which the intrinsic STOs of MCs are transiently coupled during afferent activation into a coherent oscillatory network [36] paced by GC-mediated inhibitory synaptic inputs that periodically reset the slower MC STOs. (Pulsed inhibitory inputs, including shunting inhibition, have been demonstrated to effectively reset MC STOs [28, 37–39]). During these active epochs, the network dynamics exhibit key PING-like properties (e.g., the population oscillation frequency depends on the decay time constant of the GABA(A) receptor conductance), but they also retain a dependence on the slower STO dynamics of mitral cells even when the STO frequency itself is superseded by the network oscillation. This dynamical mechanism, pyramidal resonance interneuron network gamma (PRING), is consistent with a broad range of experimental data and is modeled here.
It is important to clearly understand the specific dynamical mechanisms underlying OB field oscillations, for several reasons. These oscillations are likely to reflect the re-encoding of afferent odor information into timing-based representations for distribution to multiple postbulbar cortical and subcortical structures [40]. Therefore, in order to understand the formation and information content of these secondary representations, the dynamics of their creation must be clear. Moreover, ascending inputs from the anterior olfactory nucleus and piriform cortex, among other structures, must be integrated into this dynamical framework. Piriform cortical inputs, in particular, are understood to alter bulbar dynamics, transiently transforming the OB’s intrinsic gamma oscillations into slower beta-band oscillations coherent with those of the piriform cortex [13, 41–44]. To understand how these ascending inputs are integrated into the secondary odor representation requires a correct, mechanistic model of bulbar gamma oscillogenesis and its subversion by piriform cortical activity. Finally, field potential oscillations at many different characteristic frequencies are found all over the brain, often interacting within particular neural structures [45] and potentially serving to select and route specific information between coherently activated brain regions [46]. Elucidation of the detailed mechanics of oscillations and their transitions in the OB and its associated networks hence also will pertain to broader questions surrounding interareal communication mechanisms in the brain.
To address this question, we developed a conductance-based, dynamically detailed biophysical model of the OB network. The present model is based on our earlier two-layer model of cholinergic neuromodulation in the OB [25], but embeds these glomerular layer and intercolumnar EPL computations within an explicit spatial framework. The results from this model favor the PRING mechanism described above [35], and demonstrate that this inhibition-coupled cellular oscillator architecture supports the diverse phenomena observed in OB neurophysiological recordings. These phenomena include (1) patterned spiking activity in MCs and GCs that both is broadly heterogeneous and occurs at lower frequencies than the population rhythm, (2) tolerance to a wide range of afferent MC excitation levels, which is important for mediating the representation of different odor qualities, (3) tolerance for substantial changes in MC-GC synaptic weights, which underlie intrinsic odor learning within the OB [47–49], (4) the broad coherence of gamma-band oscillations across a physically extensive network despite the incoherent activity of some neurons within that network, (5) the phase-constraining of spikes within each cycle of the gamma oscillation [10, 13], and (6) the persistence of LFP gamma oscillations at consistent frequencies despite sparse network connectivity (connection probability p = 0.3 between MCs and GCs) and sharply heterogeneous afferent activation levels. The explicit, multiscale nature of this dynamical model further enables the elaboration, explanation, and experimental testing of the underlying mechanistic details that may underlie these observed physiological phenomena.
Stimulation with simulated odorants induced gamma oscillations that were coherent across the entire OB network. Simulated odorants comprised heterogeneous levels of input delivered to the 25 MC/PGC pairs; each MC fired at a different mean rate corresponding to the strength of its afferent input (including feedforward inhibition from its associated PGC; Fig 2A). The mean MC firing frequency in response to odor stimulation was 14 Hz (min = 4 Hz; max = 38 Hz; standard deviation (SD) = 9.8 Hz). Despite these heterogeneous firing rates, a strong and broadly coherent oscillation emerged in the gamma band (32.4 Hz, Fig 2B), consistent with in vitro recordings from olfactory bulb [19, 35]. Individual MCs responded in a mixed mode, usually spiking at mean frequencies substantially below the underlying STO frequency, but with odor-evoked spikes phase-constrained to the underlying sLFP oscillation, as observed experimentally [10, 13]. Moreover, systemwide coherence was maintained; voltage timeseries depictions of different pairs of MCs confirmed that the STOs of different MCs were synchronized with one another (Fig 2C, top panel), MC spikes were synchronized with STOs from other MCs (Fig 2C, bottom panel), and MC spikes were also substantially synchronized with spikes from other MCs (Fig 2D1 and 2D2). GC subthreshold voltages also fluctuated rhythmically and were well synchronized with one another (Fig 2E1), as were GC spikes (Fig 2E2), despite GCs’ low mean firing rates (4.6 Hz). In contrast, no gamma-band synchrony was observed in the subthreshold voltage fluctuations (Fig 2F1) or spiking activity of PGCs (Fig 2F2). Population spike histograms of MCs, GCs and PGCs with corresponding frequency power spectra are shown in Fig 2D3, 2E3 and 2F3 respectively. The population spiking activities of both MCs and GCs exhibited gamma rhythmicity, and the frequency was the same as that measured from the sLFP (32.4 Hz; Fig 2D3 and 2E3). By comparison, no rhythmicity was observed in the PGC population spike histogram, and the frequency power spectrum was flat (Fig 2F3).
Examined in aggregate, MC spikes were phase-constrained within the common, coherent gamma cycle of the OB network. The majority of MC spikes were evoked near the crest of the oscillatory sLFP (Fig 3A). GC spikes also were phase-constrained within the gamma cycle, occurring predominantly during the descending phase of the sLFP (Fig 3B). In contrast, PGC spikes were not phase-constrained, but were distributed uniformly across the gamma oscillation cycle (Fig 3C).
Because of the tight phase-locking between MC/GC spikes and the sLFP cycle, the gamma rhythm also was evident in MC/GC population spiking activities (Fig 3D). The tightly alternating relationship between MC and GC population spiking, but not PGC spiking, suggested that this temporal delimiting of MC activity arose from effective feedback inhibition delivered by granule cells. To illustrate this point, we plotted the voltage traces of a weakly-activated MC and a strongly-activated MC against their respective cumulative GC-mediated GABAA conductances (Fig 3E). The weakly-activated MC STO depolarized directly as its inhibitory conductance decayed (Fig 3E, upper panel) and the strongly-activated MC fired only after release from inhibition (Fig 3E, lower panel). Moreover, the inhibitory conductance increased again directly following the evocation of MC spikes, initiating the next excitation-inhibition cycle.
Sufficiently strong inhibitory GC input also effectively reset the phase of MC STOs (Fig 3E, upper panel, arrows), consistent with experimental observation and earlier cellular models [28, 37–39]. In principle, such resets erase the history imposed by longer-timescale internal dynamics, thereby enabling afferent input levels across the MC population to determine the depolarization rates in each MC from a common starting state, potentially governing MC spike phase as well as spike probability [39, 50]. Moreover, recurrent resets also serve to supersede the intrinsic frequency of MC STOs, enabling the network to oscillate at a frequency faster than that generated by intrinsic STO dynamics [35]. In aggregate, these reciprocal interactions between MCs and GCs synchronized MC internal dynamics and MC spikes, incorporating them into a coherent gamma oscillation in which MC spikes were reliably phase-constrained with respect to the common oscillatory sLFP of the network (Fig 3F; [10, 13]).
Functional computations in the olfactory bulb are generally independent of the physical distance between columns [2, 51–54], though their underlying biophysical mechanisms often have proximity-dependent properties. We therefore asked whether the distance-dependent spike propagation delays along MC lateral dendrites were sufficiently heterogeneous to impair the global coherence of gamma oscillations across the OB circuit. To visualize the propagation delay as a function of distance, the membrane potentials of a representative MC (MC[2][2]) were recorded from the soma and from the locations of three reciprocal synapses distributed along the lateral dendrite (at 80 μm, 235 μm, and 500 μm from the soma; Fig 4). These three synapses connected, respectively, to an adjacent GC (GC[5][4]), a GC connecting near the middle of the lateral dendrite (GC[6][3]), and a GC connecting at the end of the lateral dendrite (GC[0][1]). Subthreshold activity in the MC dendrite was slightly hyperpolarized as the recording site progressed away from the soma, but spikes propagated at essentially full amplitude (Fig 4A). Spike propagation was rapid, with less than 1 ms delay from the soma to the end of the 500 μm dendrite (Fig 4B), suggesting that heterogeneous spike propagation delays would have little effect on network synchronization. This reflects the experimental observation that spikes fully propagate along a MC lateral dendrite with little delay (Fig 2 in [55]), and is consistent with previous computational work in which spike backpropagation along MC lateral dendrites activates granule cells independently of distance [56].
It has been proposed that gamma oscillations in the OB depend on MC STOs [28, 33]; however, in the PING framework, the pyramidal (excitatory) neurons generally do not exhibit resonance. We therefore asked whether and how MC STOs contribute to the robustness, power, and regularity of gamma coherence and spike synchronization in the active OB network. To investigate this, we first removed STOs from model MCs and examined the effect of this change on network dynamics. Specifically, STOs were eliminated by replacing the persistent sodium current (INaP) in all MCs of the network with ohmic cation currents scaled to maintain the same MC firing rates under the same current injection levels (Fig 5A and 5B; [39]). The cation current was modeled as ICAT = gCAT(v−ECAT), where gCAT = 0.26 mS/cm2 and ECAT = 0 mV. Under this manipulation, the power and regularity of odor stimulus-induced network gamma were substantially reduced and sLFP oscillations became less coherent, as evidenced by reduced persistence in the autocorrelogram and a lower, flatter peak in the power spectrum (compare Fig 5C with Fig 5D). An examination of membrane potential timeseries from two pairs of MCs revealed that, although MCs without intrinsic STOs could still display subthreshold voltage fluctuations owing to phasic inhibition from granule cells, these fluctuations had smaller amplitudes and were much less regular compared with intact STOs in control cells (compare Fig 5E with Fig 5F). MC spikes also became less synchronized with one another in the absence of intrinsic STOs (compare Fig 5E with Fig 5F, bottom panels), although the mean odor-evoked MC firing rates were essentially identical (Control: 14 Hz; STO removed: 13.2 Hz). Finally, the synchronization index (SI) was reduced from 0.64 in controls to 0.53 when STOs were removed. Hence, MC resonance contributed substantially to the integrity and regularity of coherent gamma oscillations in the active OB network, even when the intrinsic STO frequency was superseded by the PING-like mechanisms of the network frequency (see below). The added stability and robustness of OB gamma oscillations contributed by these MC resonance properties resembles the advantages of resonance-induced gamma (RING) oscillations [57], with the important distinction that RING is described for resonant inhibitory interneurons, whereas in the OB network it is the excitatory principal neurons that are resonant. The mechanism of OB oscillations can be described as pyramidal resonance interneuron network gamma (PRING), thereby acknowledging the PING-like properties of the activated gamma oscillation as well as the additional properties afforded by MC resonance.
Intrinsic STO frequencies appear to yield to higher-frequency network-based oscillations owing to recurrent STO phase resets delivered by GC-mediated synaptic inhibition [28, 35]. If this interpretation is correct, and MC resonance is an important contributor to network coherence and frequency stability, then this oscillatory coherence should be disrupted if the intrinsic STO frequency becomes faster than the natural frequency of the synaptically-based network oscillation. To test this, we increased the intrinsic MC STO frequency by reducing the time constant of the activation variable of the slow potassium current (IKS) [33]. Specifically, we reduced the activation time constant of IKS from 10 ms to 5 ms, and increased the conductance densities of the IKS and INaP currents by factors of 1.6 and 1.3 respectively to maintain approximately the same STO amplitudes and MC firing rates. These modifications increased the STO frequency in an isolated MC model cell from 29 Hz (in controls) to 44 Hz in response to a 200 pA depolarizing current injection (Fig 6A and 6B). Without altering any other model parameters, this change in the intrinsic STO frequency seriously disrupted the sLFP gamma rhythm and sharply reduced gamma power in the OB network (compare Fig 6C with Fig 5C). A comparison of STO voltage timeseries with the aggregated GABAA conductances in the same MCs confirmed that GC inhibition could no longer effectively regulate MC STOs, which became irregular (Fig 6D). Phase locking between MC spikes and sLFP oscillations also was significantly reduced (SI, controls: 0.64; increased STO frequency: 0.38), although the average odor-evoked MC firing rate was virtually unchanged (controls: 14 Hz; increased STO frequency: 14.4 Hz).
If this disruption was due to a mismatch between intrinsic STO frequency and the natural frequency of the network oscillation, as predicted, rather than to some separate effect of the changes made to the model MCs, then the coherence of OB gamma oscillations should be restored if the natural frequency of the network oscillation was also increased so as to again be faster than those of the MC STOs. In PING and ING networks, the natural frequency of network oscillations depends strongly on the decay time constant of the inhibitory synapse [32, 58]. Indeed, when the GABAA receptor decay time constant of the GC→MC synapses was reduced to 3 ms (from the default 18 ms), in a network populated with MCs exhibiting the higher intrinsic STO frequency, a strong gamma oscillation re-emerged at 51.3 Hz (Fig 6E)–considerably faster than the 32.4 Hz frequency exhibited by control networks. Under these conditions, MC STOs again displayed rhythmicity and were entrained effectively by GC-mediated GABAA synaptic conductances (Fig 6F); network synchrony also was substantially restored (SI, controls: 0.64; increased STO frequency alone: 0.38; increased STO frequency + 3 ms synaptic decay time constant: 0.56). These simulations indicate that the decay rate of GC-mediated GABAA inhibition must be faster than the intrinsic MC STO frequency in order to be able to synchronize MC dynamics.
To test whether it was important that the inhibitory synaptic decay time constant be closely matched to the MC STO frequency, or that it simply be faster, we tested a network in which we paired the faster (3 ms) GABAA synaptic decay time constant with the default (29 Hz) intrinsic MC STO frequency. Under these parameters, the network oscillation frequency increased from 32.4 Hz (under control conditions; Fig 2B) to 43.4 Hz (Fig 7A). This increase in oscillation frequency was accompanied by a slight reduction in oscillatory power and coherence, as indicated by a wider spectral peak and less persistent periodicity, though the amplitudes of the two spectral peaks were comparable (compare Fig 7A with Fig 2B). Additionally, under these conditions the GABAA conductance fluctuated regularly and decayed fully within every gamma cycle owing to its fast dynamics, effectively entraining MC STOs (Fig 7B). In contrast, when the GABAA decay time constant was increased from 18 ms to 30 ms, there was no change in the peak frequency of the network oscillation (18 ms: 32.4 Hz; 30 ms: 33 Hz), though its power was reduced considerably (compare Fig 7C with Fig 2B). Within individual MCs, the slowly decaying GABAA conductance accumulated across successive gamma cycles and lost much of its rhythmicity, resulting in inconsistent effects on MCs that failed to supersede their intrinsic STO frequency preferences (Fig 7D). Accordingly, under control parameters, synaptic decay time constants faster than ~18 ms progressively increased network sLFP oscillation frequencies, whereas slower time constants had no effect (Fig 7E). These faster kinetics also maintained relatively high power spectral peaks at the gamma frequency (oscillation indices), whereas slower synaptic kinetics resulted in rapidly declining oscillation indices (Fig 7F). Mean spiking frequencies in both MCs and GCs, however, varied monotonically with respect to the rate of GABAA decay (Fig 7E), likely because slower decay rates produced an overall increase in the total integrated inhibition of MCs. Specifically, as the decay time constant was increased from 3 ms to 30 ms, the MC firing rate decreased from 21.6 Hz to 10.8 Hz, resulting in a concomitant decrease in GC firing rate from 6.9 Hz to 3.4 Hz. In sum, these results showed that a wide range of synaptic decay time constants generated reliable coherence from the OB network provided that they were lower (faster) than a threshold value determined by the intrinsic frequency of MC STOs. However, there also was a clear peak (15 ms; Fig 7F), indicating that the strongest network oscillatory power could be achieved by an optimal matching of the synaptic and STO timescales. Additionally, these results demonstrated that the network oscillation frequency was robust to substantial changes in mean spike frequencies in both MCs and GCs.
Finally, we decided to increase the intrinsic MC STO frequency by increasing the excitation levels of all MCs, rather than by altering their IKS and INaP conductance parameters as above, in order to test whether similar dynamical effects resulted. Depolarizing MCs increases their intrinsic STO frequencies both experimentally [28] and in the present model. To broadly increase MC excitation while retaining the same heterogeneous odor inputs, we decreased the level of PGC-mediated inhibition on MCs by reducing the PGC→MC synaptic weight to half of the control value (from 4 to 2). The results largely conformed to those observed when STO frequencies were increased by adjusting cellular conductance parameters (Fig 6). Under default parameters (with an 18 ms GABAA decay time constant), reducing PGC inhibitory weights by half had no effect on the network oscillation frequency (controls: 32.4 Hz; 50%WPGC-MC: 33.6 Hz), but did impair the coherence and stability of field potential oscillations and reduce the oscillation index (peak spectral power; compare Fig 8A with Fig 2B). In contrast, when using a faster GC synaptic decay time constant of 3 ms, this reduced PGC inhibition produced a coherent gamma oscillation at a higher peak frequency (controls: 32.4 Hz; 3 ms decay time constant only: 43.4 Hz; 3 ms decay time constant + 50%WPGC-MC: 51.3 Hz; Fig 8B), because the faster synaptic decay was again able to effectively reset the intrinsic MC STOs on every cycle. This result further suggests that higher overall levels of MC excitation, which generate faster intrinsic STO dynamics, would require correspondingly faster synaptic inhibition kinetics in order to maintain network stability, and thereby demonstrates the importance of maintaining a limited range of mean MC activity levels via global afferent activity normalization ([59]; corrected mechanism in [54]).
The synaptic weight of GC→MC inhibition also is an important factor in determining the stability of network gamma oscillations. To assess this effect, we varied the GC→MC synaptic weight (WGC→MC) from zero (full blockade) up to five times the default value. Under full blockade conditions, MC spikes and STOs were desynchronized (Fig 9A) and network sLFP oscillations were dramatically reduced (compare Fig 9B with Fig 2B). Whereas overall spike rates increased substantially (average odor-evoked spike rate, controls: 14 Hz; no GC inhibition: 24 Hz), synchronization among MC spikes was sharply reduced (SI, controls: 0.63; no GC inhibition: 0.30). These results further confirm that GC-mediated feedback inhibition is necessary for the synchronization of mitral cells into a coherent gamma rhythm in the OB.
In contrast, when WGC→MC was increased threefold (from 2 to 6), MC spiking activity was reduced substantially (controls: 14.0 Hz; 300% WGC→MC: 8.4 Hz) and STOs were corrupted by an irregular mixture of shorter and longer oscillation periods, though MC membrane potential fluctuations were still moderately well-coordinated (Fig 9C). The frequency power spectrum reflected this disruption, presenting a number of low-power peaks (Fig 9D); two of these (at 18.3 Hz and 29.9 Hz) were somewhat more distinct, though both remained well below control amplitudes (compare Fig 9D with Fig 2B). These results indicate that excessive inhibition of MCs by large GC→MC synaptic weights impairs network gamma oscillations by disrupting STO periodicity.
The frequency of the network sLFP oscillation and the mean spike rates of both MCs and GCs declined as WGC→MC increased from 0 to 6 and remained stable thereafter (Fig 9E). In contrast, the synchronization index rose substantially as WGC→MC increased from 0 to 2 and maintained this level for all larger synaptic weights measured (Fig 9F). The oscillation index (spectral peak amplitude) also increased greatly as WGC→MC grew from 0 to 2, but then progressively decreased once WGC→MC exceeded 3 (Fig 9F). This pattern of results indicates that the degradation of gamma oscillatory power at larger GC→MC weights was not a result of reduced phase coupling, but of disrupted STO periodicity (Fig 9C). In sum, while sufficient GC inhibition is required to reset and synchronize MC STOs, excessive GC synaptic weights are detrimental to the stability of the gamma rhythm; an optimal level of GC inhibition is required to sustain a strong and coherent gamma oscillation.
To understand in detail why larger GC→MC synaptic weights impaired gamma rhythmicity, we plotted spike time histograms for both MCs and GCs alongside the membrane potential timeseries of a representative MC and the aggregate GC-mediated GABAA conductance of that MC, all under the disruptive conditions of a 3-fold increase in WGC→MC (Fig 10A; same parameters as Fig 9C and 9D). At a timepoint marking a surge of synchronous MC spiking activity, GCs responded in turn with higher-than-average activity (Fig 10A, top and second panels, leftmost vertical line). Because of the large WGC→MC, this surge in GC activity evoked a particularly enlarged (and correspondingly persistent) GABAergic chloride conductance in MCs (Fig 10A, third panel, leftmost vertical line), which substantially hyperpolarized MC membrane potentials (Fig 10A, bottom panel, leftmost vertical line) and, in aggregate, noticeably suppressed MC firing across the network (Fig 10A, top panel, second vertical line). This reduced level of MC activity, in turn, did not induce any GC spiking in that cycle (Fig 10A, second panel, second vertical line). As the GABAergic chloride conductance continued to decay (Fig 10A, third panel, second vertical line), marginally increased numbers of spikes were generated from the MC population (Fig 10A, top panel, third vertical line), which evoked weak responses in GCs (Fig 10A, second panel, third vertical line) and hence much smaller GABAergic conductances that only minimally hyperpolarized MC membrane potentials (Fig 10A, third and bottom panels, third vertical line). After a few such “small” cycles, the MCs recovered from the effects of accumulated inhibition and a high-activity cycle occurred again (Fig 10A, all panels, rightmost vertical line). The irregularity of this recurrent process substantially distorted MC subthreshold activity and gamma rhythmicity (Fig 9D; Fig 10A, bottom panel).
If this disruption of gamma oscillations indeed resulted from an oversuppression of MCs by excessive GC inhibition, as hypothesized, then boosting MC excitability should restore the rhythmicity. To test this, we increased MC mean firing rates back to the control level by reducing the PGC→MC inhibitory synaptic weight (WPGC→MC) to 50% of its default value (controls: 14.0 Hz; 300% WGC→MC: 8.4 Hz; 300%WGC→MC + 50%WPGC→MC: 14.4 Hz). GC firing rates also were restored to control levels by this change (controls: 4.6 Hz; 300%WGC→MC: 2.8 Hz; 300%WGC→MC + 50%WPGC→MC: 4.5 Hz). Under these restored excitability conditions, MC spikes again reliably drove substantial GC responses in every gamma cycle, the GABAergic synaptic conductance changes became more regular, and the periodicity of MC subthreshold activity was substantially improved (compare Fig 10B with Fig 10A). Moreover, MC STOs were again well synchronized, and exhibited greater stability and regularity than under conditions of elevated GC inhibition but default PGC inhibition (compare Fig 10C with Fig 9C). As a result, the second spectral peak observed in Fig 9D was eliminated and a single coherent gamma peak again appeared at 34.2 Hz, comparable to the control value of 32.4 Hz (compare Fig 10D with Fig 2B). The above simulation demonstrates that the detrimental effect of excessive GC inhibition on gamma rhythmicity can be ameliorated by reduced PGC inhibition, indicating that an overall balance of excitation and inhibition is required for coherent, stable network gamma oscillations.
Finally, the synaptic weights and decay time constants of GABAA synapses are not functionally independent of one another; shorter decay time constants generate less total MC inhibition and a weaker and shorter suppressive effect, all else being equal. We therefore asked whether the optimal inhibitory synaptic weights for robust oscillations and synchronization would differ depending on the synaptic time constant. We generated a network in which the GABAergic decay time constant was reduced from 18 ms (in controls) to 3 ms (as depicted in Fig 7A and 7B), and measured network oscillation and spike frequencies and the oscillation and synchronization indices as functions of GC→MC synaptic weight. As predicted, the oscillation index (OI) peak and the SI plateau both occurred at substantially higher inhibitory synaptic weights when using the faster decay time constants (compare Fig 10E and 10F to Fig 9E and 9F). The inhibitory synaptic decay time constant therefore also must be factored into the balance between excitation and inhibition that enables stable and coherent gamma oscillations across the OB network.
The functional efficacy of feedback inhibition in the OB EPL depends on the synaptic weights of both the inhibitory GC→MC and the excitatory MC→GC synapses. If a balance between excitation and inhibition is required for strong and stable gamma oscillations across the OB, then an optimal range of excitatory MC→GC synaptic weights may also exist. However, because MC→GC synapses onto adult-born GCs are plastic [47, 48], the EPL network would be expected to tolerate a substantial range and heterogeneity among these synaptic weights. To examine the functional range of synaptic weights for the excitatory MC→GC synapses in this network, we varied the MC→GC synaptic weight (WMC→GC) from 0 up to 8 times the default value. When these synapses were blocked (i.e., WMC→GC = 0), GCs were largely inactive (0.9 Hz spontaneous background activity); other simulation results were similar to those obtained when blocking GABAergic synaptic transmission (i.e., WGC→MC = 0; Fig 9A and 9B) and are not separately reported here. When WMC→GC was reduced to 50% of the default value (from 1 to 0.5), the GC subthreshold potential was substantially hyperpolarized and lost much of its rhythmicity compared with controls (Fig 11A, upper panel), leading to significantly smaller and arrhythmic GABAergic chloride currents in MCs (Fig 11A, lower panel). Because of this reduced phasic GC inhibition, MC activity increased, but both MC spikes and STOs were relatively desynchronized (Fig 11B), and gamma oscillations were greatly impaired (compare Fig 11C with Fig 2B).
In contrast, when WMC→GC was increased to 8 times the default value (from 1 to 8), GCs were strongly excited, spiking in response to many incoming EPSPs and maintaining a level of rhythmicity comparable to controls (Fig 11D, upper panel), but delivering much larger phasic GABAergic chloride conductances onto MCs (Fig 11D, lower panel). The increased level of phasic inhibition suppressed MC spikes, but retained the synchrony and periodicity of MC STOs (Fig 11E). Accordingly, a robust and coherent gamma oscillation persisted even with an 8-fold increase in the MC→GC synaptic weight, with little change in frequency (controls: 32.4 Hz; 800%WMC→GC: 35.4 Hz; Fig 11F).
To break this effect down further, we generated raster plots of MC and GC firing under these two conditions. When WMC→GC was reduced by 50%, the mean odor-evoked GC firing rate was reduced from 4.6 Hz (in controls) to 2.6 Hz, resulting in a slight increase in the mean MC firing rate from 14 Hz (in controls) to 17.7 Hz. As noted above, network synchrony was reduced substantially (SI, controls: 0.63; 50%WMC→GC: 0.40), because neither MC nor GC spike trains were well coordinated (Fig 12A and 12B). In contrast, with an eightfold increase in WMC→GC, GC firing rates were greatly increased (controls: 4.6 Hz; 800%WMC→GC: 12.8 Hz) and GC spikes became remarkably well synchronized; this strong GC activation substantially suppressed MC firing (controls: 14 Hz; 800%WMC→GC: 3.3 Hz; Fig 12C and 12D). This substantially different balance of MC and GC activity was stable because one MC input was strong enough to produce correlated discharges in many postsynaptic GCs. Notably, under these conditions the mean MC firing rate (3.3 Hz across all MCs) was much lower than the oscillation frequency (35.4 Hz) and a majority of MCs exhibited no odor-evoked spikes (Fig 12C). The spikes of the remaining active MCs were effectively entrained by the highly synchronous GC activity, and exhibited elevated levels of synchrony (SI, controls: 0.63; 800%WMC→GC: 0.92); i.e., coherent gamma oscillations persisted despite substantial increases in lateral excitatory synaptic weights. This is a particularly important stabilizing property given that the intrinsic OB synaptic plasticity underlying odor learning relies on the potentiation of excitatory synapses [47, 48].
The average odor-evoked MC/GC firing rates and sLFP oscillation frequencies across a range of MC→GC synaptic weights are depicted in Fig 12E. As WMC→GC was increased, MC firing rates decreased while GC firing rates increased, eventually crossing. In contrast, the sLFP oscillation frequency remained stable (though unreliable at weights below 1 owing to very low spectral power; Fig 12E). The OI grew rapidly from its arrhythmic values at MC→GC synaptic weights below 1 up to a strong peak value that persisted across a fourfold range of excitatory synaptic weights, decreasing moderately thereafter (Fig 12F). This eventual decline arose as the increased activation of GCs began to impose tonic, as well as phasic, feedback inhibition that further reduced MC activation levels. In contrast, the SI increased steadily as WMC→GC increased, gradually approaching unity at higher synaptic weights (Fig 12F).
Heterogeneity in population activity levels, whether across the neurons of an active ensemble or within a given population over time, poses a challenge to the stability and consistency of dynamical systems [60–65]. For example, the frequencies of gamma oscillations driven by pure PING mechanics vary directly with the activation levels of the excitatory neurons [58], which in the olfactory bulb are strongly heterogeneous (indeed, heterogeneity in MC activation levels is the fundamental basis of olfactory sensory representations). Notably, systems of coupled oscillators often are robust to reasonable heterogeneities in excitation levels [66]; indeed, the essence of coupled oscillator systems is a dynamics by which intrinsic differences in the natural frequencies of constituent oscillators are drawn together into a common limit cycle.
To assess the robustness of the OB network gamma oscillation to variance across the afferent input levels of MCs, we altered the ranges of excitation generated across the MC population by simulated odorant stimuli. By default, steady-state odor input intensities us (nA) were drawn from a uniform distribution within a bounded range (US1, US2). We first varied the upper input bound US2 from 0.4 nA to 1.0 nA with increments of 0.2 nA, with the lower input bound US1 fixed at 0.2 nA (Fig 13). When the upper input bound was reduced from 1.0 nA (in controls) to 0.4 nA, the odor-evoked MC firing rate dropped from 14 Hz to 8.8 Hz and the MC firing rate variance was markedly reduced (SD, us ∈ (0.2, 1.0): 9.8 Hz, us ∈ (0.2, 0.4): 1.6 Hz; Fig 13A). Because of the reduced MC drive, the odor-evoked GC firing rate also declined from 4.6 Hz to 2.4 Hz, and the reduction in GC excitation generated much smaller GABAA conductance fluctuations on MCs (Fig 13B); this feedback response limited the overall change in the balance of excitation and inhibition. Despite these changes in firing rates and the amplitudes of synaptic interactions, MC oscillations remained highly synchronized under both conditions (Fig 13C), and the synchronization index was essentially unchanged (SI, us ∈ (0.2, 1.0): 0.63; us ∈ (0.2, 0.4): 0.62), and the frequency of the dominant sLFP spectral peak was only slightly reduced (us ∈ (0.2, 1.0): 32.4 Hz; us ∈ (0.2, 0.4): 28.7 Hz; compare Fig 13D with Fig 2B).
The mean odor-evoked neuronal firing rates and sLFP oscillation frequencies across a range of upper input bounds are depicted in Fig 13E. As the upper input bound increased from 0.4 nA to 1.0 nA, the mean MC firing rate increased 58.6% (from 8.7 Hz to 13.8 Hz) and that of GCs increased 76.9% (from 2.6 Hz to 4.6 Hz). In contrast, there was only a 14.3% increase in oscillation frequency (from 28.6 Hz to 32.7 Hz), demonstrating the relative robustness of OB gamma frequency to input variance. The synchronization and oscillation indices for the same range of upper input bounds are shown in Fig 13F. Both indices also demonstrated considerable stability in response to changes in the upper input bound.
We next fixed the upper input bound US2 at 1.0 nA, and varied the lower input bound US1 from 0.2 nA to 0.8 nA with increments of 0.2 nA (Fig 14). Increasing the lower input bound reduced input heterogeneity, as in Fig 13, but potentiated rather than reducing the average MC excitation level. When US1 was increased from 0.2 nA to 0.8 nA, the odor-evoked MC firing rate increased from 14 Hz to 24.6 Hz, with markedly reduced variance (SD, us ∈ (0.2, 1.0): 9.8 Hz; us ∈ (0.8, 1.0): 3.6 Hz), leading to highly synchronized MC spikes (SI, us ∈ (0.2, 1.0): 0.63; us ∈ (0.8, 1.0): 0.7; Fig 14A and 14D). Accordingly, a strong, coherent sLFP gamma oscillation was generated with a higher-amplitude spectral peak than that exhibited by controls (Fig 14D; also compare Fig 14B with Fig 2B). However, despite this large increase in the mean MC firing rate, the sLFP oscillation frequency remained remarkably stable (us ∈ (0.2, 1.0): 32.4 Hz; us ∈ (0.8, 1.0): 31.1 Hz; Fig 14C).
Coupled-oscillator networks are able to synchronize oscillators with nonuniform natural frequencies, but this robustness has limitations [61, 62, 66]. The large differences in input activation that can be generated by primary sensory receptor populations (responding to stimuli varying by orders of magnitude in physical intensity and receptive-field optimality) require regulation if they are to be constrained within the limited permissive range of the EPL’s oscillatory regime. Specifically, the range of absolute physiological variability generated in primary sensor populations must be compressed into a dynamic range that does not disrupt the functional dynamics of subsequent sensory system computations. This need is met in the early olfactory system by a series of concentration tolerance mechanisms (reviewed in [67]), culminating in a global normalization computation in the deep glomerular layer ([59]; corrected mechanism in [54]); this computation is mediated by the heterogeneous periglomerular/short-axon cell population [68, 69] and modeled herein by PGCs. To demonstrate the importance of these intensity compression mechanisms and examine the role of PGC-mediated inhibition in enabling OB gamma oscillations, we varied the PGC→MC synaptic weight (WPGC→MC) from 0 to 250% of its default value.
When PGC inhibition was entirely removed (WPGC→MC = 0), the average odor-evoked MC firing rate increased markedly, from 14 Hz (in controls) to 32.2 Hz, inducing a concomitant increase in the mean GC firing rate (from 4.6 Hz to 9.5 Hz; compare Fig 15A with Fig 3D). Firing rates within the MC ensemble also displayed a much larger variance when PGC inhibition was removed (SD, controls: 9.8 Hz; No PGC inhibition: 19.3 Hz; compare Fig 15B with Fig 2A). Importantly, the removal of PGC inhibition significantly degraded MC spike synchrony (SI, controls: 0.64; No PGC inhibition: 0.39); this reduction in SI arose because of the substantial increase in asynchronous background or noisy spiking in MCs (compare Fig 15A, upper panel, with Fig 3D, upper panel). Nevertheless, GC population activity still retained a high level of rhythmicity comparable to controls (compare Fig 15A, lower panel, with Fig 3D, lower panel), and imposed strong phasic inhibition on MCs. Examination of MC and GC population activities indicates that GCs spiked only in response to peak MC spike rates (Fig 15A, dashed vertical lines), and the resulting phasic inhibition from GCs only partially suppressed MC spikes (i.e., MC spikes persisted during peak phasic inhibition), in contrast to the complete periodic suppression of MC spikes by GC inhibition in controls (compare Fig 15A with Fig 3D). Moreover, spike rates in the most strongly driven MCs exceeded the frequency of the underlying STOs, violating the restrictions of coupled oscillator-derived synchrony and consequently wholly desynchronizing with the remainder of the MC population (Fig 15C, lower panel; Fig 15J). Because of the loss of these highly-activated MCs from the synchronous population, the oscillatory power was considerably reduced in the absence of PGC inhibition (compare Fig 15D with Fig 2B), although a sizable spectral peak arising from the less-active MC population still persisted, exhibiting little change in frequency (controls: 32.4 Hz; No PGC inhibition: 34.2 Hz). This result supports two important points: First, although PGC inhibition improves global synchrony–specifically, it improves global participation in the synchronous ensemble by limiting the absolute activation levels of MCs to within a permissive range–it is not required for the generation of the OB gamma rhythm (Fig 15D), whereas GC inhibition is clearly required for OB gamma oscillogenesis (Fig 9B). Second, and critically, these results make clear that this coupled-oscillator mechanism is capable of sustaining coherent oscillations among participating MCs–i.e., those that are both within the permissive band of afferent activation levels and adequately coupled via MC/GC synaptic weights–irrespective of the additional presence of substantial numbers of active MCs that are non-participants in the coherent assembly (Fig 15C). As MCs are known for high levels of background spiking activity, both in vitro and in vivo but especially in awake/behaving animals [70], it is critical to determine the extent to which this activity is likely to interfere with the transmission of neural information. Experimental studies and theoretical models of gamma-timescale coincidence detection in the piriform cortex have suggested that such postsynaptic temporal selectivity will naturally exclude most uncorrelated background activity in MCs from affecting third-order neuronal representations of odor information [71, 72]. However, the present model is the first to demonstrate that timing-based odor representations in the OB can persist in the presence of high levels of uncorrelated background spiking.
Increased PGC inhibition also disrupted OB oscillations (Fig 15E–15H). When the PGC→MC synaptic weight was increased twofold (from 4 to 8), the average odor-evoked MC firing rate decreased from 14 Hz to 4.5 Hz (compare Fig 15F with Fig 2A), reducing the mean GC firing rate from 4.6 Hz to 1.3 Hz. Because of the paucity of activity under this tonic inhibitory suppression, the MC-GC feedback loop was functionally disrupted; GCs responded sparsely and weakly (compare Fig 15E with Fig 3D), evoking weak and irregular GABAergic synaptic conductances onto MCs. MC STOs thereby began to desynchronize and become irregular (compare Fig 15G to Fig 2C), and both gamma rhythm and power were seriously impaired (compare Fig 15H with Fig 2B).
Both the MC and GC mean firing rates decreased rapidly as WPGC→MC increased further, whereas the sLFP oscillation frequency was stable below the control value and declined modestly at higher levels of PGC inhibition (Fig 15I), from 34.5 Hz at WPGC→MC = 0 to 24.2 Hz at WPGC→MC = 10. The synchronization index increased along with the strength of PGC inhibition up until the control value, and remained largely stable under stronger PGC→MC inhibitory weights (Fig 15J). In contrast, the oscillation index peaked around the control value and declined rapidly at higher PGC weights (Fig 15J). The discrepancy between SI and OI at large WPGC→MC values arises largely from the fact that decreasing the numbers of spiking MCs does not reduce the SI, whereas the OI is sensitive to the desynchronization of driver currents and other subthreshold activity occurring among less strongly activated neurons. This highlights the fact that a correspondence between MC spikes and LFP deflections alone does not suffice to ensure coherent gamma oscillations.
These results show that PGC-mediated inhibition can serve to constrain the majority of MCs within a permissive range of activation. This constraint both protects the relational activation differences among MCs that underlie odor quality encoding and enables these odor-activated MCs to participate in a globally coherent gamma-oscillatory ensemble that constrains MC spike timing. Moreover, this globally coordinated oscillation, and the underlying phase-constraint of STOs and spikes in a majority of MCs, is robust to the potentially disruptive impact of highly active but uncorrelated MCs, whether uncorrelated owing to overstimulation or to inadequate coupling.
Our OB network model contained 25 MCs, 25 PGCs and 100 GCs, a small fraction of the number of neurons in the biological OB; additionally, the ratio between the numbers of GCs and the numbers of MCs and PGCs is far greater than is represented in the model [1]. To test whether gamma oscillation in our model was robust to variations in this ratio, we increased the number of GCs (NGC) from 100 to 225 (15*15 array in Fig 1B) and 400 (20*20) respectively, while maintaining the number of MCs and PGCs at 25 each. To correct for the increased total inhibition that would be delivered onto MCs, we scaled down the maximal conductance of individual GC→MC synapses by the same factor such that the total GABAA conductance received by each MC remained relatively constant. When NGC was increased to 225, the mean odor-evoked MC and GC firing rates remained relatively unchanged (controls, MC: 14 Hz, GC: 4.6 Hz; NGC = 225, MC: 13 Hz, GC: 4.2 Hz). Both MC and GC spikes displayed clear synchronization, and MCs displayed appropriately sparse spiking activity (Fig 16A and 16B). A dominant spectral peak in the sLFP power spectrum persisted at almost the same frequency and power as controls (controls: 32. 4 Hz; NGC = 225: 33.6 Hz; compare Fig 16C with Fig 2B).
When NGC was increased to 400, the mean odor-evoked MC and GC firing rates also remained stable (controls, MC: 14 Hz, GC: 4.6 Hz; NGC = 400, MC: 14.2 Hz, GC: 5.2 Hz), and MC activity remained reasonably sparse (Fig 16D and 16E). A strong coherent gamma oscillation again persisted at approximately the same frequency and power as in controls (controls: 32.4 Hz; NGC = 400: 33 Hz; compare Fig 16F with Fig 2B). While this variance does not encompass either the absolute size or the MC-GC ratio of the biological system, it does indicate that gamma oscillations are not highly sensitive to variations in network size.
The olfactory bulb transforms not only the information content of the primary sensory receptor input that it receives, but also its underlying coding metric. Large variance in absolute input amplitudes across receptor populations, varying on a slow respiratory timescale of encoding, are transformed by OB neural circuitry into patterns of ensemble spiking activity among OB principal neurons (mitral cells and projecting tufted cells) that are constrained in their amplitude variance and regulated on a fast gamma-band timescale. This emergent fast timescale for signaling is reflected in the gamma-band sLFP oscillations across the OB that are evoked by afferent activation of OB principal neurons, and presumably serves to efficiently integrate olfactory sensory information into the temporally regulated information networks of the central nervous system.
However, the physiological mechanism underlying this transformation has not been clear. Field potential oscillations at many frequencies are ubiquitous across the brain, and have been attributed to several different underlying dynamical frameworks. Each such theoretical framework imposes predictable relationships and limitations upon the activities of its constituent neurons, and defines the capacities and vulnerabilities of the network to changes in input statistics or internal parameter values. Multiple such frameworks–including PING, ING, STO-driven gamma oscillations, and the PRING hybrid mechanism described herein–have been proposed to underlie OB dynamics; among these, the PRING framework best corresponds to experimental observations of OB circuit neurophysiology [28, 35, 38]. The diagnostic elements of this PRING framework are (1) resonant principal neurons that receive external excitation (unpatterned on the gamma timescale) and exhibit intrinsic STOs, (2) reciprocal connectivity of these principal neurons with spiking inhibitory interneurons that do not separately receive afferent input, (3) a PING-like network oscillation that emerges under afferent activation; its frequency is determined principally by the decay time constant of the GABA(A) receptor conductance and must be higher than that of the STOs, thereby enabling a recurrent reset of STO phase in participating principal neurons, and (4) a continued dependence on principal neuron resonance properties during these network oscillations. In the present simulations, excitatory synapses were spike-mediated; inhibitory synapses were realistically graded but also compatible with GC spiking. Using a biophysically elaborated multiscale computational model of the OB, we here assessed the capacities and limitations of this PRING framework with respect to the observed properties of the OB circuit and the requirements of the olfactory sensory modality.
First, MCs converge onto piriform cortical pyramidal neurons from positions dispersed across the OB; there is no topographical organization to their projection patterns [73]. Coincidence detection in piriform pyramidal neurons [71, 72] requires that spike timing relationships among converging MCs be regulated by a common clock, so that incoming information is not dominated by random variance. Therefore, even physically distant MCs must be regulated by this common clock, indicating that EPL oscillations would need to be coherent across the entire layer, with negligible phase differences among regions. Such spatially extensive zero-phase coherent networks are nontrivial to construct, particularly in the presence of heterogeneous levels of activity among principal neurons. Coupled-oscillator networks in general, and our model here in particular, can yield robust coherence among excitatory neurons with negligible phase drift and across a wide range of physical scales, provided that there is sufficient direct long-distance synaptic coupling between distant columns (as provided here by the long MC lateral dendrites). When long-distance synaptic coupling is reduced in density, the spatial extent of coherence regions in the OB is correspondingly reduced [35], consistent with theoretical predictions [74–76].
Second, the mechanisms generating gamma oscillations should serve to phase-constrain informative MC spike timing, presumably with respect to a timescale appropriate for the synaptic integration time constants of postsynaptic follower neurons. Indeed, MC spikes are phase-constrained at the gamma/beta timescale [10, 13], and their follower neurons in piriform cortex exhibit key properties of coincidence detectors [71]. However, MCs also exhibit high levels of uninformative background spiking, and are particularly active in awake/behaving animals [70]. It is therefore equally important that the oscillogenic mechanism of the OB be robust to high levels of uncorrelated MC spiking. In our model, MC spikes are phase-constrained by virtue of intrinsic STO dynamics [28], which are periodically reset by GABA(A)-ergic synaptic inputs. The dynamical coordination and synchronization of these STOs and spikes across the full OB model is remarkably robust to the impact of high levels of uncoordinated MC spiking input (Fig 15; see also [77]). This robustness, together with the need for multiple convergent inputs to activate piriform pyramidal neurons [78], enables postsynaptic coincidence detectors to selectively respond to informative, temporally-coordinated MC inputs while disregarding MC background activity.
Third, this common frequency and zero-phase coherence must withstand substantial heterogeneity in afferent input levels, both across the network and over time. Heterogeneous networks are a challenge to synchronize [63–65], and, under many mechanisms, differentially-activated local regions of a heterogeneously-activated, spatially extensive network will exhibit different preferred frequencies [28, 33, 35]. Weak coupling has the capacity to pull such regions into a common oscillation, though it is generally effective only across a limited range of preferred frequencies and typically requires several, sometimes many, cycles to achieve synchronization [66, 79–81]. Stronger coupling, such as the STO phase-reset phenomenon of our coupled-oscillator model, enables a rapid, history-independent coordination among diverse local (columnar) oscillators across a range of activation levels [79]. The afferent activation-dependent differences among MCs in the rate of their recovery from synchronous GC-mediated synaptic inhibition have been proposed to generate the spike phase code exported from the OB [49, 72]; however, for present purposes, the important factor is that this coupling mode renders global sLFP synchronization robust to the large differences in afferent activation levels that together constitute the primary sensory representation (Figs 2, 13 and 14). Some dynamical frameworks also are not robust to inhibitory neurons that spike, or to networks in which excitatory or inhibitory neurons fire at dissimilar rates, or at rates far below the common oscillatory frequency. All of these phenomena are features of the OB network, and are robustly supported by the present model. Finally, global synchronization across the OB must also be robust to sparse network connectivity, and to substantial differences in synaptic weights across the EPL, particularly the excitatory synaptic weights that are modified during the process of odor learning [47, 48]. The present model maintains stable oscillations and global synchronization with sparse connections and a wide range of excitatory synaptic weights (Figs 11D–11F and 12).
Fourth, notwithstanding the above, there clearly are limits to the range of absolute input amplitudes that a dynamical system can withstand. The effects of afferent input intensity (concentration) are mitigated in animals by a series of compensatory mechanisms [67] capped by a global normalization network embedded in the OB glomerular layer, essentially feeding back a global average of input intensity as inhibition onto all MCs. This global normalization function was proposed a decade ago [51, 59], but the underlying circuit mechanism has only recently been determined [54]. In the model, as predicted, reduction of this circuit-based concentration tolerance by modifying PGC inhibition increased mean activity and variance across the MC population and disrupted spike synchronization (Figs 8 and 15).
PRING oscillations exhibit these diverse and computationally important properties by virtue of their integration of PING and STO mechanics. Two prior conductance-based network models of OB gamma oscillations also have incorporated both synaptic and STO dynamics [22, 33], but each reached different conclusions owing to differences in implementation. The earlier of these models, by Bathellier et al. [22], incorporated STO dynamics in single-compartment MCs, but did not include explicit GCs; instead, MC spikes directly generated recurrent and lateral inhibition, and there was no graded contribution to synaptic inhibition. In this model, the resonant properties of MCs were found to play little role in the gamma oscillation, and the population frequency depended on the rising time constant (rather than the decay time constant) of lateral synaptic inhibition. The second such model, by Brea et al. [33], incorporated explicit MCs and GCs, and exhibited both MC STO dynamics and graded synaptic inhibition. The Brea model demonstrated that STOs can be synchronized by graded inhibition, exhibited some STO resetting by this inhibition, and allowed mean MC firing rates to be much lower than the population oscillation frequency. However, it also differed from the present PRING model in several ways. First, in the Brea model, intrinsic STO frequencies directly drove the population oscillation frequency; the time constants of synaptic inhibition played little role. To accomplish this, MC STO frequencies were raised to 60–90 Hz, significantly higher than the 20–40 Hz that has been observed experimentally [28] and implemented in the present model. In principle, these high-frequency STOs could prevent the slower synaptic inhibition from determining the population frequency of the active network, as illustrated above (Fig 7E); however, in the Brea model, the STOs directly determined network frequency even when slowed to 35 Hz (Fig S5 in [33]). Differences in the properties of synaptic inhibition and GC spiking are more likely to be the main differentiating factors. Second, synaptic inhibition in the Brea model was activated at relatively hyperpolarized potentials (-66 mV), exhibited a relatively hard threshold (activated between -66.5 mV and -65.5 mV; Fig 1A of [33]), and was delivered directly to the single somatic compartment of the model cell. In contrast, in the present model, half-activation of the graded inhibitory synapses occurred at -40 mV, the threshold was much softer (activated between -50 mV and -30 mV), and incoming inhibitory synapses were distributed along an electrotonically extensive lateral dendrite. Third, the Brea model was not readily compatible with sparse GC spiking (i.e., GCs that spike at substantially lower frequencies than the population oscillation); in contrast, the present PRING model robustly supports sparse GC spiking during population oscillations.
In sum, the present model demonstrates that the PRING mechanism elucidated in the OB network by [35], when embedded in a multiscale, dynamical biophysical model of MC circuit function, exhibits the full set of dynamical properties that either have been experimentally demonstrated in the OB or are critical theoretical predictions based on experimental data. These experiments demonstrate that OB dynamics can be best described as independent columnar oscillators, coupled by pulsed inhibition, with a network topology based on long-distance, non-topographically organized connections. This elucidation of the essential dynamics of OB oscillogenesis will substantially constrain the plausible mechanistic hypotheses for interareal dynamics, such as the transient coherence in the beta band between OB and piriform cortex that characterizes particular phases of olfactory investigation.
The “default” OB network model contained 25 mitral cells (MCs), 25 periglomerular cells (PGCs) cells and 100 granule cells (GCs; [25]). Each MC, together with an associated PGC, represented a separate OB column, each of which was associated with a particular glomerulus and hence a distinct olfactory receptor type. The number of GCs in the model was increased substantially in certain simulations. The MC, PGC and GC single-cell models were Hodgkin-Huxley type conductance-based compartmental models based on those in [25]. In contrast to the 2013 model, the present OB network incorporated physical locations for each OB column in order to model the problems of distance-dependent lateral interactions, such as the differing propagation delays of spikes along MC lateral dendrites [50]. Specifically, the OB surface was modeled as a two-dimensional (2D) space (1 mm x 1 mm), upon which MCs and PGCs (together) and GCs (separately) were arranged in grid arrays with equal spacing in the horizontal and vertical directions (Fig 1B). To avoid edge effects, the 2D network was mapped onto a torus. Each neuron was labeled with its column and row numbers in the 2D space starting from 0 (i.e., MC[i][j] denoted the MC in the ith column and the jth row). In some figures, model neurons were denoted by a single index to enable their distribution along a single axis (e.g., in raster plots). In such cases, that single index z was related to the two indices i and j as follows: z = N * i + j + 1, where N was 5 for MCs and PGCs and 10 for GCs.
Both MC-PGC and MC-GC connections incorporated dendrodendritic synapses (Fig 1A; [25]). In the model, each MC formed reciprocal synapses with its local PGC (associated with the same glomerulus); i.e., the MC excited the PGC dendritic spine whereas the PGC inhibited the MC tuft compartment via graded inhibition (Fig 1A). MCs also interacted bidirectionally with GCs along the lengths of the MCs’ lateral (secondary) dendrites, which extend for long distances across the olfactory bulb [55, 82]. Specifically, MCs delivered synaptic excitation onto GC dendritic spines while receiving feedback and lateral inhibition from these same spines (Fig 1A). Each MC connected reciprocally to a random selection of GC dendrites with a connection probability p = 0.3. To model the cable effects of distance, the location of the dendrodendritic contact along the length of the seven-compartment MC lateral dendrite [25] was determined by the distance between the MC soma and the GC in question (Fig 1B).
The MC→PGC and MC→GC synapses were mediated by both AMPA and NMDA receptors, whereas the PGC→MC and GC→MC synapses were mediated by GABAA receptors. Postsynaptic currents were modeled as in [25]:
Isyn=WgsynsB(V)(V−Esyn)
(1)
where gsyn is the maximal synaptic conductance (prior to weighting) and Esyn is the reversal potential (0 mV for AMPA/NMDA currents and -80 mV for GABAA currents; [25]). The maximum synaptic conductances were: gAMPA = 2 nS and gNMDA = 1 nS for both MC→PGC and MC→GC synapses, and gGABA = 2 nS for both PGC→MC and GC→MC synapses [25]. W denotes the synaptic weight, which scaled the maximum preweighting synaptic conductance so as to generate final maximum synaptic conductances of W*gsyn. The synaptic weight was varied systematically in simulations; default synaptic weights were: WMC→PGC = 1, WMC→GC = 1, WPGC→MC = 4, and WGC→MC = 2 (arbitrary units). The function B(V) implemented the Mg2+ block for the NMDA current, and was defined as B(V) = (1 + [Mg2+]exp(−0.062V)/3.57)−1 83]. For AMPA and GABAA currents, B(V) = 1. The gating variable s represented the fraction of open synaptic ion channels and obeyed first-order kinetics [84, 85]:
dsdt=αF(Vpre)(1−s)−βs
(2)
where F(Vpre) was an instantaneous sigmoidal function of the presynaptic membrane potential, F(Vpre) = 1/(1 + exp(−(Vpre –θsyn)/σ)). The half-activation potential (θsyn) of the synapse was set to 0 mV for AMPA/NMDA receptor synapses and -40 mV for GABAA synapses; the parameter σ was set to 0.2 for AMPA/NMDA currents and 2.0 for GABAA currents [25]. Consequently, synaptic excitation was triggered mostly by spikes (high threshold), whereas synaptic inhibition occurred below spiking threshold and depended on presynaptic voltage in a graded manner. The channel opening rate constants (α and β) were expressed as α = 1/τα and β = 1/τβ, where τα and τβ were the synaptic rise and decay time constants respectively. For AMPA receptor currents, τα = 1 ms, τβ = 5.5 ms; for NMDA receptor currents, τα = 52 ms, τβ = 343 ms; and for GABAA receptor currents, τα = 1.25 ms, τβ = 18 ms [25]. Such first-order synaptic models naturally simulate the interactions of successive presynaptic events, enabling the saturation of slow synapses [85]. Specifically, with a slow rising time constant of 52 ms, the NMDA conductance increased only slightly in response to a single presynaptic spike, but accumulated over multiple synaptic inputs owing to its slow decay time constant of 343 ms, limited by the maximum synaptic conductance. Different synaptic decay time constants have been reported by experimental studies in the OB [20, 23, 86]; importantly, the modeled time constants represent in part the “functional time constants” generated by a quasisynchronously activated population of presynaptic synapses affecting the same postsynaptic neuron. Some of these parameters were varied for purposes of particular simulations as described in the Results; in those cases, the parameter values specified above are referred to as “default” or “control” values.
Odor stimulation was modeled as in [25]. A sigmoidal function was used to model OSN inputs [33]:
IOSN=u0+0.5(us−u0)[tanh(3(t−tORN)r−3)+1]
(3)
where u0 was the pre-odor value (simulated pure air input) and us the steady-state value after odor excitation. The parameter r determined the transition rate from u0 to us (set to 100) and torn the time of odor onset. Different MCs (representing separate, independently-tuned glomeruli) received different levels of afferent activation; the corresponding values of u0 and us were drawn from uniform distributions u0 ∈ (0.1, 0.2) and us ∈ (0.2, 1.0). Additionally, all cells in the network received random excitatory inputs representing intrinsic and extrinsic sources of uncorrelated background noise. These nonspecific inputs were modeled as uncorrelated Poisson spike trains mediated exclusively by AMPA receptors; specifically, they comprised instantaneous steps followed by exponential decays with a time constant of 5.5 ms [25]. When plotting major network measures (e.g., MC/GC firing rates, oscillation frequencies, synchronization and oscillation indices) under variable parameter sets, the data reported were averaged across 10 instantiations of the network with different random seeds for these Poisson spike trains.
A simulated local field potential (sLFP) was constructed by filtering the mean (somatic) membrane potentials across all MCs [25]. Filtering was carried out numerically using a band-pass filter (10–100 Hz) with the MATLAB functions FIR1 and FILTFILT [33]. The power spectrum of the signal was obtained by a fast Fourier transform (FFT) of the filtered sLFP. MC somatic spike times were converted to spike phases using the method detailed in [25].
The synchronization (or phase-locking) index was calculated as follows [22]:
κ=1/N[∑i=1Nsin(φi)]2+[∑i=1Ncos(φi)]2
(4)
where φi was the phase of each MC spike in the network relative to the sLFP peak. This synchronization index (SI) measures the degree of phase locking between MC spikes and sLFP oscillations rather than the absolute synchrony of MC spikes in time. Nevertheless, when substantial numbers of MC spikes are evoked, the SI also is a good measure of absolute spike synchrony. When all MC spikes have identical phases, the index achieves its maximal value of unity. The oscillation index (OI) corresponded to the peak of the frequency power spectrum of the sLFP, which was normalized to the largest peak value generated from ten sets of simulations with different random seeds. The oscillation frequency was determined from the position of the spectral peak in the power spectrum [25].
The model was implemented in the neuronal simulator package NEURON, version 7.3 [87], using the Crank-Nicholson integration method and a fixed timestep of 2 μsec (0.002 ms). Shorter timesteps did not change the results. Simulations were run both on a workstation under CentOS Linux and on Linux clusters provided by the Cornell Computational Biology Service Unit’s High Performance Computing laboratory (BioHPC). Simulation output data were saved in files and analyzed using custom Matlab scripts.
|
10.1371/journal.pgen.1005225 | Casein Kinase 1 and Phosphorylation of Cohesin Subunit Rec11 (SA3) Promote Meiotic Recombination through Linear Element Formation | Proper meiotic chromosome segregation, essential for sexual reproduction, requires timely formation and removal of sister chromatid cohesion and crossing-over between homologs. Early in meiosis cohesins hold sisters together and also promote formation of DNA double-strand breaks, obligate precursors to crossovers. Later, cohesin cleavage allows chromosome segregation. We show that in fission yeast redundant casein kinase 1 homologs, Hhp1 and Hhp2, previously shown to regulate segregation via phosphorylation of the Rec8 cohesin subunit, are also required for high-level meiotic DNA breakage and recombination. Unexpectedly, these kinases also mediate phosphorylation of a different meiosis-specific cohesin subunit Rec11. This phosphorylation in turn leads to loading of linear element proteins Rec10 and Rec27, related to synaptonemal complex proteins of other species, and thereby promotes DNA breakage and recombination. Our results provide novel insights into the regulation of chromosomal features required for crossing-over and successful reproduction. The mammalian functional homolog of Rec11 (STAG3) is also phosphorylated during meiosis and appears to be required for fertility, indicating wide conservation of the meiotic events reported here.
| The formation of haploid gametes (sex cells, such as eggs and sperm) from diploid precursor cells involves two nuclear divisions but one round of chromosomal replication. In the unique first meiotic division, centromeres of sister chromatids remain connected and homologous chromosomes (homologs) segregate from each other. In most species proper homolog segregation requires that crossover recombination occur between homologs to impart tension between homologs as they move apart. A protein kinase (casein kinase 1) has long been known to regulate proper sister centromere connections by phosphorylating Rec8, a meiosis-specific sister chromatid cohesin subunit. We report here that in fission yeast this kinase has a second critical role—to mediate phosphorylation of another meiosis-specific cohesin subunit Rec11. Phosphorylation of Rec11 enhances loading of two meiosis-specific components of linear elements, which are related to the synaptonemal complex and help pair homologs. These linear element proteins lead to high-level DNA breakage and crossovers between homologs. Thus, casein kinase regulates two crucial but separate events in meiosis. The mammalian functional homolog of Rec11, called STAG3, is also phosphorylated during meiosis and appears to be required for fertility in human females. These observations suggest wide-spread conservation of the roles of casein kinase 1 and Rec11 in ensuring proper meiotic chromosome segregation and sexual reproduction.
| The two specialized nuclear divisions during meiosis convert a diploid precursor cell into one or more haploid cells (gametes). Uniquely in the first meiotic division, centromeres of homologous chromosomes (homologs) segregate from each other, whereas centromeres of sister chromatids segregate from each other only in the second meiotic division, as in mitotic divisions. Proper chromosome segregation is essential for the formation of gametes with viable chromosome complements and requires two chromosomal events special to meiosis—crossing-over between homologs, and spatially and temporally regulated cohesion between sister chromatids. For homologs to segregate properly from each other in the first meiotic division, they must pair and the sister centromeres must remain connected and move as a unit to one pole of the cell; in most species pairing is accompanied by physical exchange of DNA between homologs to form crossovers that provide the interhomolog tension required for their proper segregation. For sister centromeres to segregate properly from each other in the second meiotic division, pericentric cohesion must be established in the first division but be removed only during the second division. This is accomplished by protection of a meiosis-specific cohesin subunit only at and near the centromere before and during the first division and its degradation specifically just before the second division.
Although the outlines of these two events are known [1], how they are regulated and coordinated remains unclear. Previous data and our new results reported here show that casein kinase 1 plays an essential role in both of these events, by mediating the phosphorylation of two separate subunits of cohesin. Our observations are in fission yeast, but the wide-spread conservation of these subunits suggests that our conclusions apply broadly to eukaryotes.
Meiotic cohesin is a large protein complex composed of Smc1 and Smc3, which are common to both the mitotic and meiotic forms, and the meiosis-specific Rec8 and Rec11 subunits [2–5]. (Some species, including budding yeast and Tetrahymena, lack a clear Rec11 homolog and retain the mitotic form, Scc3, in meiosis [6,7].) This complex forms a ring that connects sister chromatids from the time of replication to the time of chromatid segregation [8,9]. To allow segregation, Rec8 is phosphorylated and then cleaved by a protease called separase [10]. During meiosis, Rec8 cleavage occurs in two steps: along the chromosome arms during the first meiotic division (MI) and in the pericentric region during the second meiotic division (MII) [11]. In MI pericentric Rec8 is protected by shugoshin (Sgo1), which recruits PP2A protein phosphatase and thereby prevents Rec8 cleavage. During MII, cohesin is no longer protected from separase, which then cleaves pericentric Rec8 to allow sister centromere segregation [12]. In the absence of Rec8, chromosome segregation is like that in mitosis: sister chromatids, rather than homologs, segregate at MI [13]. In the absence of Rec11, chromosome segregation is similar to that in a recombination-deficient mutant: sister centromeres remain connected until MII, when they segregate, but aberrant arm cohesion and a paucity of crossovers reduce proper homolog segregation at MI [5].
Rec8 and Rec11 are also required for formation of crossovers, which result from the repair of DNA double-strand breaks (DSBs) programmed to occur during meiosis [14,15]. DSBs are made by the highly conserved topoisomerase-like protein Spo11 (named Rec12 in fission yeast) [16]. To be active, Rec12 requires six essential partner proteins, which likely function as a large complex similar to that of the Spo11 complex of budding yeast [17]. In a proposed pathway, Rec8 and Rec11 cohesin subunits are loaded onto chromosomes during S phase [14,18]. Their loading allows loading of the linear element (LinE) complex, related to the synaptonemal complex of other species; LinEs contain Rec10, Rec25, Rec27, and Mug20 [14,19–21]. In accord with this pathway, LinEs are rare or absent in each of the six corresponding mutants [14,19–22]. LinE protein loading activates the Rec12 complex to make DSBs [14]. Although Rec10 is required for DSB-formation and recombination throughout the genome, the other three LinE proteins, like Rec8 and Rec11, are required more strongly in some chromosomal intervals than in others [14].
Cleavage of Rec8 to allow proper chromosome segregation requires phosphorylation of Rec8 by one or more protein kinases, including casein kinase 1 orthologs in both budding and fission yeasts [23]. In fission yeast two casein kinase 1 paralogs, Hhp1 and Hhp2 (collectively called Hhp here), function redundantly to phosphorylate Rec8 [12,23]. In their absence Rec8 cleavage is delayed and persistent sister-chromatid cohesion along chromosome arms often prevents chromosome segregation, leading to many fewer viable gametes (spores) than in wild-type cells. Because of the close connection between meiotic chromosome segregation and recombination, exemplified by the role of Rec8 in both processes [13,15], we examined meiotic recombination in Hhp-deficient mutants. We found that Hhp is indeed required for recombination but that the substrate for this process is, unexpectedly, the meiosis-specific cohesin subunit Rec11, not Rec8. Our findings indicate that Hhp regulates chromosome segregation and recombination separately, by regulating Rec8 cleavage and by activating Rec11 to promote DSB-formation and recombination. We discuss parallels in the roles of meiotic cohesin subunits common to fission yeast and mammals.
To test for a possible role of casein kinase 1 homologs Hhp1 and Hhp2 in meiotic recombination of the fission yeast Schizosaccharomyces pombe, we measured recombination in hhp1Δ hhp2Δ double deletion mutants. Recombination was reduced by factors ranging from about 5 to 170, depending on the interval measured (Table 1). Both intergenic recombination (crossing over) and intragenic recombination (gene conversion) were reduced in the double mutant but only slightly in each single mutant, indicating that Hhp1 and Hhp2 have redundant roles in recombination, as previously reported for chromosome segregation [12]. Similar differential reductions, depending on the interval measured, are observed in cohesin- and LinE-deficient mutants [19,24,25], leading us to suspect that the Hhp substrate required for recombination is a cohesin subunit or LinE protein. Our results below bear out this suspicion.
Because mitotic growth and viable spore yield are severely impaired in hhp1Δ hhp2Δ mutants [26,27], we used an Hhp1 mutant (hhp1-as encoding the Met84 → Gly alteration) sensitive to 1-NM-PP1, an analog of the purine moiety of ATP, by alteration of its ATP-binding site [28,29]. In conjunction with hhp2Δ we could thereby allow Hhp function (in the absence of analog) or block Hhp function (in its presence). As expected from the results above, recombination was strongly reduced in the presence of the analog: we observed ~15- and 100-fold reductions in the two intervals tested (Table 2), comparable to the reductions seen in hhp1Δ hhp2Δ. Recombination was also reduced to essentially the same extent in the absence of the analog, a condition that allowed much higher viable spore yield and nearly wild-type chromosome segregation (Table 2 and S1 Table). This fortuitous result, presumably a reflection of hhp1-as allowing adequate phosphorylation of some substrates but not others, allowed us to conduct experiments under conditions allowing nearly wild-type mitotic growth and high viable spore yield. We discuss later the putative separation of functions of hhp1-as.
When we coupled the hhp1-as hhp2Δ mutations with rec12Δ, we obtained results indicating that, in the absence of analog, recombination but not Rec8 cleavage was defective in the hhp1-as hhp2Δ mutant. Because S. pombe has only three chromosomes and has a mechanism that enhances proper segregation of non-recombinant homologs [30], rec12Δ mutants form ~25% as many viable spores as wild type (Table 2) [31]. This yield was not further reduced by the hhp1-as hhp2Δ mutations in the absence of analog, but it was greatly reduced in its presence, as expected from failure of Rec8 to be cleaved under this condition [12,32]. In the absence of analog, the failure of rec12Δ to further reduce the viable spore yield of hhp1-as hhp2Δ indicates that without analog hhp1-as hhp2Δ blocks recombination but not chromosome segregation. We infer that one or more recombination-promoting protein(s) is not properly phosphorylated by Hhp1-as in the absence of the analog and that in the presence of the analog Rec8 is also hypo-phosphorylated.
Meiotic recombination requires both formation and repair of DSBs. hhp1Δ hhp2Δ mutants have DNA repair defects in vegetative cells [26]. To determine if meiotic DSB repair is blocked in the hhp1-as hhp2Δ mutant in the absence of the analog, we artificially introduced DSBs with the I-SceI homing endonuclease, controlled by the meiosis-specific rec12 promoter [33]. DSBs were introduced in the ade6 gene at the site of the ade6-3061 mutation, which can recombine with another mutation ade6-52 to generate Ade+ recombinants. In the absence of analog the frequency of recombinants was indistinguishable in wild type and in hhp1-as hhp2Δ, indicating that under this condition Hhp is not required for DSB repair (S2 Table). rec8Δ, however, reduced the recombinant frequency by a factor of ~3, suggesting that Rec8 is required for DSB repair of I-SceI DSBs, as it is at some chromosomal sites in budding yeast [6]. We infer that during meiosis Hhp is required for DSB formation, but we found no evidence for its having a role in DSB repair under this condition. This conclusion is consistent with the hhp mutant having high viable spore yield but low recombination-proficiency in the absence of analog (Table 2).
To directly test for a role of Hhp in DSB formation, we assayed DSBs by Southern blot hybridizations of DNA extracted from hhp1-as hhp2Δ mutants induced for meiosis in the absence of analog; DNA from wild type and rec8Δ was analyzed for comparison (Fig 1 and S1 Fig). In wild type there were six prominent, meiosis-dependent DSB hotspots on the 501 kb NotI restriction fragment J, as seen before [34,35]. DSBs were barely detectable at these sites in hhp1-as hhp2Δ, as is the case in rec8Δ and rec11Δ [14,18]. DSBs were also barely detectable in hhp1-as hhp2Δ at the strong DSB hotspot ade6-3049, as is the case in rec8Δ and rec11Δ [14,18]. In contrast, DSBs were detectable, though at reduced levels, at some hotspots on the 1500 kb NotI restriction fragment C in hhp1-as hhp2Δ and in rec8Δ and rec11Δ (Fig 1B) [14,18]. These results show that Hhp is required for most meiotic DSB formation, but residual DSBs with a spatial pattern similar to that in rec8Δ and rec11Δ, which are indistinguishable [14,18], remain in the absence of Hhp function.
Because the residual patterns of DSB formation and recombination in hhp1-as hhp2Δ resemble those in rec8Δ (Fig 1) [14,18] and because Rec8 is an Hhp substrate [12,23], we tested the hypothesis that the Hhp substrate required for DSB formation is Rec8. One of the Hhp-dependent phosphorylation sites on Rec8 (S412) is critical for cleavage of Rec8 to allow chromosome segregation [12]. The non-phosphorylatable mutant rec8-S412A was, however, as recombination-proficient as wild type (S3 Table) and had DSB patterns on NotI fragments J and D similar to those of wild type. Similar recombination results were obtained with six additional rec8 mutants lacking seven, 12, 13, 17, or 18 phosphorylation sites (S3 Table). In S. cerevisiae, Rec8 phosphorylation is also not essential for meiotic recombination, although crossing-over is about one-half as frequent and delayed relative to wild type [36]. These results suggest that Rec8 is not, under the conditions used here, the major substrate of Hhp required for DSB formation and recombination, although minor effects cannot be ruled out.
To search for additional Hhp substrates during meiosis, we immunoprecipitated TAP-tagged versions of Hhp1 and Hhp2 and analyzed the precipitated proteins by mass spectrometry (S2 Fig, S4 Table and S5 Table). In meiotically induced cells, but not in mitotically growing cells, we found that Hhp1 co-immunoprecipitated with Hhp2-TAP and, conversely, Hhp2 with Hhp1-TAP. We confirmed this interaction in meiotic extracts by standard co-immunoprecipitation and Western blotting (S3 Fig). Physical interaction between Hhp1 and Hhp2 was previously observed in checkpoint-activated mitotic cells [37]. Importantly, we found the cohesin subunits Rec8, Psm1, and Psm3 (orthologs of Smc1 and Smc3 in other species) in precipitates of Hhp1 and Hhp2 from meiotic cells (S4 Table). To test the possibility that other cohesin subunits interact with Hhp, we analyzed Rec11-TAP precipitates, in which we found known cohesin subunits and Hhp2, suggesting that cohesin and Hhp physically interact. Interestingly, we did not detect interaction between Hhp and cohesin in extracts from mitotically growing cells, suggesting that this interaction is stronger during meiosis or is meiosis-specific (S4 Table).
To test directly for phosphorylation of Rec11, we determined the mobility of Rec11-TAP by gel electrophoresis (Fig 2). The mobility of Rec11 from wild-type cells was considerably increased by phosphatase treatment (Fig 2A, lanes 1 and 2 vs. lanes 3 and 4), indicating that Rec11 indeed was phosphorylated during meiosis. In the presence of analog the mobility of Rec11 from the hhp1-as hhp2Δ mutant induced into meiosis was greater than that from wild-type cells (lane 5 vs. lane 1); its mobility was further increased by phosphatase treatment (lanes 7 and 8 vs. lanes 5 and 6). As expected, the mobility of Rec11 from wild type (hhp+) was the same with or without analog (lane 1 vs. lane 2), demonstrating that the increased mobility of Rec11 from the mutant was due to inhibition of Hhp and not to an off-target effect. In the absence of analog the mobility of Rec11 from the hhp1 mutant was also greater than that from wild type (Fig 2B, lanes 1 and 4 vs. lane 2) and was further increased by adding analog (Fig 2B, lane 2 vs. lane 3), as also evident in Fig 2A (lane 5 vs. lane 6).
These results indicate that Rec11 phosphorylation in the hhp1-as hhp2Δ mutant is reduced even in the absence of the inhibitor, consistent with the hhp1-as hhp2Δ mutant having a recombination-deficient phenotype in the absence of inhibitor (Table 2). They also indicate that Rec11 phosphorylation is reduced even more, but not completely, in the presence of analog. This outcome is consistent with viable spore yield being somewhat reduced in the absence of analog but reduced much more in the presence of analog (Table 2). Residual Rec11 phosphorylation in the presence of analog (Fig 2A, lane 5 vs. lane 7) may depend on residual Hhp1-as function or on another protein kinase.
To determine the nature of Rec11 phosphorylation during meiosis, we analyzed the immunoprecipitates of Rec11-TAP for phosphopeptides via mass spectrometry in two independent experiments. We found that Rec11 is phosphorylated on eight serine (S10, S22, S34, S43, S150, S439, S496 and S880) and two threonine (T60 and T70) residues during meiosis (Fig 2C).
To test the potential functional significance of Rec11 phosphorylation, we generated rec11 mutants encoding alanine at these ten phosphorylation sites (allele rec11-10A) or the phosphomimetic aspartate at those positions (rec11-10D) (Fig 2C). The recombination-proficiency of the rec11-10A mutant was lower than that of wild type by factors of ~5–10, whereas that of the rec11-10D mutant was near that of wild type (Tables 3). These data indicate that phosphorylation of Rec11 at one or more of these sites is important for recombination. [Since in recombination assays the phosphomimetic aspartate is less deleterious than alanine (Table 3), we presume that the deficiency in the rec11-10A mutant results from reduced phosphorylation, although other structural deficiencies of Rec11 cannot be excluded. The abundance of both Rec11 mutant proteins during meiosis was similar to that of wild-type Rec11 (S4 Fig), indicating that the mutations do not significantly alter the stability of Rec11.] Because the recombination-deficiency of the rec11-10A mutants was not as great as that of the rec11Δ null mutant or of the hhp1-as hhp2Δ mutant in the absence of analog, Rec11 presumably has additional phosphorylation sites important for its promotion of recombination; these may be the residual sites phosphorylated in hhp1-as hhp2Δ (Fig 2). Alternatively, Rec11 may have phosphorylation-independent functions important for recombination, still present in the rec11-10A mutants but lacking in rec11Δ.
If phosphorylation of Rec11 depends on Hhp and is important for recombination, the rec11 phosphomimetic mutation, rec11-10D, might suppress the loss of Hhp function in the hhp1-as hhp2Δ mutant. Indeed, rec11-10D partially suppressed hhp1-as hhp2Δ for both gene conversion (ade6 intragenic recombination) and crossing-over (ade6—arg1 intergenic recombination). The rec11-10D mutation slightly raised the Ade+ recombinant frequency in the hhp1-as hhp2Δ mutant from 28 to 51 (per million viable spores; p = 0.05 by one-tailed t-test) and the ade6—arg1 crossover distance from 2.0 to 5.3 (cM; p < 0.001 by Fisher’s exact test) (Table 3). Suppression may be only partial owing to sites of phosphorylation other than those mutated in the rec11-10D mutant or to the phosphomimetic mutations being only partially effective, as is frequently observed [38,39].
As expected from the greater recombination-proficiency of rec11-10D than of rec11-10A, DSBs were more abundant in rec11-10D than in rec11-10A (Fig 1, middle two pairs of lanes in each panel). Only at one DSB hotspot, denoted “c” on NotI fragment C (Fig 1B, middle panel), were DSBs close to wild-type levels in rec11-10A; DSBs at this hotspot were also most abundant in hhp1-as hhp2Δ, rec8Δ, and rec11Δ (Fig 1) [14,18], showing that DSB formation at this hotspot is largely independent of these factors. At other hotspots, DSBs were slightly reduced in rec11-10D relative to wild type, but they were reduced more in rec11-10A though in most cases not to the level in hhp1-as hhp2Δ, rec8Δ, or rec11Δ (as noted above, DSB levels are indistinguishable in rec8Δ and rec11Δ [14,18]). These data are in accord with the recombination data in Table 3 —the rec11-10A mutation reduces both DSB formation and recombination more than the rec11-10D mutation but not as much as the hhp1-as hhp2Δ, rec8Δ, and rec11Δ mutations.
Because Hhp is required for cohesin removal and proper meiotic divisions [12,40], we tested whether Rec11 phosphorylation is required for these steps of meiosis. Although Rec8 phosphorylation is essential for its cleavage and removal and thus for segregation of chromosomes during meiosis I [12,40], meiotic divisions occurred normally in Rec11 phosphorylation-site mutants 10A and 10D, and no defect in Rec8 removal at the onset of anaphase I was observed (S5 Fig). Mutating one of the two separase cleavage sites on Rec8 (Rec8-RD1) leads to only a minor defect in chromosome segregation during meiosis [41]. We did not observe any further impairment of meiotic chromosome segregation when we analyzed cells expressing both Rec11-10A and Rec8-RD1 (S6 Fig). These results suggest that the Rec11 phosphorylation sites identified in our study are required for meiotic recombination but not for segregation of chromosomes during meiotic divisions.
Rec8 and Rec11 appear to act before the LinE proteins, because the formation of LinEs or nuclear LinE protein foci largely depends on Rec8 and Rec11, but Rec8 focus-formation does not depend on LinE proteins [22]; Rec11 focus-formation has not, to our knowledge, been similarly tested. We therefore tested the dependence on Hhp of focus-formation by LinE proteins, using fluorescence microscopy of Rec10-GFP and Rec27-GFP, both of which are functional [14,19,20]. The number of foci formed by each protein and the fraction of cells with visible foci were significantly reduced in hhp1-as hhp2Δ in the absence of analog (p < 0.01 by Fisher’s exact test) as well as in rec11-10A (p < 0.01) (Fig 3 and S7 Fig). Comparison of either time point for wt (3 and 3.5 hr) with either time point for hhp1-as hhp2∆ (4 and 4.5 hr) or for rec11-10A (3 and 3.5 hr) shows that the mutants have fewer foci and cells with foci. [We used later time points for the hhp1-as hhp2∆ mutant because meiosis is delayed in the mutant (S8 Fig).] The residual levels were similar to those reported in rec8Δ and rec11Δ mutants, which behave similarly in other recombination-related assays [19,20]. In contrast, Rec11-GFP appeared to localize to the nucleus to nearly the same extent in wild type and in hhp1-as hhp2Δ (S9 Fig) but not in rec8∆ (S10 Fig) as noted previously [42]. Rec11-GFP did not consistently form distinct foci, making quantification difficult. Nevertheless, the frequency of nuclei with Rec11-GFP fluorescence appeared to be similar in wild type and in hhp1-as hhp2Δ. These results show that the loading of LinE proteins depends on Hhp and likely on the phosphorylation of Rec11, which itself depends on Hhp (Fig 2).
Our results reveal that casein kinase 1 homologs in fission yeast, Hhp1 and Hhp2 (Hhp), have, in addition to their known substrate Rec8 [12,40], a second substrate that must be phosphorylated by Hhp during meiosis to promote DSB formation and recombination. We infer that the second substrate is Rec11, since inactivating Rec11 phosphorylation sites (by Ser or Thr → Ala changes) reduced recombination more in wild type cells than in Hhp-deficient cells (Table 3). Furthermore, phosphomimetic alterations (Ser or Thr → Asp) in Rec11 left the cells recombination-proficient, though not quite to the wild-type level, and partially suppressed the recombination-deficiency of Hhp-deficient cells (Table 3). Rec8 mutants lacking Hhp-dependent phosphorylation sites are deficient for cohesin cleavage [12,40] but are recombination-proficient (S3 Table). Conversely, Rec11 mutants lacking phosphorylation sites are deficient for recombination and DSB formation (Table 3; Fig 1 and S1 Fig) but are segregation-proficient (S5 Fig and S6 Fig). Therefore, our data combined with the cited published data show that Hhp phosphorylates Rec8 to regulate cohesin cleavage for proper chromosome segregation and mediates phosphorylation of Rec11 to activate DSB formation for recombination. A recent independent report drew the same conclusion [43]. These two actions of Hhp are separable, since lack of Rec8 phosphorylation leaves recombination (Rec11 action) intact (S3 Table) and reduction of Rec11 phosphorylation leaves chromosome segregation (Rec8 action) intact (S5 Fig and S6 Fig). Below, we discuss the implications of these findings for the mechanism of meiotic recombination and chromosome segregation, and their co-ordination. The conservation of Rec8 and Rec11 in most species suggests that these two separate roles of the cohesin subunits regulate meiotic chromosome dynamics in widely divergent species, including humans.
We were surprised that the hhp1-as (M84G) ATP-analog-sensitive mutant had a dramatic, differential phenotype even in the absence of added analog (Tables 2 and 3; Figs 1, 2, and 3)—it strongly reduced recombination but had much less effect on viable spore yield or chromosome segregation (Tables 2 and 3). We infer that this mutation differentially alters the substrate specificity or activity of Hhp1 such that in the absence of analog the mutant Hhp1 adequately phosphorylates Rec8 but not Rec11 (at least not completely) (Tables 2 and 3; Figs 1, 2, and 3). This fortuitous result greatly aided our experiments because hhp1Δ hhp2Δ mutants grow poorly and have very low viable spore yields [26–28], whereas the hhp1-as mutants grow like wild type and have quite high viable spore yield in the absence of analog (Table 2). There are precedents for such differential inactivation of protein kinases by ATP-binding site mutations. For example, mutation of the “gatekeeper” residue in the ATP-binding site can reduce kinase activity even without analog present [29]. In addition, differential regulation of cellular events can arise from differential threshold levels for kinase activity [44,45].
During meiosis, Hhp plays major roles both in the timely removal of cohesin from chromosome arms at the onset of anaphase I, via phosphorylation of Rec8, and in DSB formation and recombination, via phosphorylation of Rec11. Since both of these processes are meiosis-specific, it is not surprising that we found physical interaction between Hhp and cohesin only in meiotic cells (S4 Table and S3 Fig); this association may aid the coordination of Rec8 cleavage and recombination. The abundance of hhp1 and hhp2 transcripts is greatly increased during meiosis [46], a feature consistent with Hhp playing especially important roles during meiosis. Hhp directly phosphorylates Rec8 [12,23], and it may directly phosphorylate Rec11 or activate another protein kinase that does; our data do not distinguish these possibilities, but it was recently reported that Hhp1 and Hhp2 can directly phosphorylate Rec11 [43]. Hhp clearly regulates separately two events essential for the formation of viable gametes and species propagation.
We found that the Hhp mutants (hhp2Δ coupled with hhp1-as or hhp1Δ) reduced recombination more in some intervals than in others (Tables 1, 2, and 3, and S3 Table), much like the region-dependent reductions by rec8 and rec11 mutations, including deletions [18,25]. Furthermore, hhp1-as abolished meiotic DSBs at some hotspots but not at others (Fig 1 and S1 Fig). The residual DSB patterns are reminiscent of those of cohesin and certain LinE mutants [14,18]. These observations lead us to propose that Rec11 phosphorylation is required for the loading of the putative Rec25-Rec27-Mug20 complex and Rec10 at DSB hotspots [14]. This proposal is consistent with the reduction, but not elimination, of Rec27-GFP foci in the hhp1-as hhp2Δ mutant (Fig 3 and S7 Fig). Rec11 phosphorylation seems not to be required for Rec11 to localize to the nucleus and possibly to load onto chromosomes, since Rec11-GFP nuclear localization was similar in wild type and in the hhp1-as hhp2Δ mutant (S9 Fig).
Our results, coupled with previous data [14], suggest the following scheme for formation of meiotic DSBs and crossovers (Fig 4). Rec8 is loaded onto chromosomes during S phase. Rec11 is concurrently or subsequently loaded in a Rec8-dependent manner (S10 Fig) [42] and then phosphorylated by Hhp, which allows the preferential loading of Rec25-Rec27-Mug20 at DSB hotspots. Rec10 is loaded at low levels independently of any of these proteins and, in their absence, activates the Rec12 complex to form DSBs at low level across the genome and at a few DSB hotspots. The Rec25-Rec27-Mug20 complex, whose loading at high levels at DSB hotspots depends on Rec8 and phosphorylated Rec11, along with Rec10 strongly activates the Rec12 complex to form DSBs at the hundreds of DSB hotspots dependent upon these proteins [14]. In this view, the main role of Hhp is to promote high level DSB formation at hotspots, which collectively account for about 70% of all DSBs and about half of all crossovers across the genome [47].
The proteins discussed here are widely conserved across all eukaryotic phyla, including humans. To our knowledge, casein kinase homologs are present, often as multiple paralogs, in all eukaryotes examined, and all except apparently the protist Tetrahymena thermophila have meiosis-specific Rec8 cohesin subunits [7]. In only rare cases, such as the budding yeast Saccharomyces cerevisiae and T. thermophila, is there no identified meiosis-specific homolog of the Rec11 cohesin subunit [6,7,48,49]. Vertebrates harbor three Rec11-like STAG (stromal antigen) proteins, STAG1, STAG2, and STAG3. STAG3 is meiosis-specific, is required for meiotic sister chromatid cohesion and chromosome axis formation, and is closely related to the Rec11 protein studied here [50–52]. Murine STAG3 is phosphorylated during meiosis, and this modification appears to be required for meiotic progression [53]. STAG3-deficient mice, both male and female, are sterile and display severe meiosis I defects [54,55]. A STAG3 frameshift mutation is apparently the cause of premature ovarian failure in humans [56]. Thus, our observations on Rec11 phosphorylation and its role in meiotic chromosome behavior are likely to pertain to meiosis and fertility in most species, including humans.
Strains were constructed by standard meiotic crosses [57]; genotypes of strains and the figures and tables in which each was used are in S6 Table. Mutations were introduced into cloned genes using the QuikChangeII kit (Agilent Technologies), which were inserted into the genome by transformation to antibiotic-resistance [28]. Transformants were confirmed by PCR-based analysis and, in some cases, by nucleotide sequencing.
Meiotic crosses and analysis of random spore colonies were conducted as described [57]. The ATP analog 1-NM-PP1 (Toronto Research Chemicals) was added to the sporulation agar (SPA) at 50 μM; required amino acid, purine, and pyrimidine supplements were added at 100 μg/ml. Ade+ recombinant frequencies were determined by differential plating on yeast extract agar (YEA) with and without guanine (200 μg/ml). To determine intergenic recombinant frequencies, spore suspensions were plated on YEA; colonies were transferred with toothpicks to YEA supplemented with adenine (100 μg/ml), incubated overnight, and replicated to appropriate media to determine phenotypes. Recombinant frequencies were converted to genetic distance (cM) using Haldane’s equation [x = -½ ln(1-2R), where x is the distance in Morgans (M) and R is the recombinant frequency].
To prepare large meiotic cultures, cells were grown to log phase at 25° C in supplemented liquid minimal medium (EMM2), washed in H2O, resuspended in supplemented EMM2 without a nitrogen source, and incubated for 18–19 hr, at which time NH4Cl was added to 5 mg/ml and the temperature raised to 34° C to inactivate the Pat1-114 temperature-sensitive protein kinase [58]. Cells were harvested at appropriate times after induction and analyzed for DNA content by flow cytometry (to determine the time of replication) and for other features as described below.
To determine DSBs, meiotic cells were washed, concentrated by centrifugation, and embedded in agarose plugs, which were sequentially treated with lytic enzymes, proteinase K, RNase, and appropriate restriction enzymes [58]. The digested DNA was separated by gel electrophoresis and analyzed by Southern blot hybridization.
To analyze phosphorylation and electrophoretic mobility of Rec11, cells expressing Rec11-TAP were arrested in G1 by nitrogen starvation, and meiosis was induced by shifting the culture to 34°C. Four hr later, cells from 20 mL of culture were concentrated by centrifugation, suspended in IPP150 buffer (50 mM Tris-HCl pH 8.0, 150 mM NaCl, 10% glycerol, 0.1% NP-40), and homogenized using glass beads (0.4–0.6 mm diameter). Extracted proteins were immunoprecipitated with IgG Sepharose 6 Fast Flow beads (GE Healthcare), treated with alkaline phosphatase (Thermo Scientific) as indicated and separated by electrophoresis through 5% polyacrylamide gels containing SDS (0.1%). Proteins were transferred to a PVDF membrane (Millipore), and Rec11-TAP and Hhp2-TAP were detected using rabbit antiperoxidase antibody linked to peroxidase (Dako; 1:30,000 dilution) in 0.1% PBS-T (8 gm NaCl, 0.2 gm KCl, 1.44 gm Na2HPO4, 0.24 gm KH2PO4, 1 mL Tween-20 per L). Hhp1-PK9 was detected using mouse-anti-PK (V5) antibody (Serotec; 1:2000 dilution) and goat anti-mouse IgG-HRP secondary antibody (Santa Cruz Biotechnology; 1:5000 dilution) in 0.1% PBS-T.
Cells from six-liter (mitotic) or fifteen-liter (meiotic) cultures of strains expressing TAP-tagged proteins were collected by centrifugation; meiotic cultures were harvested at 2.5–3.5 hr after induction of meiosis. Yeast cell powder was made from frozen pellets using a SamplePrep 6870 Freezer Mill (SPEX, Inc.). Proteins were extracted using IPP150 buffer containing complete protease and phosphatase inhibitors (Roche) and 1 mM PMSF (Sigma). All washing steps were performed in Poly-Prep columns (Bio-Rad) by gravity flow. IgG Sepharose™ 6 Fast Flow beads (500 μl; GE Healthcare) were washed with IPP150 buffer, mixed with protein extract, and rotated for 2 hr at 4°C. Beads were washed with IPP150 buffer and then with TEV cleavage buffer (TCB: 10 mM Tris-HCl pH 8.0, 150 mM NaCl, 10% glycerol, 0.1% NP-40, 0.5 mM EDTA, 1 mM DTT). Protein cleavage was performed in 2 ml of TCB buffer supplemented with 400 units of AcTEV protease (Life Technologies) for 2 hr at 16°C. The eluate (2 ml) was supplemented with 6 μl of 1 M CaCl2 and mixed with 6 ml of calmodulin binding buffer 1 (CBB1: 10 mM Tris-HCl pH 8.0, 150 mM NaCl, 10% glycerol, 0.1% NP-40, 1 mM imidazole, 1 mM Mg(OAc)2, 2 mM CaCl2, 10 mM β-mercaptoethanol). Calmodulin Sepharose 4B beads (150 μl; GE Healthcare) were washed with CBB1 buffer, added to a mixture of eluate and CBB1 buffer, and incubated for 2 hr at 4°C. The beads were washed with CBB1 and calmodulin binding buffer 2 (CBB2: 10 mM Tris-HCl pH 8.0, 150 mM NaCl, 1 mM Mg(OAc)2, 2 mM CaCl2, 1 mM β-mercaptoethanol). The proteins were step-eluted using one bed volume of elution buffer (EB: 10 mM Tris-HCl pH 8.0, 150 mM NaCl, 1 mM Mg(OAc)2, 2 mM EGTA, 1 mM β-mercaptoethanol). Eluted proteins were separated by SDS-PAGE and silver stained. Eluates from peak fractions were analyzed by LC-MS/MS as described in S1 Text and by Cipak et al. [59].
Haploid pat1-114 cells were arrested by nitrogen starvation for 16 hr and released into meiosis at 34°C by inactivation of Pat1 and addition of nitrogen. Cells were harvested at the indicated time points after meiotic induction, fixed with 70% ethanol (Fig 3 and S7 Fig) or with 99.8% methanol (S9 Fig and S10 Fig), and stained with DAPI; nuclei were counted in 100 cells at each time point. The Rec10-GFP protein was visualized in unfixed cells (S7 Fig) as follows. Cells from 100 μl of culture were collected by centrifugation, washed once with water, and spread on a cover slip coated with poly-L-lysine. Slides with a drop of mounting medium containing DAPI were covered with the cover slips, and the cells were analyzed the next day. All images were from a single focal plane; out-of-focus cells were not scored. Images for Fig 3 and S7 Fig were obtained with a Zeiss Axio Imager.Z2 microscope equipped with a Plan Apochromat 63x/1.4 oil-immersion lens and an AxioCamMRm camera; images were analysed with ZEN 2011 software. Images for S9 Fig and S10 Fig were obtained with a Zeiss Apotome microscope equipped with a 100X /1.4 oil-immersion lens and were analyzed using Axiovision software. All inductions and localization analyses were performed at least twice. Haploid strains were analysed because diploid hhp1-as hhp2∆ cells grow poorly even without analog and produce grossly abnormal asci.
|
10.1371/journal.ppat.1004091 | Structural Differences Explain Diverse Functions of Plasmodium Actins | Actins are highly conserved proteins and key players in central processes in all eukaryotic cells. The two actins of the malaria parasite are among the most divergent eukaryotic actins and also differ from each other more than isoforms in any other species. Microfilaments have not been directly observed in Plasmodium and are presumed to be short and highly dynamic. We show that actin I cannot complement actin II in male gametogenesis, suggesting critical structural differences. Cryo-EM reveals that Plasmodium actin I has a unique filament structure, whereas actin II filaments resemble canonical F-actin. Both Plasmodium actins hydrolyze ATP more efficiently than α-actin, and unlike any other actin, both parasite actins rapidly form short oligomers induced by ADP. Crystal structures of both isoforms pinpoint several structural changes in the monomers causing the unique polymerization properties. Inserting the canonical D-loop to Plasmodium actin I leads to the formation of long filaments in vitro. In vivo, this chimera restores gametogenesis in parasites lacking actin II, suggesting that stable filaments are required for exflagellation. Together, these data underline the divergence of eukaryotic actins and demonstrate how structural differences in the monomers translate into filaments with different properties, implying that even eukaryotic actins have faced different evolutionary pressures and followed different paths for developing their polymerization properties.
| Malaria parasites have two actin isoforms, which are among the most divergent within the actin family that comprises highly conserved proteins, essential in all eukaryotic cells. In Plasmodium, actin is indispensable for motility and, thus, the infectivity of the deadly parasite. Yet, actin filaments have not been observed in vivo in these pathogens. Here, we show that the two Plasmodium actins differ from each other in both monomeric and filamentous form and that actin I cannot replace actin II during male gametogenesis. Whereas the major isoform actin I cannot form stable filaments alone, the mosquito-stage-specific actin II readily forms long filaments that have dimensions similar to canonical actins. A chimeric actin I mutant that forms long filaments in vitro also rescues gametogenesis in parasites lacking actin II. Both Plasmodium actins rapidly hydrolyze ATP and form short oligomers in the presence of ADP, which is a fundamental difference to all other actins characterized to date. Structural and functional differences in the two Plasmodium actin isoforms compared both to each other and to canonical actins reveal how the polymerization properties of eukaryotic actins have evolved along different avenues.
| Actins are the most abundant and among the most conserved proteins in eukaryotic cells and play indispensable roles in a plethora of key cellular events, including muscle contraction, cell division, shape determination, transport, and cell motility [1], [2]. Actins are highly conserved in opisthokonts with <10% divergence between yeast and man. The six mammalian actin isoforms differ from each other by a maximum of 6% of the sequence, and are virtually identical across species. Nevertheless, these subtle differences are enough to determine isoform-specific functions [3]. Common to most actins is their capacity to form long filaments. However, in a number of phylogenetically distinct organisms, such as Trypanosoma and Plasmodium spp., actin filaments have not been observed [4], [5]. Unlike other members of the phylum Apicomplexa, which comprises single-celled eukaryotic intracellular parasites, the malaria parasites have two actin isoforms, which at the sequence level are <80% identical with canonical (opisthokont) actins and each other. This is a remarkable difference, considering the near identity among canonical actins (Fig. S1). An important question is how this divergence at the amino-acid level translates into different structures – and how this, in turn, influences polymerization.
Most studies on apicomplexan actins have concentrated on their role in gliding motility, a unique mode of migration, essential for the parasite to infect new cells. However, like in other eukaryotes, parasite actins must have several cellular functions. Actin polymerization is indispensable for gliding and likely involved in host cell invasion and egress [6]–[8]. Despite evidence for this crucial role of filamentous actin, long filaments have only been visualized in Theileria [9], which appears not to use actin filaments for host cell invasion [10]. The presence of regular actin filaments in Plasmodium is uncertain [11]–[13]. In vitro, apicomplexan actins form short, ∼100-nm long filaments, which undergo rapid treadmilling [14]–[17]. Recently, specific antibodies revealed filament-like structures in motile forms of Plasmodium [12], [13]. Toxoplasma gondii actin, which is 93% identical to Plasmodium actin I, has been reported to polymerize at concentrations 10-fold lower than canonical actins [15], and most recently, it has been proposed to polymerize in an isodesmic manner without a lag phase or a critical concentration, which is unique among all actins or actin homologs studied to date [18]. Yet, most of the cellular actin is present as monomers [19], implying that filaments occur only transiently, and polymerization is under tight control of regulatory proteins or governed by distinct properties of the monomer. On the other hand, it has been estimated that 2/3 of Plasmodium actin in merozoites – the infective blood stage form, which does not exhibit gliding motility – could be present as short filaments [20].
Plasmodium actin I is abundant and expressed throughout the life cycle of the parasite, whereas actin II is present only in the gametocytes and mosquito stages [21]–[24], including sporozoites [25], the highly motile form of the parasite, transmitted to the vertebrate by the mosquito. Actin I is an indispensable part of the parasite motor machinery responsible for the unique gliding motility of the parasite. Actin II has at least two functions in the mosquito stages, as revealed by reverse genetics analyses. It is required both during male gametogenesis and in the zygote stage [23], [24]. However, no clear molecular function has been assigned for actin II.
To understand the properties of the divergent Plasmodium actins, we have determined their monomer crystal structures and analyzed their filament assembly using electron microscopy (EM). We show that, unlike in any other cell reported so far, the two isoforms differ substantially from each other in their ability to form filaments and that both oligomerize in the presence of ADP. Their functional uniqueness is further highlighted by the finding that Plasmodium actin II has a distinct role in male gametogenesis that cannot be complemented by actin I. Finally, we show that a chimera of Plasmodium actin I and canonical actin can form long filaments and, importantly, restores the function of actin II in gametogenesis.
The most peculiar property of apicomplexan actins is their apparent inability to form long, stable filaments. This is a fundamental difference to all actins studied so far, and in the lack of structural information, the reasons for the poor polymerizability are not understood. It has been shown by atomic force microscopy that the dimensions of jasplakinolide (JAS)-stabilized P. falciparum actin I filaments, purified from merozoites, are different with respect to their helical symmetry from canonical actins [26]. However, the structure of actin II filaments has not been studied before. We visualized the structures formed by both Plasmodium actin isoforms using EM (Fig. 1). When polymerized in the presence of ATP in high-salt conditions at room temperature overnight, actin I forms only short, irregular structures of approximately 100 nm in length (Fig. 1 A and F), while actin II forms significantly longer filaments (average length 650 nm) (Fig. 1 B,C,F). In the presence of JAS, which in vitro stabilizes actin filaments, both actin I and II form long, rather straight filaments (Fig. 1 D–F).
To evaluate whether the helical assemblies of the two Plasmodium actins are different, we used cryo-EM. Filaments of both parasite actins were embedded in vitreous ice (Fig. 2). First, we inspected the averaged power spectra derived from segments of 330 and 56 filaments for actin I and II, respectively. These look virtually identical because they can only be compared at 1/60 Å−1 resolution (Fig. 2 A). We then characterized the structure of the filaments in real space and performed k-means classification of helical segments [27]. Inspection of the classes allows a direct measurement of the cross overs or half-pitch of the two-start helix, which represent the distance the filament requires to undergo a 180° rotation. For actin I and II, cross-over distances cluster around 406±16 Å and 364±10 Å, respectively (Fig. 2 B), which is confirmed from Eigen images that represent the half-pitch of the two-start helix (Fig. 2 C). Hence, actin II has symmetry parameters identical to α-actin, for which a cross-over distance of 371 Å has been reported [28], while actin I possesses a significantly longer cross-over distance, which is in agreement with the earlier work performed on actin I using atomic force microscopy [26].
To understand whether the longer cross-over distance is a result of a change in the helical rise and/or the helical rotation, we determined the low-resolution 3D structure of actin I (Fig. 3 A) using single-particle based helical reconstruction [29], [30]. We refined the helical symmetry (Fig. 3 B) and determined the low-resolution filament structure at 25-Å resolution (at FSC 0.5 cutoff, Fig. 3 C). This showed that the cross-over distance change is mainly due to a change of helical rotation from −166.6° [28] to −167.5°, which corresponds to the predicted rotation change if the helical rise remains constant at 27.7 Å. Despite the obvious difference in helical symmetry, at the current resolution, the cryo-EM structure of the actin I subunit looks very similar to the canonical actin filament (Fig. 3 A). This is the first time that such large differences in the properties of filaments have been observed for actin isoforms of any species. Thus, higher resolution is needed to further characterize the molecular interactions giving rise to the different polymerization propensities and the observed helical rotation changes.
The divergent polymerization properties of Plasmodium actins in vivo may be partly accounted for by differences in the activities of the actin-binding proteins, which in these parasites are also poorly conserved and have partially divergent functions compared to canonical counterparts [31]–[38]. However, differences in the actin monomer structure must be responsible for the observed differences in filament structure in vitro and also for interactions with regulatory proteins and, thus, functional differences in vivo. Therefore, high-resolution structures are required for understanding the biological and molecular differences of the parasite actins compared to their canonical homologs. Such information will also aid us in evaluating the suitability of Plasmodium actins as drug targets.
We set out to determine the crystal structures of P. falciparum actin I and P. berghei actin II. The sequence identities between the counterparts from P. falciparum and P. berghei are 99% for actin I and 92% for actin II. The gelsolin G1 domain was used to stabilize the monomers and facilitate crystallization of both actins [39], [40]. The actin I structure was refined to a resolution of 1.3 Å and actin II to 2.2 Å. These high-resolution structures allow for a very detailed comparison of the Plasmodium actins with each other and with other actins (Figs. 4, S2, and S3). Although Plasmodium lacks a gelsolin homolog, the mammalian G1 is bound between subdomains 1 and 3 in both Plasmodium actins, similarly to other actin–G1 complexes [39], [41] (Fig. S2 D and E). For all comparisons, we have used canonical actin structures that have also been determined in complex with G1, in order to rule out structural rearrangements caused by gelsolin binding.
In most canonical actin structures, the C terminus is folded as an α-helix, which interacts with the bottom part of subdomain 1. This is true also for Plasmodium actin II (Fig. 4 A). In actin I, however, the C terminus turns towards solvent and is disordered (Fig. 4 B). The large hydrophobic cleft between subdomains 1 and 3 is defined as a ‘hotspot’ for regulatory protein binding [42] (Fig. 4 A and B). A smaller hydrophobic patch in the direct vicinity of the C terminus is, in addition, involved in binding at least profilin [43]. In actin I, the large hydrophobic residues in this smaller cleft, including Trp357, have adopted different conformations compared to canonical actins (Fig. 4 B), possibly influencing the binding of profilin, which we have previously proposed to bind actin in a different manner in Plasmodium compared to opisthokonts [31]. In actin II, these hydrophobic residues, like the C terminus, are in the canonical conformations. All in all, despite the obvious similarity to other actins, especially actin I shows appreciable structural deviations, in particular in regions involved in binding of regulatory proteins [42]. This may provide possibilities for structure-based drug design targeted at the Plasmodium actin–regulatory protein interfaces.
A key question concerning the properties of the major Plasmodium actin isoform, as well as other apicomplexan actins, is why they, unlike all the extensively studied actins, form only short, unstable filaments. Comparing the crystal structures of the monomers to the recently determined high-resolution cryo-EM structures of canonical actin filaments [28], [44]–[46] can provide clues to answer this question. In canonical F-actin, the axial interactions, meaning the longitudinal contacts between the actin monomers in each of the two protofilaments, are tight and mainly electrostatic, as revealed by cryo-EM studies [28], [45]. The most important axial interactions are discussed below.
The DNase I binding (D-)loop in subdomain 2 (residues 39–61 in actin I) is one of the most important regions for polymerization and, in the filament, inserts into the hydrophobic cleft between subdomains 1 and 3 of the neighboring monomer [28], [45], [47], [48]. In both Plasmodium actin crystal structures, the D-loop is disordered, and the tip of it is not visible in the electron density maps (Figs. 4 and S3). Interestingly, the most notable changes in the D-loop sequence concern the first and the last residues of a segment (residues 42–49) that forms a short α-helix in some G-actin structures with ADP bound [49] and has been modeled as a helix also in one of the recent F-actin structures [45]. This part of the D-loop inserts deep into the neighboring monomer in the filament, contacting the so-called proline-rich loop (residues 109–115). Residue 42, which is a glutamine or threonine in most other actins, is a proline in both Plasmodium actins. A proline restricts the conformation of the chain and, although also known as a ‘helix breaker’, is often also seen as the first residue in α-helices. At the C terminus of this segment, residue 49, which is a glycine in canonical actins, is a glutamate in both Plasmodium actins. These replacements at both ends of this segment would be expected to increase the helical propensity of the tip of the D-loop.
In canonical F-actin, Thr325 in the loop between α-helix 11 and β-strand 18 (residues 323–327; Fig. 4 A and B) in subdomain 3 of one monomer interacts with Glu242 in the loop connecting β-strands 15 and 16 in subdomain 4 (residues 242–247; Fig. 4 A and B) of the neighboring monomer [28]. In actin I, this loop and the threonine side chain have turned away from the optimal position for this interaction (Figs. 4 B and S3). In actin II, Ser325 is positioned such that hydrogen bonding to the apposing glutamate can be easily achieved. The difference in the conformation can be explained by the substitutions Y280F and M284K in actin I compared to canonical actins and actin II.
The third main site contributing to axial interactions is formed by the loop connecting helices 9 and 10 (residues 284–290 in subdomain 3; Fig. 4 A) inserting between subdomains 2 and 4 of the neighboring monomer. This area is almost fully conserved in both parasite actins, the only substitution being that of Met283 by a lysine (284) in actin I.
All in all, the largest differences in the axial contacts concentrate to the D-loop and concern equally both Plasmodium actins. Smaller differences in subdomain 3 may, however, partly explain the different polymerization propensities of the two parasite actin isoforms.
The lateral contacts between the two protofilaments concern mainly interactions between subdomains 1 and 4 and the beginning of the D-loop in subdomain 2 interacting with the so-called hydrophobic loop (residues 263–275) (Figs. 4 A–B, S1, and S2 D–E) between subdomains 3 and 4 of the apposing monomer [28], [45], [48]. In a recent high-resolution cryo-EM filament structure [45], Arg206 in subdomain 4 interacts with Ser271 of a neighboring monomer. In actin I, Ser271 is replaced by Ala272, and Lys207 and Glu188 are involved in a short hydrogen bond with salt-bridge character (Fig. 4 C). A similar interaction occurs between Arg206 and Asp187 in Latrunculin A-bound α-actin [50]. In the absence of latrunculin, the distance between these residues in α-actin is longer and the geometry suboptimal for a salt bridge (Fig. 4 C). Latrunculins prevent actin polymerization, presumably by limiting the flexibility of subdomains 2 and 4 [50],[51], and the salt bridge seen in the latrunculin-actin structure, similar to Plasmodium actin I, may be one reason for this. The longer side chain of Glu188 in actin I compared to Asp187 in canonical actins may facilitate the interaction with Lys207. Actin II has Tyr187 in place of Asp187/Glu188, which affects the orientation of Arg206. Actin II, furthermore, has Cys272 in the place of Ser271 of canonical actins, allowing for hydrogen bonding upon polymerization.
At the N terminus of the D-loop, Arg40 and His41 of canonical actins are replaced by lysine and asparagine in actin I and lysine and methionine in actin II. In the filament, Arg40 forms a salt bridge with Glu271 [28], and the replacement to lysine can subtly weaken this interaction. His41, in turn, interacts with Ser266 [28]. Although the asparagine in actin I can also contribute to a hydrogen bond, the interaction may be weaker due to the shorter side chain. In F-actin, Lys114 in the proline-rich loop (residues 109–115 in subdomain 1, following actin I numbering) forms a salt bridge with Glu196 of a neighboring monomer [28]. Uniquely, the neighboring residue is Gly115 in actin I (Fig. 4 D), which can have a large effect on the mobility of the proline-rich loop. Actin II has a non-glycine residue (threonine) at this position, similarly to canonical actins (alanine or serine).
Taken together, especially residues involved in lateral contacts, crucial for the stability of the filaments, are altered in Plasmodium actin I but more conserved in actin II. The most important differences concern subdomain 4, the D-loop, and the proline-rich loop. The conformations of all these affect each other allosterically and can also modulate the ATPase activity [52]–[54].
Because of the observed differences in sequence, conformation, and helical symmetry between Plasmodium and canonical actins and the importance of the D-loop-mediated contacts for polymerization, we constructed a chimera, in which the entire D-loop of actin I was replaced with that of α-actin. Strikingly, this chimera in the absence of any stabilizing agents forms filaments with an average length of 1.6 µm, which is longer than either actin I or II filaments (Figs. 5 A–B and 1 F). However, the appearance of the chimera filaments is not as regular as the filaments formed by canonical actins or actin II.
To further characterize the filaments formed by the actin I–α-actin chimera, we analyzed electron cryo-micrographs of JAS-stabilized chimera filaments and subjected them to the same classification and symmetry analysis as outlined above (Fig. S4). Based on the determined half-pitch distance and the symmetry analysis, we conclude that the symmetry parameters of the chimera filaments are in close agreement with the wild-type actin I filaments. Therefore, the symmetry parameters appear to be retained from the wild-type actin I, and thus, the D-loop of canonical actin merely confers increased filament stability.
The crystal structure reveals that the 3D structure of the chimera in monomeric form is very similar to that of wild-type actin I (Figs. 5 C and S3). However, there are some notable differences. The D-loop in the chimera is slightly more ordered. In addition, the C-terminal helix is visible and can be superimposed with the C terminus of actin II and most canonical actins. The Trp357 side chain in the hydrophobic patch is not flipped as in wild-type actin I (Figs. 4 B and 5 C). This observation is in line with earlier reports on allostery between the conformations of the D-loop and the C terminus [55], [56] and demonstrates the ability of even large residues in the hydrophobic core of actin to move, which is a requirement for the rearrangements taking place in the monomer upon polymerization.
In addition to residues already discussed, a notable difference between the D-loop of Plasmodium actin I and canonical actins is at position 54, which has a tyrosine in opisthokont actins and Plasmodium actin II but phenylalanine in actin I. In the chimera, the OH group of Tyr54 is hydrogen bonded to the main chain of Lys51 and could also interact with the tip of Lys62 (Fig. 5 C). In wild-type actin I, Phe54 has moved away from Lys62 and seems to push Lys51 to a slightly different conformation. In actin II, Tyr53 (corresponding to Phe54 in actin I) is able to interact with Lys61 (Lys62 in actin I), but otherwise, this region is rather different in conformation compared to both actin I and canonical actins. This is due to conformational changes in subdomain 4, involving Thr203, Arg206, and seemingly originating from the bulky side chain of Tyr187, which is replaced by aspartate in canonical actins and glutamate in actin I, as discussed above. According to a recent filament structure [45], Tyr54 becomes stacked between the side chains of Asp52 and Lys62 of the same monomer and its only possible rotamers would allow an OH–π interaction with the phenyl ring of Tyr170 of the neighboring monomer. The side chain of Phe54 in actin I is not able to participate in this kind of an interaction due to the missing hydroxyl group. Thus, the replacement of tyrosine by phenylalanine at this position may affect both the rigidity of the D-loop in the monomeric state as well as the stability of the filament.
The state of the nucleotide-binding pocket can be described using two parameters: (i) the “phosphate clamp”, which is the distance between the Cα atoms of Gly15 and Asp157, and (ii) the “mouth”, which is the distance between the Cα atoms of Gln59 and Glu207 [46]. The phosphate clamp distance in Plasmodium actins does not differ significantly from other actins. However, the mouth of the binding pocket is significantly more closed in both parasite actins: 9.85 Å in actin I and 9.04 Å in actin II, compared to an average of 10.87±0.4 Å in 9 other actin–G1 structures used for comparison. In three recent high-resolution F-actin structures [28], [45], [48], the mouth distance varies between 7.88 and 9.66 Å. Both parasite actins reach the more closed conformation by slightly different means. In actin I, the largest differences in conformation to canonical actins are in subdomain 4 and in actin II, subdomain 2 (Figs. 4 and S3).
Both structures reported here have calcium and ATP bound to the nucleotide-binding cleft between subdomains 2 and 4 (Figs. 4 and 6). In the nucleotide-binding residues, there is one interesting difference in actin I compared to canonical actins and actin II; residue 17, which is hydrophobic (methionine/leucine) in all canonical actins and actin II, is an asparagine in actin I (Fig. 6 A and B) and also T. gondii actin. This side chain is close to the α- and β-phosphates of ATP. In actin I, the distance of the Asn17 Nδ2 atom to the β-phosphate O1 atom (∼3.75 Å) is too long for hydrogen bonding in this conformation. However, together with its own main chain N, that of Gly16, and Nζ of Lys19, the Asn17 side chain could form an oxyanion hole to stabilize a negative charge on the β-phosphate O1 atom (Fig. 6 A). In addition, it could interact with the α-phosphate. The Asn17 side chain is flexible in the crystal structure, as evident from the electron density maps and B factors, and based on the shape of the electron density as well as anisotropic ellipsoids, seems to move in concert with active-site water molecules as well as the nearby Tyr338, which is in a double conformation.
A catalytic mechanism based on a nucleophilic attack of a water molecule activated by His162 and Gln138 has been proposed for ATP hydrolysis in actin [57]. The reaction itself is rather simple, containing only proton transfer steps. The complication arises from the fact that G-actin – currently the only form of which atomic-resolution information can be achieved – is a poor catalyst, and conformational changes upon polymerization are needed for achieving the catalytically competent conformation of the active site. Water 39 (actin I numbering) has been proposed to be the nucleophile initiating the reaction, and depending on the bound metal, the position of this water changes [57]. In actin I, this water is 4.75 Å away from the γ-phosphate, and the angle between the β-γ bridging O, Pγ, and water 39 is 152.4°, which is amenable for a nucleophilic attack (Fig. 6 B). The catalytic site water structure in actin II is different compared to actin I (Fig. 6 C), as the presumed catalytic water (147 in the actin II structure) has moved towards His161 and Pro109 and has weak electron density and a high B factor. The distance of this water to Pγ is 5.43 Å, and it is in an almost in-line position (162.4°). In fact, according to the electron density and the distances to neighboring atoms, this water may be a second conformation of water 357, which is directly hydrogen bonded to His161. Together, these differences in the active site architectures between the two parasite actins and compared to other actins indicate that the catalytic activity and the exact mechanism of ATP hydrolysis may differ in them, likely resulting also in differences in polymerization.
Given the structural differences in the catalytic sites, we set out to test if the Plasmodium actins differ from each other and canonical actins in their ability to hydrolyze ATP and/or release phosphate. We first measured phosphate release in the presence of Mg2+ (Fig. 6 D). It should be noted that, as phosphate release at least in canonical actins is much slower than hydrolysis, this method gives only indirect information about the hydrolysis rate. In these conditions, both parasite actins release phosphate faster than α-actin. Of the two parasite actins, actin I has a slightly higher rate (∼2 times higher than actin II). Curiously, the chimera with the α-actin D-loop is approximately 3-fold more active than wild-type actin I and 23-fold more active than α-actin. We also made two point mutations to actin I; G115A, which we predicted to affect the flexibility of the proline-rich loop and, thus, the rate of hydrolysis, and F54Y, which we hypothesized might affect the rigidity of the D-loop in monomeric state. Neither of these mutants formed long filaments without JAS (Fig. 1 F). In the presence of Mg2+, the F54Y mutant shows identical behavior to wild-type actin I. G115A, however, has a reduced rate, similar to actin II. Also the kinetics differ from each other in the different actins. Whereas actin I and the point mutants release phosphate in a linear way, α-actin, actin II, and the chimera display more complex kinetics, having an initial very short, faster, non-linear phase, followed by a linear phase.
We next measured phosphate release in the Ca2+-bound, presumably mainly monomeric, forms (Fig. 6 D). As expected, α-actin showed an even lower release of phosphate in the Ca2+-bound compared to the Mg2+-bound form. Both actin I and actin II have activities equal compared to each other and approximately 5-fold higher than α-actin. The chimera releases phosphate 3-fold less in the presence of Ca2+ than with Mg2+, but the rate is still significantly higher than that of muscle actin or both wild-type parasite actins with Ca2+. Of the point mutants, F54Y has practically no activity with Ca2+ (identical to α-actin), whereas G115A is slightly more efficient in the presence of Ca2+ than Mg2+. Altogether, these data show that the Plasmodium actins have a different mechanism of ATP hydrolysis and/or subsequent phosphate release compared to canonical actins, which are poor catalysts in the monomeric form and adopt the catalytic conformation only upon polymerization, which is a prerequisite for non-equilibrium polymerization kinetics enabling directional growth [58]. We were also able to pinpoint amino acid residues responsible for these differences.
In order to evaluate the oligomeric state of the parasite actins in the presence of ATP/ADP and different ions, we used native PAGE. With ATP bound, actin I spontaneously forms short polymers (from tetramers up to 11–12-mers) in ∼48 h when stored on ice (Figs. 7 and S5). Actin II stays mainly monomeric over the same period of time, although minute amounts of oligomers (dimers–octamers) appear. Interestingly, in the presence of ADP, oligomerization starts instantly, and the majority of both actins is oligomeric (dimers–10-mers) immediately after a 1-h hexokinase treatment at 298 K to remove ATP. After incubation of the actins at 298 K for 1 h without hexokinase, only minute amounts of oligomers can be visualized for actin I, and no visible oligomerization of actin II takes place (data not shown). After 48 h, the proportion of larger oligomers of the ADP forms is much higher, and monomers as well as lower oligomers are practically non-existent. The formation of oligomers is not caused by oxidation, as using even a large excess of the reducing agent TCEP in the sample does not reduce the amount of oligomerization (Fig. 7 A). Both ATP and ADP forms of α-actin remain monomeric in the same conditions. However, short oligomers of ATP-α-actin have been reported below the critical concentration for polymerization [59].
Because the distribution, when separated on a gel, does not necessarily reflect the equilibrium between different species in solution, we also used dynamic light scattering (DLS) to visualize the size distribution and polydispersity of the actin mono- and oligomers in solution over time (Fig. S6). The resolution of DLS is far from that of the native gel assay, and it is only possible to detect size differences of approximately 5–6 fold. Therefore, e.g. monomers, dimers, and trimers will appear as a single, polydisperse peak. 6 h after purification, actin I is seen in mainly two separate peaks of average hydrodynamic radii of approximately 2 and 6 nm (Fig. S6 A). 2 nm would be very close to the expected hydrodynamic radius of the monomer. After 11 h, nearly all of actin I is in particles with a radius of ∼5 nm (Fig. S6 B). As time goes by, the distribution becomes divided between particles of below 3 nm (close to a monomer) and larger oligomers with an average radius of 11–12 nm (Fig. S6 C and D). The polydispersity of the sample after 11 h is very high, indicating that the sample contains a mixture of monomers and small oligomers, and the polydispersity diminishes again, as the sample gains a multimodal distribution, indicating that the smallest oligomers disappear over time, leaving behind a pool of monomers in addition to the higher oligomers, consistent with our native PAGE data. As seen also in the native gels, actin II retains a higher fraction of monomers over 48 h, but also gains a fraction of significantly higher oligomers, which are, however, infrequent and very heterogeneous in size (Fig. S6 E–H).
In order to probe the effects of Mg2+ and K+ ions on the oligomerization behavior, we also performed native PAGE in the presence of two concentrations (1 and 5 mM) of MgCl2 as well as 5 mM MgCl2 and 50 mM KCl (Fig. S7). In the presence of ATP, Mg2+ slightly reduces the amount of the short oligomers for both actin I and II compared to the Ca2+ forms (Fig. S7 A and B). However, some actin I is visible at the bottom of the well at the top of the gel, which would imply filaments too long to enter the gel. This could not be seen in the Ca2+ gels for either the Plasmodium proteins or α-actin, but was much more pronounced for α-actin with Mg2+. In the presence of ADP, there is a clear shift towards longer oligomers in actin I, and after 48 h, part of actin I stays in the well, not entering the gel in the presence of 5 mM MgCl2 both with and without KCl (Fig. S7 C and D). Thus, Mg2+ alone seems to be sufficient for polymerization.
We used genetically modified parasites to address the question whether the observed structural differences translate into different properties of the proteins in vivo. While it is not possible to delete actin I due to its essential functions, a knock-out of the actin2 gene has been done, resulting in a block of male gametogenesis [23], [60]. We reasoned that a replacement of actin2 with actin1 would display the mutant phenotype if the two actin isoforms have different biological functions, while restoration of gametocyte development would indicate a similar function.
Male gametogenesis in the malaria parasite is a unique event, involving the formation of flagellar gametes. This event, called exflagellation (Fig. 8 A and Video S1), is easily scored under the microscope, allowing us to use it as a quantitative method. In our approach, actin1 was expressed under the control of the actin2 flanking regions. We used a recipient line, in which the complete open reading frame (ORF) of actin2 had been deleted. Therefore, these parasites do not exflagellate [60]. This line was separately transfected with two constructs, both aiming at integration into the actin2 locus. The complementation construct (act2com) restored the actin2 ORF, which allowed expression of the cognate gene comparable to wild type. In the replacement construct (act2rep), a fragment corresponding to the actin1 ORF was used instead of the actin2 ORF. The constructs were otherwise identical and were integrated in the locus via a single crossover homologous recombination event in the 5′ flanking region of actin2 (Fig. 8 B). In both cases, clonal lines were obtained. We compared the act2com and act2rep parasite lines with wild-type parasites in the exflagellation assay (Fig. 8 C). In the wild-type and act2com parasites, the number of exflagellation events was similar, indicating that complementation with actin2 restored the function of the gene. However, while some normal exflagellation events were detected also in the act2rep parasites, the numbers were significantly reduced compared to the act2com parasites (Fig. 8 C), strongly suggesting that actin II has unique functions, which actin I cannot fulfill during male gametogenesis.
As it became apparent that actin I polymerization properties in vitro could be altered by exchanging its D-loop to that of α-actin, we decided to investigate if this modification would also have an impact on the in vivo function of actin I. We produced transgenic parasites using the same strategy as described above; inserting the actin I–α-actin D-loop chimera into the actin2 locus, producing the act1chi parasites (Fig. 8 B). Surprisingly, this revealed that the exchange of the D-loop had a remarkable impact on exflagellation (Fig. 8 C). In the act1chi parasites, exflagellation was significantly increased compared to the act2rep strain and restored to values close to the act2com strain. These data show that the D-loop has a critical role in the function of these actins, but actin II, as shown by the structural data, has acquired other properties that contribute to its higher filament stability. Furthermore, and to our surprise, it seems that the molecular function of actin II may be dependent on the ability of the protein to form filaments.
An actin cytoskeleton was long thought to be a feature unique to eukaryotic cells, and this view was revisited only two decades ago upon the discovery of the first bacterial actin and tubulin homologs [61]–[64]. The ancient phylum Apicomplexa is likely separated from opisthokonts by an evolutionary distance of a billion years, and the diversion of Plasmodium spp. took place hundreds of millions of years ago [65]. Therefore, looking at the divergent properties of Plasmodium actins provides us with insight into the early stages of actin evolution. The ability of actin to polymerize must have evolved very early – before its involvement as tracks for molecular motors [66]. This may explain some features of both the polymerization propensity and the divergent actin-myosin motor in Apicomplexa. Also the minimal set of actin-binding proteins in Apicomplexa suggests that a common ancestor had a limited polymerization propensity, and the various regulatory proteins in higher eukaryotes have evolved as the polymerization properties of actin itself have been fine-tuned, creating a need for additional regulation. For the biological functions of actin in Apicomplexa, the development of similar polymerization properties has strikingly not been of importance.
The two Plasmodium actins differ in their polymerization propensities, filament stability, and filament helical symmetry – the hallmark of canonical F-actin. The second, stage-specific actin isoform of Plasmodium that forms long filaments with canonical F-actin symmetry is unique among Apicomplexa. Curiously, at the sequence level, actin II is as divergent from Plasmodium actin I as it is from all other actins. It has been suggested [23], and our structural data support the view, that the actin2 gene has arisen only after the diversion of Plasmodium from other Apicomplexa, and the protein seems to have gained a higher filament stability independent of the evolution of higher eukaryotic actins.
In canonical actins, polymerization is tightly coupled to ATP hydrolysis, such that structural rearrangements upon polymerization enable the active site to adopt a conformation optimal for catalysis [67]. Two key factors have been described necessary for achieving the catalytically competent conformation upon the transition from monomeric to filamentous state. These are: (i) a rotation of the outer domain (subdomains 1 and 2), resulting in flattening of the monomer and (ii) bending down of the proline-rich loop in subdomain 1 [45], [48]. The Plasmodium actins hydrolyze ATP also in the monomeric form, releasing phosphate more efficiently than canonical actins, and oligomerize readily in the presence of ADP, which is a fundamental difference to all other actins characterized, and must be a result of different atomic structures.
Interestingly, Plasmodium actin I has a unique glycine at the end of the proline-rich loop. This allows more flexibility for this loop, which apparently increases the catalytic rate in the presence of magnesium but, surprisingly, has an opposite effect in the calcium-bound form (Fig. 6 D). Also the more closed conformations of subdomains 2 and 4 in the parasite actins may facilitate ATP hydrolysis but also reduce the conformational change – or flattening – required upon insertion of the monomer into the filament. An interesting difference that may also contribute to catalysis is Asn17 close to the α- and β-phosphates of ATP in the active site. Intriguingly, the bacterial actin homolog MreB [68] shares this residue with Plasmodium actin I, whereas canonical actins and also Plasmodium actin II have a hydrophobic residue at this position. Thus, this asparagine may be a relict from an early, polymerization incompetent ancestor.
The structural features described above may explain the parasite actins' unconventional response to ADP. Surprisingly, the state of the nucleotide seems to determine polymerization propensity, but not in the same way as in canonical actins. The tight link between ATP hydrolysis and polymerization in higher eukaryotes has probably been refined during the hundreds of millions of years after the diversion of Apicomplexa. Our data and a recent report proposing an isodesmic polymerization mode for apicomplexan actins [18] suggest that the same has also happened for allosteric regulation of conformational changes taking place upon polymerization. However, it is clear that higher resolution data on the Plasmodium actin filaments are needed in order to find out what kind of conformational changes the parasite actins undergo upon polymerization and what is the arrangement of the protomers in the filament, leading to the altered symmetry compared to canonical F-actin. On the other hand, the distribution of oligomers, as seen on the native gels and DLS (Figs. 7, S5, and S6), suggests that polymerization may involve a nucleation step, the nucleus being either a dimer or trimer, which are the species that disappear early in the process. Thus, we hypothesize that the ADP state may favor nucleation, making ATP hydrolysis a rate-limiting step for polymerization.
The D-loop plays a key role in the conformational changes upon polymerization as well as the conformation and stability of F-actin [28], [45], [47], [48]. Both previous work [26] and our EM analyses reveal differences in the helical architecture of actin I compared to α-actin. In the crystal structures, several regions important for intra-filament contacts in canonical actin filaments show substantial differences between the parasite and opisthokont actins, and the polymerization propensity and filament stability are overall likely a sum of numerous atomic details in the monomers. Yet, the sequence of the α-actin D-loop alone is sufficient to restore the ability of actin I to form long filaments, without altering the symmetry compared to the JAS-stabilized wild-type actin I filaments. Thus, whereas the longitudinal contacts by the D-loop are important for stability, the shape and symmetry of the filaments are determined by other factors.
Actin II shows us that stability can be obtained by other means than the D-loop, probably involving lateral interactions. In addition to the differences we have described in the residues involved in lateral contacts, a candidate responsible for increased stability is residue 200, which is a glycine in Plasmodium actin I and T. gondii actin [69] but serine or threonine in canonical actins as well as Plasmodium actin II and Theileria actin [9], all of which form long filaments. It has been reported that the double mutant G200S/K270M in T. gondii actin leads to an increased filament length when using phalloidin-labeled filaments [69]. However, we were not able to visualize long filaments of this mutant of Plasmodium actin I in polymerizing conditions without JAS (data not shown), indicating that several small changes are cumulatively responsible for the increased stability of actin II filaments.
The tip of the D-loop can adopt a helical conformation, albeit it is disordered in the vast majority of all G-actin structures, and appears mainly intrinsically disordered in solution in all nucleotide states of G-actin [70]. The likely higher helical propensity of the D-loop in Plasmodium actins may affect polymerization and filament stability in at least two different ways. If the helical conformation is more likely to occur in the filamentous form, this might actually facilitate polymerization, which would be in line with the proposed low critical concentration [15] or isodesmic model for polymerization of parasite actins [18]. However, it has also been proposed that the helical form occurs only transiently in the filament or that it is favored in the ADP form and leads to filament destabilization [54], [70]. In this way, a higher helical propensity would contribute to the lower stability of the parasite actin filaments.
Tyrosine hydrogen bonds can contribute substantially to protein stability [71]. Tyr54 is a phosphorylation target and plays a regulatory role in many actins [72]–[74]. For Dictyostelium actin, phosphorylation of this tyrosine increases the critical concentration and controls cell shape changes and spore formation [72]–[74]. In Mimosa pudica L., a contact sensitive plant, where actin is heavily phosphorylated, tyrosine phosphatase inhibitors inhibit the fragmentation of actin filaments during leaf bending [75]. In addition to affecting binding to other proteins, phospho-Tyr54 stabilizes the D-loop conformation [74]. Upon polymerization, this region undergoes a large conformational change, and it seems that the OH group of Tyr54 may be involved in stabilizing interactions [45] that the Phe54 side chain could not fully compensate for. Only 11 of over 300 known actin sequences contain a phenylalanine at this position, and no other substitutions are known. Most of these 11 sequences are actins from Plasmodium or Trypanosoma, both species where actin filaments have not been observed in vivo. Despite the apparent importance of tyrosine at this position for normal actins, a single mutation to phenylalanine in Dictyostelium actin does not affect its polymerization properties [74]. In line with this, we also could not observe long filaments of the actin I F54Y mutant (Fig. 1 F). However, the large effect of the F54Y mutation on the phosphate release rate of actin I (Fig. 6 D) suggests that this residue, indeed, may significantly affect the conformation and flexibility of the D-loop.
Together, the above described structural properties may lead to a higher polymerization propensity but also lower filament stability in the parasite actins by lowering the energy barrier of the transition between monomeric and filamentous forms. Yet, the fact that the replacement of the D-loop alone is sufficient for stabilizing the filaments formed by actin I, while retaining their unique symmetry, is surprising, taking into account how similar the D-loops of the two Plasmodium actins with different stabilities are. This implies that, starting from an unstable filament forming ancestor, actin II has reached its present form mainly using other means than the D-loop for gaining additional filament stability.
Male gametogenesis is a complex, rapid series of cellular events including escape from the host cell, three mitotic divisions, and axoneme assembly, leading to the formation of eight flagellar and highly motile gametes from each gametocyte within 10–20 min from activation. Both actin isoforms are present in male gametocytes of P. berghei [23], but their function in these events is not understood. Actin II is not expressed in the asexual blood stages [23]. Its deletion blocks male gametogenesis, and therefore, these mutant parasites cannot be transmitted through the mosquito [23]. Still, it has not been possible to pinpoint the exact role of actin II. We show that the function of actin II cannot be complemented by actin I, proving distinct molecular functions for the two actins and suggesting that their unique structures and the differences in their ability to form filaments directly translate into different functional characteristics in vivo. By generating transgenic parasites expressing the actin I–α-actin chimera, we found that this mutant protein was able to function almost as well as actin II in vivo. This strongly confirms the in vitro experiments and supports the notion that the D-loop has a significant role in determining the polymerization properties of the parasite actins. Furthermore, we can hypothesize that the reason two actins evolved in Plasmodium, was the need to have actins with different propensities to polymerize in cells lacking a large repertoire of actin-binding proteins.
Another example of distinct general and reproductive actin isoforms can be found in plants, where it was recently shown that animal cytoplasmic but not muscle actins can take over the functions of the plant vegetative actins [76]. Remarkably, also three actins from single-celled protists could carry out the same tasks, suggesting that the properties required for fulfilling the cytoplasmic actin functions during spatial development in multicellular organisms were present already early on in the evolutionary history. However, it seems that the polymerization properties of both Plasmodium and all other actins have evolved separately, starting from a poorly polymerizing ancestor. It would be interesting to see if either of the Plasmodium actins can support spatial development in either plants or animals.
The current hypothesis is that actin I in Plasmodium is required for gliding motility, and the filaments involved need to be short and short-lived. Our data support this, as actin I forms only very short polymers. For the suggested role of actin I in gliding, the formation of long, stable filament seems undesirable [69]. Actin II clearly is able to form long filaments, which may be needed for functions specific to actin II within the mosquito stages, although such functions have not yet been specified. Intriguingly, Plasmodium appears to be the only apicomplexan parasite that has faced the evolutionary pressure for acquiring a second actin isoform that forms stable, long filaments.
Our data provide a structural basis for understanding the different functional properties of the two actin isoforms of Plasmodium spp. These structures represent the, so far, most divergent and primitive actins characterized, and we show that the two isoforms have the most unique biochemical properties, structures, and biological functions of all known actin isoforms. High-resolution structural information will serve as a starting point for understanding these functions in detail and for evaluating the suitability of parasite actins and actin-binding proteins as drug targets.
Purification of G1 was performed as described [40]. Endogenous pig skeletal muscle α-actin was purified as described [36], [77]. P. falciparum actin I (PlasmoDB PF3D7_1246200) and P. berghei actin II (PlasmoDB PBANKA_103010) were expressed in Sf21 cells at 300 K, as described before [36]. A chimera, where residues 40–61 of the P. berghei actin I were replaced by the corresponding residues from α-actin, was cloned into pFastBac HT A (Invitrogen) and expressed in the same way as the wild-type actins. Two point mutations (G115A and F54Y) were introduced to actin I by incorporating the corresponding mutation to the 5′ end of the primers. The parental plasmid was cleaved with DpnI and recirculated with the T4 DNA ligase. The protein coding sequences were confirmed by DNA sequencing. The purification of the wild-type actin–G1 complexes was performed as described [40]. The chimera–G1 was also purified as described before for the two wild-type actins [40], except that HEPES (pH 7.5) was used in the lysis buffer, and size exclusion chromatography was performed in 10 mM HEPES (pH 7.5), 50 mM NaCl, 5 mM dithiothreitol (DTT), 0.2 mM CaCl2 and 0.5 mM ATP. Peak fractions containing the chimera–G1 were pooled and concentrated to 5.6 mg ml−1 for crystallization.
The purification of all the actin variants without G1 was performed essentially as described [40] except for a few modifications, as listed. For actin I, the lysis was carried out in 10 mM HEPES (pH 7.5), 5 mM CaCl2, 250 mM NaCl, 1 mM ATP, 5 mM β-mercaptoethanol, 15 mM imidazole, and size exclusion chromatography was performed in 15 mM HEPES (pH 7.0), 0.5 mM ATP, 5 mM DTT, and 0.2 mM CaCl2. The pH of the lysis buffer for actin II was 8.7, and size exclusion chromatography was performed in 25 mM Tris-HCl (pH 7.5), 0.5 mM ATP, 5 mM DTT, and 0.2 mM CaCl2. For the chimera, lysis was carried out in 20 mM HEPES (pH 7.5), 5 mM CaCl2, 250 mM NaCl, 1 mM ATP, 5 mM β-mercaptoethanol, 15 mM imidazole, and size exclusion chromatography was performed in 15 mM HEPES (pH 7.0), 0.5 mM ATP, 5 mM DTT, and 0.2 mM CaCl2. For DLS and filament length measurements, size exclusion chromatography was performed in 5 mM HEPES (pH 7.5), 0.5 mM ATP, 2 mM DTT, and 0.2 mM CaCl2. DLS was measured using a Wyatt DynaPro platereader-II and 15 or 30 µl of actin I and II at concentrations between 8.5–24 µM at 298 K. The measurements were performed in triplicate and the samples stored at room temperature between the measurements.
ADP-actin was prepared by incubating 50 µl of 10 µM actin with 1–2 mg of hexokinase-agarose beads (Sigma-Aldrich, #H-2653) in 15 mM HEPES pH 7.5, 1 mM ATP, 1 mM tris(2-carboxyethyl)phosphine (TCEP), 0.2 mM CaCl2, 2 mM D-glucose for 1 h at 298 K. As a control reaction, Plasmodium actins I and II were incubated in identical conditions without D-glucose and hexokinase and subsequently run on native PAGE. The residual ATP contamination in ADP stocks was removed by treating them in a similar fashion.
Native PAGE was performed using a running buffer of 25 mM Tris-HCl (pH 8.5), 195 mM glycine, 0.5 mM ATP or ADP, and 0.1 mM CaCl2 or MgCl2. The sample buffer consisted of 25 mM Tris-HCl (pH 8.5), 195 mM glycine, 10% (v/v) glycerol (final concentrations). Actin samples were loaded at a concentration of 6.7 µM in a volume of 10 µl. Commercial TGX 4–20% gradient gels (Biorad) were pre-run for 30 min at 277 K, 100 V before applying the samples. Samples were run for 7 h using the same voltage settings and temperature, with corresponding nucleotides and divalent cations in the running buffer. The gels were stained the next day with Coomassie Brilliant Blue R250.
Relative mobilities were determined by measuring the distance of the bands from the top of the image and dividing this value by that of the monomeric band. In the absence of a reference monomeric band in ADP-ActI (48 h), the absolute value from ATP-ActI (0 h) was used as a reference. The absolute mobilities of the other visible bands in these images had a difference of <2.5%. Gel images were processed and band intensities extracted using ImageJ [78]. A rolling ball background subtraction was applied before manually extracting the intensities.
Actin samples were prepared for the phosphate release assay by treating 10–15 µM purified actin with DOWEX 1X8 to remove nucleotides and free phosphate. After the removal of the nucleotide and phosphate, ATP was replenished by adding a small volume of a concentrated stock solution. Buffer controls were treated in a similar fashion, in order to reset the level of free phosphate and nucleotide compared to the samples. The concentration to be used for determining the release rate was measured from the nucleotide-free solutions in order to reduce the effect of pipetting errors. After the DOWEX treatment, samples were divided in triplicate wells of a UV-transparent 96-well plate (Corning) containing reagents from the EnzChek Phosphate Release Assay (Molecular Probes) without using the reaction buffer, which contains MgCl2 at a final concentration of 1 mM. For calcium measurements, the final reaction contained 1 mM CaCl2 and 0.1 mM MgCl2. The total omission of MgCl2 was not possible, since the coupled enzyme requires magnesium. For magnesium measurements, the respective concentrations were 0.13 mM CaCl2 and 1 mM MgCl2. Formation of the 2-amino-6-mercapto-7-methylpurine from the coupled reaction was measured as absorbance at 360 nm with a kinetic interval of 60 s over a period of 5 h at 298 K. The total measurement volume was 200 µl. Phosphate release rates were calculated from linear parts of the plot (100 to 200 min) using GraphPad PRISM 5.03.
Crystallization and diffraction data collection of both wild-type actin–G1 complexes has been described [40]. The chimera–G1 complex was crystallized similarly, and the final crystallization condition contained 100 mM Tris-HCl (pH 8.0), 8% (w/v) polyethylene glycol (PEG) 20 000, and 2% (v/v) dioxane. Before flash-cooling in liquid nitrogen, the crystal was shortly soaked in 100 mM Tris-HCl (pH 8.5), 14% (w/v) PEG 20 000, 2% (v/v) dioxane, 0.5 mM ATP, 50 mM NaCl, 0.2 mM CaCl2, and 10% (w/v) PEG 400. A diffraction data set to 2.5-Å resolution was collected on a Pilatus 6M detector at the beamline P11, PETRA III (DESY), Hamburg, using a wavelength of 0.92 Å at 100 K. The data (Table 1) were integrated with XDS [79] and scaled with XSCALE [79] using XDSi [80]. The actin II–G1 structure was solved by molecular replacement with Phaser [81] using the α-actin–G1 complex as a search model (PDB code 1P8Z [82]). For actin I–G1 and chimera–G1, the actin II and actin I in complex with gelsolin, respectively, were used as molecular replacement models. The refinement was carried out with PHENIX.refine [83] and manual model building in Coot [84], and structure validation using the MOLPROBITY server [85]. For actin I, actin II, and chimera–G1 complexes, 99.8%, 99.8%, and 99.4% of the amino acids, respectively, were in the allowed regions of the Ramachandran plot. The final electron density maps as well as data and refinement statistics are presented in Fig, S2 A–C and Table 1. The structure figures were prepared using PyMOL and Chimera [86].
Actin (7–13 µM) was polymerized overnight at room temperature. Polymerization was induced by adding 1/10 volume of 10× polymerization buffer [50 mM Tris-HCl (pH 8.0) or HEPES (pH 7.5), 500 mM KCl, 20 mM MgCl2 (in cryo-EM 40 mM MgCl2), 50 mM DTT, and 10 mM ATP] with or without 5–7 µM JAS. In order to concentrate the filaments, actin II and the chimera in F-buffer were spun for 45 min at 435,000 g, and remaining pellet was resuspended into polymerization buffer. 2–3-µl aliquots of the samples were diluted in the polymerization buffer before applying them on glow-discharged grids (CF-300CU, Electron Microscopy Sciences) and stained with 1% (w/v) uranyl acetate or potassium phospho-tungstate (pH 7.0). The grids were examined with Tecnai G2 Spirit (100 kV) or FEI Tecnai F20 microscopes (200 kV). Filament lengths were measured using ImageJ [78]. Many of the longest (>1 µm) measured filaments are fragments, as both ends were not always visible in the images.
Polymerized samples were applied in 3-µl aliquots onto freshly glow-discharged holey carbon grids (Quantifoil R 2/2) at 295 K and 70% humidity and vitrified in liquid ethane using a Leica EM GP vitrification robot. Specimens were held in a Gatan 626 cryoholder maintained at 93 K for imaging in a FEI Tecnai F20 microscope operated at 200 kV. Micrographs were recorded under low dose conditions on a Gatan Ultrascan 4000 CCD camera at a magnification of 69,000 to give a final pixel size of 2.21 Å.
The contrast transfer function (CTF) of the micrographs was determined using CTFFIND [87]. A total of 330 (actin I), 56 (actin II) and 457 (chimera) filaments were selected using e2helixboxer.py from the EMAN2 suite [88]. For classification, segments were excised using a mean step size of 30 Å and an additional random shift along the helix between -15 and 15 Å to avoid high-resolution artifacts in the class average power spectra introduced by regularly shifted images. The segments were further corrected for their CTF by phase flipping, and aligned to the vertical axis. This resulted in 4,581 segments for actin I, 968 for actin II, and 8,052 for the chimera actin. Two-dimensional (2D) classification of helical segments was performed using the SPARX k-means algorithm [27]. The segments were iteratively classified and aligned against a subset of class-averages chosen based on their quality with a total of four iterations. At each cycle, multiple copies of the chosen references were created by applying integer y-shifts ranging from −15 Å to +15 Å in order to be able to reduce the Y-shift search range during alignment to less than half of the step size in order to avoid summation of successive images on a filament shifted at the same axial position. The total number of class averages used to measure the cross-over distance was 40 for actin I and the chimera actin, and 20 for actin II. In addition, Eigen images were calculated and the corresponding pitch distances were measured. For 3D structure determination of actin I filaments, 2,182 segments were excised using a regular step size of 70 Å, convolved by their respective CTF and further reconstructed as described [30] using the software SPRING [89]. In addition, symmetry refinement was performed using the IHRSR method [90] by systematically varying the initial helical rises and azimuthal rotations from 26 to 30 Å (step 1 Å) and from 164 to 170° (step 1°), respectively. More specifically, 25 iterations of refinement were computed with SPIDER, using a solid cylinder of 100 Å in diameter as a starting model. The symmetry parameters were refined with the hsearch program after the second refinement iteration, using a step size of 0.03 Å for helical rise and of 0.05° for azimuthal rotation.
The actin2 complementation and replacement constructs were made in a derivative of the pL0006 vector, which encodes human DHFR conferring resistance to the drug WR99210 [91], [92]. The design of the constructs is described in detail elsewhere [24], and the three different constructs were produced following the same strategy. Briefly, 2.7 kilobase pairs of the promoter and 728 base pairs of the 3′-flanking region of the P. berghei actin2 gene were amplified from gDNA and cloned into the vector. For the act2rep construct, P. bergei actin I complete ORF including start and stop codon was amplified from gDNA and cloned between the actin2 promoter and the 3′ flanking region of actin2. The same strategy was followed for the act2com construct using the P. berghei actin2 ORF and the act1chi construct. The plasmids were linearized before transfection of the recipient act2−::mCherry parasite line [60]. Parasites were cloned as described [93]. Correct integration was verified by PCR genotyping and Southern blotting. Exflagellation was scored after diluting blood from an infected mouse in exflagellation medium [23] and incubating the samples for 10–20 min at 292 K. The exflagellation events were counted under a light microscope.
The structure factors and coordinates for all three crystal structures have been submitted to the PDB under the codes 4cbu, 4cbw, and 4cbx. The actin I EM map has been deposited to the EMDB under the accession code EMD-2572.
Figure S1 shows an alignment of apicomplexan and canonical actin sequences. Figure S2 shows the electron density maps around the ATP-binding site of the Plasmodium actins and the chimera and the gelsolin complexes for actin I and II in two orientations. Figure S3 depicts root mean square deviations between Plasmodium and canonical actin structures. Figure S4 shows the cryo-EM analysis of the actin I–α-actin chimera filaments. Figure S5 shows native PAGE analysis of the Plasmodium actins in the calcium-bound form. Figure S6 shows the DLS analysis of the oligomerization of the parasite actins over time. Figure S7 shows native gels of the Plasmodium actins in the magnesium-bound form. Video S1 shows an exflagellation event of a male P. berghei gametocyte.
|
10.1371/journal.ppat.1006131 | Blocking two-component signalling enhances Candida albicans virulence and reveals adaptive mechanisms that counteract sustained SAPK activation | The Ypd1 phosphorelay protein is a central constituent of fungal two-component signal transduction pathways. Inhibition of Ypd1 in Saccharomyces cerevisiae and Cryptococcus neoformans is lethal due to the sustained activation of the ‘p38-related’ Hog1 stress-activated protein kinase (SAPK). As two-component signalling proteins are not found in animals, Ypd1 is considered to be a prime antifungal target. However, a major fungal pathogen of humans, Candida albicans, can survive the concomitant sustained activation of Hog1 that occurs in cells lacking YPD1. Here we show that the sustained activation of Hog1 upon Ypd1 loss is mediated through the Ssk1 response regulator. Moreover, we present evidence that C. albicans survives SAPK activation in the short-term, following Ypd1 loss, by triggering the induction of protein tyrosine phosphatase-encoding genes which prevent the accumulation of lethal levels of phosphorylated Hog1. In addition, our studies reveal an unpredicted, reversible, mechanism that acts to substantially reduce the levels of phosphorylated Hog1 in ypd1Δ cells following long-term sustained SAPK activation. Indeed, over time, ypd1Δ cells become phenotypically indistinguishable from wild-type cells. Importantly, we also find that drug-induced down-regulation of YPD1 expression actually enhances the virulence of C. albicans in two distinct animal infection models. Investigating the underlying causes of this increased virulence, revealed that drug-mediated repression of YPD1 expression promotes hyphal growth both within murine kidneys, and following phagocytosis, thus increasing the efficacy by which C. albicans kills macrophages. Taken together, these findings challenge the targeting of Ypd1 proteins as a general antifungal strategy and reveal novel cellular adaptation mechanisms to sustained SAPK activation.
| As fungi-attributed human deaths are increasing, there is an urgent need to develop new antifungal treatments. Two-component related proteins, such as the Ypd1 phosphorelay protein, have been heralded as antifungal targets as they are not found in humans and because inactivation of YPD1 in several different fungi causes sustained SAPK activation and cell death. However, we have discovered that inactivation of YPD1 in the major human pathogen, Candida albicans, actually enhances virulence. Furthermore, we reveal that this fungus adapts to the sustained activation of the Hog1 SAPK triggered by Ypd1 loss by mounting distinct mechanisms that actively reduce the level of phosphorylated Hog1. These findings question the validity of Ypd1 proteins as broad-spectrum antifungal targets and provide insights into the cellular adaptation to sustained SAPK activation.
| Candida albicans is the leading cause of systemic fungal infections in humans resulting in over 400,000 deaths each year in immuno-compromised patients [1]. The ability of C. albicans to adapt to host-imposed stresses encountered during infection is an important virulence trait [2]. Central to fungal stress responses are the stress-activated protein kinases (SAPKs), which are conserved eukaryotic signalling enzymes that allow cells to adapt to environmental change [3, 4]. In C. albicans, the Hog1 SAPK is activated in response to diverse, physiologically relevant, stress conditions, and cells lacking Hog1 are acutely sensitive to such stresses [5–7]. Consistent with the vital role of the Hog1 SAPK in stress survival, C. albicans cells lacking HOG1 display significantly attenuated virulence in systemic, commensal, and phagocyte infection models [8–11].
All SAPK activation mechanisms reported to date result in the phosphorylation of conserved threonine and tyrosine residues located within the TGY motif of the catalytic domain of the kinase [3]. Such pathways are tightly regulated as the nature of the response is dependent on the extent and period of SAPK activation. For example, in the model yeast Saccharomyces cerevisiae, transient activation of the Hog1 SAPK is vital to survive osmotic stress [4], whereas sustained activation triggers programmed cell death [12]. Similarly in human cells transient activation of the p38 SAPK promotes stress-induced gene expression and cellular proliferation [13], whereas sustained SAPK activation triggers apoptosis [14]. In contrast, much less is known regarding the regulation and cellular consequences of sustained SAPK activation in C. albicans.
Despite the availability of several antifungal drugs, the high mortality rate associated with C. albicans systemic infections and the emergence of drug resistant strains highlights the urgent clinical need for new anti-fungal therapies [15]. Although Hog1 is an essential virulence determinant in C. albicans, the conservation with highly related SAPKs in human cells suggests that Hog1 itself may be unsuitable as an antifungal target. Instead, there has been much interest in identifying fungal-specific regulators of SAPKs as potential drug targets. Candidate targets include two-component related phosphorelay systems which constitute an important mechanism employed by fungi, but not mammals, to sense and relay specific stress signals to SAPK modules [16]. In S. cerevisiae, this system is comprised of a hybrid histidine kinase (Sln1), an intermediary phosphorelay protein (Ypd1), and a response regulator protein (Ssk1) (Fig 1). Following osmotic stress, the Sln1 histidine kinase is inactivated, which halts phosphorelay through Ypd1, and consequently leads to the rapid dephosphorylation of Ssk1 [17]. Dephosphorylated Ssk1 is a potent activator of the Ssk2/Ssk22 MAPKKKs which regulate Hog1 activation [18, 19]. Significantly, loss of either Sln1 or Ypd1 function in S. cerevisiae is lethal [20], due to the accumulation of unphosphorylated Ssk1 and the resulting sustained Hog1 activation which triggers apoptosis-mediated cell death [12]. It is likely that sustained Hog1 activation can also not be tolerated in the human fungal pathogen, Cryptococcus neoformans, as Ypd1 is essential for the viability of cells containing Hog1 [21]. The essential nature of Ypd1 in these fungi, and the many reports illustrating the importance of two-component proteins in fungal pathogenicity, has fuelled interest in targeting Ypd1 for antifungal drug development (reviewed in [22]).
C. albicans has seven two-component proteins; three histidine kinases (Sln1, Chk1, Nik1), three response regulators (Ssk1, Skn7, Crr1/Srr1), and a single phosphorelay protein (Ypd1) [23]. In S. cerevisiae, Ypd1 plays a pivotal role in mediating all phosphorelay events from the upstream histidine kinases to the downstream Ssk1 and Skn7 response regulators [17, 24], which supports the concept that Ypd1 is an appropriate antifungal target. Hence, in this study, we investigated the impact of inactivating Ypd1 upon stress signalling and virulence of C. albicans. As reported recently, we found that C. albicans can survive deletion of YPD1 [25]. Here we extend this finding by illustrating that C. albicans survives the sustained SAPK activation following Ypd1 loss by evoking multiple mechanisms to reduce the level of phosphorylated Hog1. Furthermore, we demonstrate that inactivation of Ypd1 during infection actually increases the virulence of C. albicans in a number of infection models, revealing that Ypd1 may not be a suitable target for anti-fungal drug development.
C. albicans contains a single homologue of the S. cerevisiae phosphorelay protein Ypd1 [26]. Although deletion of Ypd1 results in a lethal phenotype in both S. cerevisiae and C. neoformans [20, 21], a recent study revealed that YPD1 is not an essential gene in C. albicans, with ypd1Δ cells instead displaying a slow growth phenotype [25]. To investigate this further we created a C. albicans strain, tetO-YPD1 (Fig 2A), in which one allele of YPD1 was deleted and the remaining allele placed under the control of a doxycycline-repressible promoter [27]. Northern analysis confirmed that treatment of tetO-YPD1 cells with doxycycline caused a rapid decrease in YPD1 mRNA levels (Fig 2B). However, whilst repression of YPD1 expression did result in a slower growth rate (Fig 2C, upper panel), the cells were viable. Furthermore, consistent with the previous study [25], we were able to generate a viable homozygous ypd1Δ null mutant which displayed a slower growth rate compared to wild-type cells (Fig 2C, lower panel).
Deletion of YPD1 is lethal in S. cerevisiae due to constitutive SAPK activation. Consistent with previous findings [25], we found that repression of YPD1 expression in tetO-YPD1 cells (Fig 2D, upper panel; S1 Fig), or deletion of YPD1 (Fig 2D, lower panel), also stimulated high levels of Hog1 phosphorylation in C. albicans. Together these data indicate that the inhibitory effect of YPD1 on SAPK activation in S. cerevisiae is conserved in C. albicans. However, it was possible that the phosphorylated Hog1 detected did not result in activation of Hog1-dependent downstream events. To test this possibility the expression levels of the Hog1-dependent genes, GPD2 and RHR2 [28], important for glycerol biosynthesis were examined upon repression or deletion of YPD1 (Fig 2E). Both genes were found to be up-regulated and, furthermore, as expected, increased intracellular glycerol concentrations were observed in ypd1Δ cells in the absence of stress (Fig 2F). Thus collectively, these data confirm that the phosphorylated Hog1 kinase triggered by inactivation of Ypd1 in C. albicans is active. However, in contrast to S. cerevisiae and C. neoformans constitutive Hog1 activation does not result in loss of viability in C. albicans.
Analyses of cells with loss of Ypd1 function, either by repression or deletion of the YPD1 gene, revealed identical morphological abnormalities and stress-resistance profiles. For example, loss of YPD1 resulted in swollen pseudohyphal-like cells (Fig 3A), possibly due to the increased intracellular levels of the osmolyte glycerol that occurs upon inactivation of Ypd1 (Fig 2F). In addition, as reported previously [25], cells lacking Ypd1 were highly flocculent as demonstrated by their rapid sedimentation rate (Fig 3A). Interestingly, cells lacking Ypd1 displayed acute sensitivity to sodium arsenite, increased resistance to the organic peroxide tert-butyl hydroperoxide (t-BOOH), and wild-type levels of resistance to osmotic stress (Fig 3B). Loss of YPD1 also resulted in increased sensitivity to the cell wall perturbing agent calcofluor white (Fig 3B). This is consistent with Hog1 activation in ypd1Δ cells (Fig 2D), as hog1Δ cells display significant resistance to this drug [6].
Loss of Ypd1 function is predicted to perturb phosphorelay to all three response regulator proteins in C. albicans; Ssk1, Skn7 and Crr1/Srr1 [23, 29, 30]. In S. cerevisiae it is the accumulation of the unphosphorylated Ssk1 response regulator which triggers hyperactivation of the Hog1 SAPK [18]. Hence, to investigate which phenotypes associated with Ypd1 loss in C. albicans were due to Ssk1-mediated activation of Hog1, hog1Δ ypd1Δ and ssk1Δ ypd1Δ double mutant strains were created. All of the ypd1Δ-associated phenotypes, described above, were found to be dependent on both Hog1 and Ssk1. For example, the swollen pseudohyphal-like morphology associated with ypd1Δ mutant cells was repressed in the absence of either HOG1 or SSK1 (Fig 3C). In fact, the hog1Δ ypd1Δ double mutant cells instead displayed the morphological defects characteristic of hog1Δ cells (Fig 3C). The high glycerol levels characteristic of ypd1Δ cells were dependent on Hog1 (Fig 3D), and the stress-phenotypes associated with deletion of YPD1 were not maintained in hog1Δ ypd1Δ or ssk1Δ ypd1Δ double mutant cells (Fig 3E). Indeed, in all the conditions examined, cells lacking HOG1 and YPD1 displayed similar stress phenotypes as hog1Δ cells, and ssk1Δ ypd1Δ cells were phenotypically similar to ssk1Δ cells. Importantly, confirming the link between Ypd1 and Hog1 activity, reintegration of HOG1 or SSK1 into the hog1Δ ypd1Δ and ssk1Δ ypd1Δ mutants, respectively, resulted in cells that were phenotypically identical to the ypd1Δ strain (Fig 3C and 3E). Furthermore, the hyper-phosphorylation of Hog1 detected in ypd1Δ cells was absent in ssk1Δ ypd1Δ cells, but was restored upon reintegration of SSK1 (Fig 3F). This result confirms that Ssk1 is essential for the high basal level of Hog1 phosphorylation in ypd1Δ cells. Taken together, these results are consistent with the model that accumulation of unphosphorylated Ssk1 triggers the ypd1Δ-dependent sustained activation of Hog1 in C. albicans and, moreover, that this activation underlies the morphological and stress phenotypes associated with loss of Ypd1.
We next investigated the molecular mechanism underlying the ability of C. albicans to survive sustained Hog1 activation. In S. cerevisiae, the lethality associated with YPD1 loss can be by-passed by the artificial over-expression of either of the protein tyrosine phosphatases, Ptp2 or Ptp3, which normally dephosphorylate and negatively regulate Hog1 activity [31, 32]. Indeed, YPD1 (tyrosine phosphatase dependent) was initially identified in a synthetic lethal screen for S. cerevisiae mutants whose growth was dependent on the expression of PTP2 [33]. Similarly, Ptp2 in C. neoformans [34] and Ptp2 and Ptp3 in C. albicans [35], have been reported to negatively regulate the respective Hog1 SAPK pathways in these pathogenic fungi. Hence, it was possible that the ability of C. albicans ypd1Δ cells to retain viability was linked to the activity of the Ptp2 and Ptp3 phosphatases. Strikingly, we found that the expression of PTP3 is significantly induced in ypd1Δ cells compared to wild-type cells, and PTP2 is also up-regulated albeit to a lesser extent (Fig 4A). Similar findings were also observed upon doxycycline-mediated repression of YPD1 in tetO-YPD1 cells (Fig 4B and 4C). Significant induction of PTP3 occurred with similar kinetics as the increase in Hog1 activation, and some induction of PTP2 was also evident (Fig 4B and 4C). Previous studies revealed that arsenite is a potent inhibitor of protein tyrosine phosphatases that regulate SAPK pathways in both yeast and humans [36–38]. Interestingly, C. albicans cells lacking Ypd1 are acutely sensitive to arsenite (Fig 3B), raising the possibility that arsenite-mediated inhibition of Ptp2 and/or Ptp3 causes catastrophic levels of Hog1 phosphorylation and ultimately cell death. Indeed, in agreement with this hypothesis, sodium arsenite treatment massively increased Hog1 phosphorylation in ypd1Δ but not wild-type cells (Fig 4D). To confirm that arsenite primarily activates Hog1 through phosphatase inhibition, we next examined Hog1 phosphorylation in cells expressing a mutant version of the MAPKK Pbs2 (Pbs2DD), where the activating phosphorylation sites of Pbs2 are mutated to phosphomimetic aspartate residues. The Pbs2DD mutant protein yields basal activation of Hog1 but prevents further upstream signalling events to the SAPK [10]. Consistent with arsenite-dependent inhibition of Hog1-specific phosphatase(s), significant induction of Hog1 phosphorylation was observed in Pbs2DD cells following arsenite treatment (Fig 4E). Furthermore, this arsenite-induction of Hog1 requires Pbs2 activity as arsenite-induced Hog1 phosphorylation does not occur in cells expressing an inactive Pbs2AA kinase where the activating phosphorylation sites of Pbs2 are mutated to alanine residues mimicking hypophosphorylation (Fig 4E, compare PBS2 and PBS2DD with the PBS2AA lanes). Thus, these results strongly suggest that the ability of C. albicans cells to survive sustained activation of Hog1 upon loss of Ypd1 function is due to the action of Ptp phosphatases that reduce the levels of phosphorylated Hog1.
To test this directly, we sought to delete both PTP2 and PTP3 in tetO-YPD1 cells with the prediction that doxycycline-mediated repression of YPD1 in this background would result in catastrophic levels of Hog1 activation and cell death. Initially we focused on deleting PTP3 as this gene shows the greatest induction upon loss of YPD1 (Fig 4A and 4B). Deletion of PTP3 in tetO-YPD1 cells resulted in notably higher levels of Hog1 phosphorylation upon repression of YPD1 expression (Fig 4F). Moreover, deletion of PTP3 had a dramatic impact on the growth of tetO-YPD1 cells upon doxycycline-mediated repression of YPD1, but not in the absence of doxycycline (Fig 4G). These results support the model that induction of PTP3 promotes C. albicans survival following Ypd1 loss by limiting Hog1 phosphorylation. However, repression of YPD1 in ptp3Δ cells did not result in a lethal phenotype which is consistent with previous findings that Ptp2 and Ptp3 function redundantly to regulate C. albicans Hog1 [35]. Upon attempting to generate tetO-YPD1 ptp3Δ ptp2Δ cells we found that only one copy of PTP2 could be deleted. This was unexpected as a ptp2Δ ptp3Δ double mutant has previously been characterised [35]. Strikingly, however, tetO-YPD1 ptp3Δ ptp2/PTP2 cells exhibited a much higher basal level of Hog1 activation than that seen cells lacking PTP3 (Fig 4F). Following doxycycline-mediated repression of YPD1 in ptp3Δ ptp2/PTP2 cells, further substantial increases in Hog1 phosphorylation were detected compared to that observed in tetO-YPD1 ptp3Δ cells (Fig 4F). Indeed, consistent with the additive effect of deleting one copy of PTP2, the growth of tetO-YPD1 ptp3Δ ptp2/PTP2 cells was barely detectable upon repression of YPD1 expression (Fig 4G). However, it is important to note that such cells also display a slow growth phenotype under −DOX conditions when YPD1 is expressed (Fig 4G). Nonetheless, taken together, these data support the model that the induction of PTP3 and PTP2, together with the basal expression of PTP2, facilitate C. albicans survival following loss of the Ypd1 phosphorelay protein.
Although C. albicans can clearly tolerate sustained Hog1 activation, cells lacking YPD1 display reduced fitness compared to wild-type cells (Fig 2C). Strikingly, however, we noted that ypd1Δ cells, when maintained on rich media plates, lost the morphological abnormalities associated with inactivation of Ypd1 function. Specifically, the swollen pseudohyphal morphology observed in freshly isolated ypd1Δ cells (Day 1), was largely replaced with normal budding cells after incubation on rich media plates for 13 days (Fig 5A). Because of these findings we examined whether the stress phenotypes associated with YPD1 loss also changed over time. In agreement with our previous findings (Fig 3), freshly isolated ypd1Δ cells (Day 1) exhibited a slow growth rate (indicated by small colony size), and increased sensitivity to sodium arsenite and calcofluor white. However, by day 13 the slow growth phenotype was lost, and the sensitivity to sodium arsenite and calcofluor white mimicked that exhibited by wild-type cells (Fig 5B). Consistent with this loss of morphological and stress sensitive phenotypes, we found that the high basal level of Hog1 phosphorylation, triggered by loss of YPD1, decreased over a 13 day period (Fig 5C). Hog1 phosphorylation levels were significantly reduced by day 10, and by day 13 levels were similar to that seen in wild-type cells. Consistent with a reduction in Hog1 activation, the high basal level of expression of the Hog1 target gene, GPD2, also declined over time (Fig 5D). The decrease in Hog1 phosphorylation was not due to a reduction in Hog1 protein and/or HOG1 mRNA levels which remained constant over the 13 day experiment (Fig 5C and 5D). This suggests that the sustained SAPK phosphorylation caused by loss of Ypd1 triggers adaptation within the cell that results in lower levels of activated Hog1.
As we had found the negative regulators PTP3 and PTP2 to be induced in ypd1Δ cells, we asked whether a further induction in their expression could contribute to the adaptation mechanism resulting in the time-dependent decline in Hog1 phosphorylation levels. In agreement with our previous findings (Fig 4A), we found PTP3 and PTP2 to be up-regulated in ypd1Δ cells at day 1 and this was maintained at day 7 (Fig 5B). However, by day 10 the levels of PTP3 and PTP2 in ypd1Δ cells had returned to wild-type levels. Hence, although C. albicans appears to adapt to sustained SAPK activation in the short term by triggering the up-regulation of PTP2 and PTP3, this up-regulation is temporary. Indeed, the induction of PTP2 and PTP3 is only observed in cells in which significant levels of Hog1 activation are seen (compare Fig 5C and 5D). This suggests that C. albicans adapts to Hog1 activation in the long term by via another mechanism(s) independent of Ptp2 and Ptp3. Although the specific adaptation mechanism has not been identified, these results illustrate that C. albicans cells adapt to YPD1 loss over time by reducing the levels of phosphorylated Hog1 such that the negative effects of sustained SAPK activation are ablated, and ypd1Δ cells become phenotypically similar to wild-type cells.
We asked whether the reduction in Hog1 phosphorylation levels following long-term sustained Hog1 activation was irreversible. Cells lacking ypd1Δ that had been maintained for 11 days on solid rich media, were then either kept on this media (13 day) or patched onto fresh media in the absence (-NaCl) or presence (+NaCl) of 0.3M NaCl, a stress condition known to transiently activate Hog1 (Fig 6A). Cells were taken from these plates after 2 days, sub-cultured in liquid media, and phosphorylation and total levels of Hog1 examined (Fig 6A). As expected, a high basal level of Hog1 phosphorylation was absent in ypd1Δ cells at 13 days (Fig 6B). However, passage of ypd1Δ cells over plates containing NaCl, but not lacking NaCl, resulted in a restoration of Hog1 phosphorylation (Fig 6B). To determine whether NaCl treatment could fully restore Hog1 activation in ypd1Δ cells, we compared the level of Hog1 phosphorylation in ypd1Δ cells over the 13 day time course with that seen following passage over NaCl plates. Whilst exposure of ypd1Δ cells to NaCl does restore a high basal level of Hog1 phosphorylation, this is not to the same level as that seen in day 1 samples (Fig 6C). Therefore, these data reveal that a transient exposure of ypd1Δ cells to osmotic stress can partially over-ride the adaptation mechanism(s) that prevents sustained Hog1 phosphorylation. Wild-type cells did not exhibit an increase in Hog1 phosphorylation after passage over media containing NaCl, illustrating that the phosphorylation observed in ypd1Δ cells is due to a re-establishment of sustained Hog1 activation, rather than stress-induced activation of the SAPK (Fig 6B). Furthermore, the restoration of sustained Hog1 phosphorylation in ypd1Δ cells, resulted in the return of a swollen pseudohyphal morphology (Fig 6D). Collectively, the results indicate that the cellular adaptation mechanisms that reduce Hog1 activation in cells lacking YPD1 can be partially over-ridden in the presence of stress that requires Hog1 function.
The lethality associated with the deletion of YPD1 in S. cerevisiae and the fungal pathogen C. neoformans [20, 21], has led to much interest in this phosphorelay protein family as a potential prime antifungal target [22]. Hence, we next examined the impact of doxycycline-mediated repression of YPD1 on C. albicans virulence. The doxycycline-regulatable gene expression system has been successfully used to control C. albicans gene expression during infection in both a mouse model of systemic candidiasis [27, 39], and in a Caenorhabditis elegans infection model [40]. Furthermore, repression of YPD1 expression after infection has three clear advantages over testing the virulence of ypd1Δ cells directly. Firstly, this avoids the problems of accurately obtaining an inoculum size with the highly flocculent and filamentous ypd1Δ strain. Secondly, complications arising from lack of efficient dissemination of a highly flocculent strain are circumvented. Finally, doxycycline-mediated repression of YPD1 following infection more closely mimics the drug-induced inactivation of a particular fungal target following infection.
Initially we employed the C. albicans-C. elegans liquid medium pathogenesis assay [40] to examine the impact of YPD1 repression on virulence. Worms were infected with tetO-YPD1 C. albicans cells which had been grown in the absence of doxycycline and thus expressing YPD1 (Fig 2B). Subsequently, the animals were transferred to liquid medium in the presence (+DOX) or absence (-DOX) of doxycycline. Strikingly, we observed a significant increase in C. elegans killing after infection with tetO-YPD1 cells in the presence of doxycycline in which YPD1 expression is repressed (Fig 7A). A total of 41% of infected animals died after 72 h in liquid medium containing doxycycline compared to 11% mortality in media lacking doxycycline (P<0.001). Importantly, consistent with previous reports [40], doxycycline did not affect the pathogenicity of wild-type C. albicans strains towards C. elegans (S2 Fig). Hence, taken together, these results suggest that doxycycline-mediated repression of YPD1 expression actually enhances the virulence of C. albicans towards C. elegans.
To investigate whether these results could be corroborated in a distinct infection model, we employed the three day murine intravenous challenge model of C. albicans infection [41, 42]. This model combines weight loss and kidney fungal burden measurements following 3 days of infection to give an ‘outcome score’. A higher outcome score is indicative of greater weight loss and higher fungal burdens and thus increased virulence. Twelve mice were infected intravenously with tetO-YPD1 cells grown in medium lacking doxycycline. Following infection, one group of mice (n = 6) were orally dosed with doxycycline daily (+DOX) to repress YPD1 expression, and the second group of placebo treated mice (n = 6) given water (-DOX). Significantly, inhibition of YPD1 expression during infection resulted in greater weight loss, increased kidney fungal burdens, and thus higher outcome scores compared to cells which continue to express YPD1 (Fig 7B). Consistent with previous studies [27, 39], doxycycline treatment alone does not impact on C. albicans virulence in this mouse model of systemic candidiasis (S3 Fig). Histological analysis of the kidneys revealed that there was significant inflammation associated with infection in the doxycycline treated animals (Fig 7C panel i), with little inflammation seen for the placebo treated animals (Fig 7C panel iv). Clusters of filamentous fungal cells were obvious in the doxycycline-treated mouse kidneys (Fig 7C panels ii & iii), whereas only isolated fungal cells were found in the placebo treated kidneys (Fig 7C panels v & vi). This indicates that the CFU measurements underestimate the increased fungal burden following repression of YPD1 expression, presumably as a consequence of the increased filamentation. Nonetheless statistical analysis revealed that the difference between +DOX and −DOX cells was significant for all three parameters, including fungal burden (Fig 7B). Collectively, these studies illustrate that repression of YPD1 expression during infection in two distinct animal models enhances the virulence of C. albicans. These results have clinical significance as they predict that a drug designed to block two-component signalling in C. albicans could actually promote the virulence of this major human fungal pathogen.
To further investigate how Ypd1 loss promotes C. albicans virulence, we employed live cell video microscopy to follow the impact of inhibition of YPD1 expression on C. albicans-macrophage interactions. Specifically, tetO-YPD1 cells were treated with (+DOX) or without (-DOX) doxycycline for 3 h prior to co-incubation with murine J774.1 macrophages, which were then grown in media +/-DOX, respectively. Quantitatively, there were no significant differences between the migration speed of J774.1 macrophages towards tetO-YPD1 cells treated or not with doxycycline, or rate of engulfment of fungal cells (S4 Fig). Importantly, however, doxycycline mediated-repression of YPD1 resulted in C. albicans cells that displayed a significantly enhanced ability to kill macrophages. It was observed that 82±4.1% of macrophages were killed following co-incubation with tetO-YPD1 cells in the presence of doxycycline (+DOX), compared to 60±3.0% macrophages killed in the absence of doxycycline (-DOX) (P<0.01) (Fig 7D). The ability of C. albicans to transition to the hyphal form following phagocytosis is pivotal in triggering macrophage death [43]. Thus, the rate of hyphae formation of tetO-YPD1 cells following phagocytosis by macrophages was measured. Hyphal growth was significantly faster in tetO-YPD1 cells grown in the presence of doxycycline (0.35±0.032 μm/min) than in the absence of doxycycline (0.26 ± 0.023 μm/min) (Fig 7E). These results illustrate that repression of YPD1 expression promotes hyphal growth following phagocytosis which in turn likely enhances the ability of C. albicans to kill macrophages. Given the importance of macrophages in immune responses to C. albicans infections, the increased capacity of C. albicans cells lacking Ypd1 to kill macrophages likely contributes to the enhanced virulence observed in a murine model of systemic infection.
The generation of fungal pathogen-specific drugs is hindered by the conservation of many potential drug-targets in the human host. Thus the complete absence of two-component related proteins in metazoans, but their presence in fungi, has rendered such pathways attractive drug-targets [22, 44, 45]. In S. cerevisiae, and the human fungal pathogen C. neoformans, loss of the two-component phosphorelay protein Ypd1 causes lethality due to the sustained activation of their respective SAPK pathways [20, 21]. However, as reported during the course of this work [25], a major human fungal pathogen, C. albicans, can tolerate sustained activation of the Hog1 SAPK pathway triggered by loss of Ypd1. We have significantly advanced our understanding of this observation in three main ways. Firstly we find that the constitutive activation of Hog1 in cells lacking YPD1 is mediated through two-component mediated regulation of the Ssk1 response regulator, and that the pleiotropic phenotypes associated with Ypd1 loss are dependent on Ssk1-mediated Hog1 activation. Secondly, we have provided novel insight into the mechanisms by which C. albicans survives and adapts to sustained SAPK activation. Thirdly, we show that inactivation of YPD1 promotes, rather than reduces, the virulence of C. albicans, and we provide evidence to suggest that this is mediated at least in part through effects on fungus-phagocyte interactions. A model summarising the major findings from this work is depicted in Fig 8.
One question raised by this study is why is sustained SAPK activation tolerated in some fungi but not others? In S. cerevisiae, prolonged SAPK activation is reported to lead to cell death by causing an increase in reactive oxygen species (ROS) thus triggering apoptosis [12]. Although we do not know whether sustained Hog1 activation triggers an increase in intracellular ROS in C. albicans, it is worth noting that C. albicans is considerably more resistant to oxidative stress than S. cerevisiae [46]. In C. neoformans, lethality due to Hog1 hyperactivation is attributed to the over-accumulation of intracellular glycerol [47]. In C. albicans, sustained Hog1 activation in ypd1Δ cells also triggers an increase in intracellular glycerol levels, presumably contributing to the associated swollen cell phenotype. Interestingly, this is tolerated in these mutant cells and indeed it is noteworthy that C. albicans can withstand higher levels of osmotic stress (which triggers glycerol biosynthesis) than many other fungal species [46]. However, perhaps a key factor in maintaining the viability of ypd1Δ cells, is the significant induction of the tyrosine phosphatase encoding genes PTP2 and PTP3 which negatively regulate Hog1 phosphorylation [35]. Interestingly, the lethality associated with ypd1Δ-mediated sustained Hog1 activation in S. cerevisiae can be by-passed by artificial over-expression of either PTP2 or PTP3 [31, 32], suggesting that C. albicans actually adopts this strategy to prevent catastrophic levels of Hog1 activation. Consistent with this hypothesis is the observation that treatment of C. albicans ypd1Δ cells with the tyrosine phosphatase inhibitor, arsenite, triggers a significant increase in levels of Hog1 phosphorylation and cell death. Furthermore, deletion of PTP3 together with deletion of one copy of PTP2, triggers dramatic increases in the level of phosphorylated Hog1 in tetO-YPD1 cells and furthermore, virtually abolishes cell growth following doxycycline-mediated repression of YPD1. It is not clear why we could not generate tetO-YPD1 cells lacking both PTP3 and PTP2, as a double ptp3Δ ptp2Δ mutant has previously been characterised [35], although it is possible that basal YPD1 expression levels are altered in the tetO-YPD1 background. Notably, however, we find that the basal level of expression of PTP2 plays a major role in preventing Hog1 activation under non-stress conditions, as significant increases in the levels of Hog1 phosphorylation are seen upon deleting one copy of PTP2. Similar findings have been reported in both S. cerevisiae and C. neoformans where basal levels of Hog1 are significantly increased in ptp2Δ mutant cells [31, 34]. Taken together, the data presented in this paper strongly support the model that induction of PTP3 and PTP2 expression, together with the basal expression of PTP2, allow C. albicans to survive loss of the Ypd1 phosphorelay protein.
The viability of C. albicans cells lacking Ypd1 has allowed an investigation of how fungal cells adapt to long-term sustained SAPK activation. Remarkably, we find that cells evoke a mechanism that prevents the long-term constitutive phosphorylation of Hog1. Moreover, this decrease in Hog1 phosphorylation is actually accompanied by a reduction in the levels of PTP2 and PTP3, the negative regulators of Hog1, raising the possibility that an upstream signalling branch to Hog1 is inhibited instead. Furthermore, this mechanism is reversible as it can be over-ridden following a transient exposure to stress that requires Hog1 activity. Do these observations that C. albicans can survive and adapt to sustained SAPK activation have physiological relevance? As a human commensal, this fungal pathogen is continuously exposed to host-imposed stresses. Indeed, Hog1 is vital for C. albicans to exist commensally in the gut [11], cause systemic infections [8, 10], and survive phagocytosis [9]. Thus, the ability of this pathogen to adapt to sustained SAPK activation by actively modulating Hog1 phosphorylation levels may be important to promote survival in certain host niches that continuously generate a stressful environment. Furthermore, we propose that the capacity of C. albicans to restore levels of SAPK phosphorylation, upon subsequent exposure to conditions that require Hog1-mediated stress responses, underpins the flexibility needed to allow adaptation of C. albicans to the range of environments encountered within the host [48]. Interestingly, long-term SAPK activation can be tolerated in other organisms. For example, in some human cell types sustained SAPK activation promotes either cell survival [49] or cellular differentiation [50], rather than apoptosis. However, whether similar mechanisms are evoked in humans and C. albicans, to adapt to sustained SAPK activation is unknown.
The main findings reported here in this paper focused on cells in which HOG1 is expressed from its native chromosomal locus. However, during the course of this work we noted that the cellular response of C. albicans to YPD1 loss differed depending on whether the HOG1 gene was present at its native locus, or re-integrated at the RPS10 locus in hog1Δ ypd1Δ+HOG1 cells. In both instances, cells adapted to long term SAPK activation by reducing the levels of phosphorylated Hog1 (S5A and S5B Fig). However, when HOG1 was reintegrated at the RPS10 locus, this was also accompanied by a reduction in HOG1 mRNA levels (S5C Fig). Consistent with the decrease in Hog1 levels, hog1Δ ypd1Δ+HOG1 cells changed dramatically over time to phenocopy hog1Δ cells (S5D Fig). These results contrast significantly with those from ypd1Δ cells when HOG1 is expressed from its normal locus where the cells change over time to phenocopy wild type cells (compare Fig 5 and S5 Fig). Indeed, the ypd1Δ-dependent down-regulation of HOG1 located at the RPS10 locus was independent of HOG1 promoter and terminator sequences (S5E Fig), suggesting that this is a genome position-effect rather than a gene-specific phenomenon. Insertion of genes at the RPS10 locus has been employed by numerous labs to successfully generate re-integrant strains [51, 52]. Significantly, the present study has highlighted the importance of studying a gene at its native chromosomal locus when investigating the role and/or regulation of the gene.
Two-component proteins represent attractive antifungal targets as these signal transduction proteins are absent in metazoans [22, 44, 45]. Indeed, deletion of any one of the three C. albicans histidine kinases, or the Ssk1 response regulator, attenuate virulence in mouse systemic infection models. Furthermore, the Ypd1 phosphorelay protein has generated significant interest as an antifungal target, due to the lethality associated with its deletion in both S. cerevisiae and the human fungal pathogen C. neoformans [20, 21]. However, in this study we report that inhibition of YPD1 expression during infection actually increases the virulence of C. albicans in two distinct infection models; a C.elegans pathogenesis model [40] and a murine model of systemic candidiasis [41, 42]. Histology images from murine kidneys show that, similar to that seen in vitro, repression of YPD1 during infection results in clusters of highly filamentous cells. Enhanced filamentation in vivo has previously been shown to increase virulence [39], and thus the increased virulence associated with YPD1 loss in vivo may be due to the formation of clusters of hyper-filamentous cells. Moreover, we show that drug mediated inhibition of YPD1 expression increases the efficacy by which C. albicans can kill macrophages. This is likely related to the observation that Ypd1 loss also results in increased filamentation within this infection model, as hyphae formation within the macrophage promotes C. albicans-mediated killing of macrophages by triggering pyroptosis [53], and by mechanically rupturing the macrophage cell membrane [54].
The molecular basis underlying the increased virulence seen upon repressing YPD1 expression is unknown, but given that Hog1 is essential for C. albicans virulence in several infection models [8–11], it is possible that the concomitant increases in Hog1 activity as a consequence of Ypd1 loss promotes C. albicans survival in the host (Fig 8). It is important to note here that the filamentous phenotype exhibited by cells lacking Ypd1 is due to the concomitant sustained Hog1 activation. To explore whether the enhanced virulence seen upon Ypd1 loss was dependent on Hog1, we compared the virulence of wild-type, hog1Δ and hog1Δ ypd1Δ C. albicans cells in the three day murine infection model described above. Although it was not feasible to analyse ypd1Δ cells in this model due to the highly flocculent nature of this strain, hog1Δ and hog1Δ ypd1Δ cells demonstrated equally impaired virulence in the three-day murine systemic infection model (S1 Table). Thus the impaired virulence exhibited by hog1Δ cells is not improved upon deletion of YPD1 and, therefore, is consistent with the hypothesis that the enhanced virulence observed upon loss of YPD1 may be dependent on the concurrent increased activation of the Hog1 SAPK to promote fungal stress resistance and/or filamentation.
Importantly, the data presented here illustrating that Ypd1 loss potentiates C. albicans virulence, indicates that antifungals administered to treat C. albicans infections that target Ypd1 have the potential to actually enhance the virulence of this major pathogen. It is also noteworthy that SAPKs are important for the virulence of other human and plant fungal pathogens [55]. Consequently, any compound capable of inhibiting two-component signalling that leads to SAPK activation, may promote the survival of other fungal pathogens when encountering host-imposed stresses. Thus, in conclusion, our findings question the validity of the Ypd1 protein as a broad-spectrum antifungal target. Indeed, the significant differences in stress-signalling outputs between S. cerevisiae and C. albicans, underscores the importance of directly studying predicted ‘essential’ genes in pathogenic fungi, rather than in model yeast.
All the C. albicans strains used in this study are listed in Table 1. Cells were grown at 30°C in YPD rich medium [56]. Addition of 20μg/ml doxycycline was used to repress expression from the tetO promoter.
The strains used in this study are listed in Table 1, and oligonucleotides used for their construction are listed in S2 Table.
Glycerol concentrations were determined using The Free Glycerol Determination Kit (Sigma-Aldrich), following the manufacturer’s instructions. Three independent biological replicates were performed.
Protein extracts were prepared from mid-exponential phase cells and phosphorylated Hog1 was detected by western blot analysis with an anti-phospho-p38 antibody (New England Biolabs) as described previously [6]. Blots were stripped and total levels of Hog1 were determined by probing with an anti-Hog1 antibody (Santa Cruz Biotechnology), and in some cases protein loading determined using an anti-tubulin antibody (DSHB, University of Iowa).
RNA preparation and Northern blot analyses were performed as described previously [6]. Gene-specific probes were amplified by PCR from genomic DNA using oligonucleotide primers specific for YPD1, GPD2, RHR2, PTP2, PTP3, HOG1 and ACT1 (S1 Table).
C. albicans strains were grown at 30°C to exponential phase and then 10 fold serial dilutions were spotted onto YPD plates containing the indicated compounds. Plates were incubated at 30°C for 24 h.
Differential interference contrast images were captured using a Zeiss Axioscope microscope as described previously [6].
All animal experiments were conducted in compliance with United Kingdom Home Office licenses for research on animals (project license number PPL 60/4135), and were approved by the University of Aberdeen Animal Welfare and Ethical Review Body (AWERB). Animal experiments were minimised, and all animal experimentation was performed using approaches that minimised animal suffering and maximised our concordance with the 3Rs.
|
10.1371/journal.pntd.0002494 | Natural Terpenoids from Ambrosia Species Are Active In Vitro and In Vivo against Human Pathogenic Trypanosomatids | Among the natural compounds, terpenoids play an important role in the drug discovery process for tropical diseases. The aim of the present work was to isolate antiprotozoal compounds from Ambrosia elatior and A. scabra. The sesquiterpene lactone (STL) cumanin was isolated from A. elatior whereas two other STLs, psilostachyin and cordilin, and one sterol glycoside, daucosterol, were isolated from A. scabra. Cumanin and cordilin were active against Trypanosoma cruzi epimastigotes showing 50% inhibition concentrations (IC50) values of 12 µM and 26 µM, respectively. Moreover, these compounds are active against bloodstrean trypomastigotes, regardless of the T. cruzi strain tested. Psilostachyin and cumanin were also active against amastigote forms with IC50 values of 21 µM and 8 µM, respectively. By contrast, daucosterol showed moderate activity on epimastigotes and trypomastigotes and was inactive against amastigote forms. We also found that cumanin and psilostachyin exhibited an additive effect in their trypanocidal activity when these two drugs were tested together. Cumanin has leishmanicidal activity with growth inhibition values greater than 80% at a concentration of 5 µg/ml (19 µM), against both L. braziliensis and L. amazonensis promastigotes. In an in vivo model of T. cruzi infection, cumanin was more active than benznidazole, producing an 8-fold reduction in parasitemia levels during the acute phase of the infection compared with the control group, and more importantly, a reduction in mortality with 66% of the animals surviving, in comparison with 100% mortality in the control group. Cumanin also showed nontoxic effects at the doses assayed in vivo, as determined using markers of hepatic damage.
| In addition to the primary metabolism necessary for life, plants have a secondary metabolism that generates compounds, which aid in their growth and development. A common role of secondary metabolites is defense mechanisms to fight off animals, pest and pathogens. Pharmacognosy takes advantage of the rich source of compounds produced by plants, selecting and processing natural products for medicinal use, resulting in a wide range of anticancer, anti-inflammatory and anti-infective drugs currently in use. Chagas disease and leishmaniasis are parasitic diseases caused by protozoa transmitted by sucking insects. According to the World Health Organization (WHO) they are considered as Neglected Tropical Diseases that are especially endemic in low-income populations in developing countries. The drugs currently in use were developed more than 50 years ago and have severe drawbacks. The authors found that chemical compounds called sesquiterpene lactones, obtained from plants of the Asteraceae family, have anti-protozoan activity against different parasite stages. The identified compounds were also able to control parasitic infection in mice without any toxic effects. Thus, these sesquiterpenes lactones could be interesting chemical compounds having pharmacological properties, whose chemical structure could be modified leading to potent and safer drugs to treat these parasitoses.
| Natural products have been a major source of drugs mainly for treating infectious diseases and cancer [1]. About 75% of anti-infective drugs approved from 1981 to 2002 are derived from natural sources. Many of them were isolated from plants and have shown antiparasitic activity. The first antimalarial drug, quinine, was isolated from Cinchona spp. and led to the development of other antimalarial drugs such as chloroquine, which is currently in use. More recently, the sesquiterpene lactone (STL) artemisinin has been isolated from the Chinese plant Artemisia annua, which has been used for over 2000 years to treat malaria. At present, this natural compound and its derivatives are used for treating chloroquine-resistant malaria.
Both Chagas disease and leishmaniasis are protozoan diseases that cause significant morbidity and mortality in Latin America, whereas leishmaniasis also worldwide. According to the World Health Organization (WHO), they are considered, among others, Neglected Tropical Diseases (NTDs) mainly affecting poor people in developing countries [2]. These parasitoses are often forgotten by governments and the pharmaceutical industry, due to economic reasons and a relatively limited market. Therefore, the development of new drugs for the treatment of these parasitic diseases remains a highly desirable goal.
Chagas' disease or American trypanosomiasis is caused by the protozoan parasite Trypanosoma cruzi, which is transmitted by blood-sucking insects. This parasitic disease affects 8 million people, mostly in endemic areas of Latin America, but has now spread to other continents [3]. Up to 30% of patients develop heart failure and people usually die from sudden death caused by arrhythmias. The chronic phase of the disease can cause damage to the esophagus, colon or the autonomic nervous system in more than 10% of patients. It is estimated that Chagas' disease killed more than 10,000 people in 2008 [3]. In Argentina, 1.5 to 2 million people are affected and it is estimated that 15 people die of this parasitosis every week [4].
Leishmaniasis is caused by kinetoplastids from the genus Leishmania and is transmitted by sand flies. WHO estimates that almost 12 million people worldwide are infected with Leishmania spp. and 350 million are at risk of contracting this parasitic disease [5]. Cutaneous leishmaniasis is endemic in Northern Argentina covering an area of 500,000 km2 where it is common to find people coinfected with T. cruzi and Leishmania sp. [6]. Its incidence has increased during the last two decades mainly due to Leishmania braziliensis [7]. The clinical manifestations of these parasitoses depend on the Leishmania species involved, presenting three different clinical forms: cutaneous, mucocutaneous and visceral leishmaniasis.
Most of the drugs currently in use for the treatment of American trypanosomiasis and leishmaniasis have severe drawbacks. The available treatments for Chagas disease are limited to the nitroaromatic compounds, benznidazole and nifurtimox, which were released in the 1970s. Even though these two drugs are active in the acute stage of infection, they are ineffective in the treatment of the chronic phase. They have toxic side effects and are not active against all T. cruzi strains [8]. The current antileishmanial therapy includes the use of pentavalent antimonials, amphotericin B, miltefosine or paromomycin, which have disadvantages in terms of the route of administration, parasite resistance, cost, teratogenic effects, length of treatment and toxicity [9]. Among the natural compounds, terpenoids display a wide range of biological activities such as anticancer and anti-inflammatory actions and are effective against infective agents such as viruses, bacteria and parasites [10], [11]. Several terpenoids have been reported to have trypanocidal and leishmanicidal activities [12]–[15]. Recent research has shown the potential role of terpenoids as promising compounds against neglected protozoan diseases such as Chagas' disease and leishmaniasis [16], [17]. An exhaustive and updated revision has been recently performed by Schmidt and coworkers [18]. These authors revised the antiprotozoal activities of the major biogenetic subclassses of terpenes focusing on STLs, diterpenes and triterpenes. Within the STLs, germacranolides, guaianolides, xanthanolides and pseudoguaianolides, have exhibited significant antiprotozoal activity, showing the potential of this class of compounds [18].
In the last decades, Asteraceae has been regarded as a promising family of plants because of the amount and variety of active compounds produced by the secondary metabolism. STLs and triterpenes isolated from this family have been reported as having trypanocidal and leishmanicidal activities [16]–[18], [19]. In Argentina, the genus Ambrosia that belongs to this family is represented by three species: A. tenuifolia, A. scabra and A. elatior. Considering the limitations of current therapies for Chagas disease and leishmaniasis and our previous promising findings on the antiprotozoal activity of terpenoids from the genus Ambrosia [20]–[23], the aim of the present work was to isolate further bioactive compounds from A. elatior and A. scabra.
The aerial parts of Ambrosia elatior L. (BAF 707) and Ambrosia scabra Hook. & Arn. (Asteraceae) (BAF 711) were collected in Buenos Aires Province, Argentina in May 2009. The botanical identification was performed by Dr. Gustavo Giberti and a voucher specimen of each species was deposited at the Museo de Farmacobotánica, Facultad de Farmacia y Bioquímica, Universidad de Buenos Aires.
Trypanosoma cruzi epimastigotes (RA and K98 strains) were grown in a biphasic medium. Cultures were routinely maintained by weekly passages at 28°C. T. cruzi bloodstream trypomastigotes from RA and K98 strains [24] were obtained from infected CF1 mice by cardiac puncture at the peak of parasitemia on day 15 postinfection. Trypomastigotes were routinely maintained by infecting 21-day-old CF1 mice. T. cruzi amastigotes were obtained from cultured cells infected with tripomastigotes. Leishmania braziliensis and Leishmania amazonensis promastigotes (MHOM/BR/75/M2903 and MHOM/BR/75/M2269 strains, respectively) were grown in liver infusion tryptose medium (LIT). Cultures were routinely maintained by weekly passages at 26°C. Parasites were passaged 24 or 48 h previous to the experiments.
Inbred male C3H/HeN mice were nursed at the Microbiology Department, Faculty of Medicine, University of Buenos Aires. All procedures requiring animals were performed in agreement with institutional guidelines and were approved by the Review Board of Ethics of IDEHU, CONICET, and conducted in accordance with the Guide for the Care and Use of Laboratory Animals of the National Research Council of Argentina [25].
Extraction of the aerial parts of A. elatior and A. scabra was done by maceration with dichloromethane∶methanol (1∶1), as previously described [22].
The organic extract of A. elatior (AE-OE) was fractionated by column chromatography on Silica gel 60 with a gradient of hexane, ethyl acetate and methanol. Nine fractions (F1AE–F9AE) of 500 ml each were collected and taken to dryness. Each fraction was tested for trypanocidal activity on T. cruzi epimastigotes. Fractions F5AE, F6AE and F7AE were taken with ethyl acetate and afforded a crystalline compound named compound A, which was assayed for trypanocidal and leishmanicidal activities.
Fractionation of organic extract of A. scabra has been previously described [23]. A pure compound, compound B, was obtained from fractions F5AS(70–74) by crystallization from ethyl acetate. Compound C was obtained as a white crystalline precipitate from fractions F5AS(75–77), while fractions F5AS(125–135) afforded a white amorphous powder (compound D).
The structure elucidation of compounds A–D was performed by proton nuclear magnetic resonance (1H NMR) and carbon NMR (13C NMR) (Inova NMR spectrometer; Varian, Palo Alto, CA) 500 MHz in CDCl3 (for compounds A, B and C) and CDCl3:CD3OD (8∶2) (for compound D), heteronuclear single quantum correlation (HSQC); heteronuclear multiple bond correlation (HMBC); correlated spectroscopy (COSY); electron impact-mass spectrometry (EI-MS) (Agilent 5973) and infrared spectroscopy (Bruker FT-IR IFS25).
Growth inhibition of T. cruzi epimastigotes and Leishmania spp. promastigotes was evaluated by a [3H] thymidine uptake assay as previously described [20]. Parasites were adjusted to a cell density of 1.5×106/ml and cultured in the presence of A. elatior organic extract, the fractions and purified compounds for 72 h at final concentrations ranging from 1 to 100 µg/ml. Benznidazole (5 to 20 µM; Roche) and Amphotericin B (0.27–1.6 µM; ICN) were used as positive controls (data not shown). The percentage of inhibition was calculated as 100−{[(cpm of treated parasites)/(cpm of untreated parasites)]×100}.
The trypanocidal effect of the pure compounds was also tested on bloodstream trypomastigotes of RA and K98 strains as previously described [20]. Briefly, mouse blood containing trypomastigotes was diluted in complete liver infusion tryptose medium to adjust the parasite concentration to 1.5×106/ml. Parasites were seeded (150 µl/well) by duplicate into a 96-well microplate, and 2 µl of each compound/well at different concentrations or control drug (benznidazole) was added per well. Plates were incubated for 24 h and the remaining live parasites were counted on a hemocytometer. Results are expressed as [live parasites in wells after compound treatment/live parasites in untreated wells]×100.
To evaluate the effect of the compounds on intracellular forms of T. cruzi, 96-well plates were seeded with nonphagocytic Vero cells at 5×103 per well in 100 µL of culture medium and were incubated for 24 h at 37°C in a 5% CO2 atmosphere. Cells were washed and infected with transfected blood trypomastigotes expressing β-galactosidase [26] at a parasite/cell ratio of 10∶1. After 24 h of coculture, plates were washed twice with PBS to remove unbound parasites and each pure compound was added at different concentrations in 150 µl of fresh complete RPMI medium without phenol red (Gibco, Rockville, MD). Controls included infected untreated cells (100% infection control) and uninfected cells (0% infection control). The assay was developed by the addition of chlorophenolred-β-D-galactopyranoside (CPRG) (100 µM) and 1% Nonidet P40, 5 days later. Plates were then incubated for 4–6 h at 37°C and the absorbance was measured at 595 nm in a microplate reader (Bio-Rad Laboratories, Hercules, CA). Percentage inhibition was calculated as 100–{[(absorbance of treated infected cells)/(absorbance of untreated infected cells)×100} and the IC50value was estimated.
Compounds A and B were combined with each other, as well as compound A and benznidazole, to evaluate a potential interaction among them. Trypanocidal activity was evaluated on RA epimastigotes by a [3H] thymidine uptake assay as previously described [20]. The fractional inhibitory concentrations (FICs) were calculated as the ratio of the IC50 of one compound in combination and the IC50 of the compound alone. The FIC index (FICI) for the two compounds was the FIC of compound A plus the FIC of compound B. The fractional inhibitory concentration index (FICI) was interpreted as follows: FICI≤0.5 synergy, FICI>4.0 antagonism, FICI = 0.5–4 addition [27].
Vero cells were assayed to determine viability by the MTT method [22]. Cells (5×105) were settled at a final volume of 150 µl in a flat-bottom 96-well microplate and were cultured at 37°C in a 5% CO2 atmosphere in the absence or presence of increasing concentrations of the pure compounds. After 24 h, 3-(4,5-dimethylthiazol-2yl)-2,5-diphenyltetrazolium bromide (MTT) was added at a final concentration of 1.5 mg/ml. Plates were incubated for 2 h at 37°C. The purple formazan crystals were completely dissolved by adding 150 µl of ethanol and the absorbance was detected at 595 nm in a microplate reader. Results were calculated as the ratio between the optical density in the presence and absence of the compound multiplied by 100.
Groups of five C3H/HeN male mice (6 to 8 weeks old; 27.2±0.9 g) were infected with 500 bloodstream T. cruzi trypomastigotes by the intraperitoneal route. Five days after infection, the presence of circulating parasites was confirmed by the microhematocrit method [23]. Mice were treated daily with compound A or benznidazole (1 mg/kg of body weight/day) for five consecutive days (days 5 to 10 postinfection) by the intraperitoneal route. Drugs were resuspended in 0.1 M phosphate buffered saline (PBS, pH 7.2), and this vehicle was also employed as a negative control. Levels of parasitemia were monitored every 2 days in 5 µl of blood diluted 1∶5 in lysis buffer (0.75% NH4Cl, 0.2% Tris, pH 7.2) by counting parasites in a Neubauer chamber. The number of deaths was recorded daily.
Groups of five C3H/HeN male uninfected mice were treated with PBS or compound A as described above, in order to evaluate the potential in vivo toxicity of the compound. On day 7 post treatment, blood samples were collected by cardiac puncture. Serum activities of alanine aminotransferase (ALT), aspartate aminotransferase (AST) and lactate dehydrogenase (LDH) were determined as markers of hepatic damage. Assays were carried out by ultraviolet spectrophotometry following the specifications of the kit's manufacturer (Wiener Lab, Buenos Aires, Argentina).
The results are presented as mean±SEM. The level of statistical significance was determined by using one-way analysis of variance (ANOVA), with GraphPad Prism 5.0 software (GraphPad Software Inc., San Diego, CA). Long rank test was used for survival curves. Comparisons were referred to the control group. P values of <0.05 were considered significant.
The dichloromethane∶methanol organic extract (AE-OE) from the aerial parts of Ambrosia elatior was evaluated in vitro against T. cruzi epimastigotes. This extract was active with a growth inhibition of 93.7±2.0% at a concentration of 100 µg/ml. The extract fractionation by chromatographic techniques using Silica gel 60 with a gradient of hexane, ethyl acetate and metanol yielded nine fractions (F1AE to F9AE), which were assayed for their in vitro trypanocidal activity on epimastigotes at 100, 10 and 1 µg/ml (Table 1). The results of this experiment showed that at a concentration of 10 µg/ml, fractions F5AE, F6AE and F7AE were the most active with percentages of growth inhibition of 94.7±0.5%, 75.3±4.8% and 76.8±1.4%, respectively (Table 1).
A pure compound (compound A) was further isolated from fractions F5AE, F6AE and F7AE. It was identified by spectroscopic methods as the STL 1α,3β,4β,5β,8β,10β-3,4-dihydroxy-11(13)-pseudoguaien-12,8-olide or cumanin (Figure 1).
An exhaustive search in the active fraction F5AS from A. scabra resulted in the isolation of three other compounds, B, C and D. The analyses of spectroscopic data allowed to identify compounds B and C as the STLs (2′R,3aS,6s,8s,8aR)-octahydro-8-hydroxy-6,8-dimethyl-3-methylene-spiro[7H-cyclohepta[b]furan-7,2′(5′H)-furan]-2,5′(3H)-dione psilostachyin and its 5-epimer cordilin or epipsilostachyin, respectively. Compound D was identified as the sterol glycoside sitosterol-3-O-β-D-glucopyranoside or daucosterol (Figure 1).
We analyzed the antiparasitic activity of the isolated compounds at different stages of Trypanosoma cruzi and Leishmania sp. The results of the trypanocidal activity of cumanin, cordilin and daucosterol on T. cruzi epimastigote forms are shown in Figure 2. The STLs were active with 50% inhibition concentration (IC50) values of 12 µM and 4 µM for cumanin, and 26 µM and 44 µM for cordilin against RA and K98 epimastigotes, respectively. Daucosterol exhibited lower activity with a growth inhibition of 40.7±2.8% at 100 µg/ml (174 µM).
The combinatory effect of the two most active STLs, psilostachyin and cumanin, on epimastigotes of T. cruzi was analyzed by an isobologram. The fractional inhibitory concentration index (FICI) was 0.97±0.11 (Mean±SD) for the combination of these drugs, indicating there is neither antagonism nor synergism; however, an additive effect could be assessed in the trypanocidal activity between cumanin and psilostachyin (Figure 3). We neither found antagonism nor synergism interaction between cumanin and benznidazole, the current reference drug to treat Chagas disease (FICI: 0.95±0.05) (Mean±SD) (Figure 3).
The trypanocidal effect of the pure compounds was tested on bloodstream trypomastigotes obtained from infected mice. Parasites were seeded into a 96-well microplate and different concentrations of each compound were added to each well. After 24 h incubation, the remaining live parasites were counted on a hemocytometer. In Figure 4 it can be appreciated that the two STLs, cumanin and cordilin, showed activity against RA trypomastigotes with IC50 values of 180 µM and 90 µM, respectively. The sterol glycoside was active with an IC50 of 142 µM.
In order to analyze the relevance of the compounds in different T. cruzi populations, K98 trypomastigotes were incubated in the presence of cumanin, cordilin and daucosterol, showing IC50 values of 170 µM, 83 µM and 184 µM, respectively (Figure 4).
To analyze the effect of the compounds on the replicative forms, Vero cells were infected with transfected trypomastigotes expressing the β-galactosidase gene. After 24 h incubation, all the parasites outside the cells were removed by washing and different concentrations of the compounds were added to the wells. Five days later the cells were disrupted with detergent and β-galactosidase activity was determined with chlorophenol red-β-D-galactopyranoside, as a direct estimation of the number of parasites. Since the effect of psilostachyin was not previously reported, the four isolated compounds were included in the assay. Figure 5 shows that cumanin and psilostachyin were able to inhibit amastigote replication. Both STLs were active with approximately 95% inhibition at 25 µg/ml, and IC50s of 8 µM and 21 µM, respectively. Neither cordilin nor daucosterol were active against the intracellular forms of T. cruzi.
After the observed effect of cumanin, psilostachyin, cordilin and daucosterol on T. cruzi, we analyzed the inhibitory activity of these compounds on Leishmania promastigotes. The IC50 values against L. amazonensis were 3 µM, 10 µM and 55 µM, for cumanin, psilostachyin and cordilin, respectively. We found that cumanin revealed leishmanicidal activity with growth inhibition values greater than 80% at a concentration of 5 µg/ml (19 µM), against both L. braziliensis and L. amazonensis (Figure 6). Psilostachyin and cordilin displayed high leishmanicidal activity against L. braziliensis (82.8±7.8 and 83.03±0.01% at a concentration of 1 µg/ml, respectively).
These results showed that cumanin was active against the two species reported as the main cause of tegumentary leishmaniosis in Argentina: L. braziliensis and L.amazonensis [6].
To analyze the effect of the active compounds in vivo, a group of mice was inoculated with a deadly number of trypomastigotes of the RA strain and parasitemia was determined 5 days after to confirm the effectiveness of the infection. Mice were then treated with cumanin or benznidazole, the drug currently used to treat infected humans, for 5 consecutive days. Parasitemia and survival were periodically recorded. Figure 7A shows how cumanin was able to dramatically reduce the number of circulating parasites in the acute phase of T. cruzi infection compared with the PBS-treated control. When we analyzed the area under the parasitemia curve (AUC) we observed an important decrease in the number of circulating parasites in cumanin treated mice comparing with controls (AUC: 382 and 800, respectively). This protection was more important at the peak of parasitemia, on day 22 postinfection, when control mice showed 8 times more parasites than cumanin-treated ones (8.2 and 1.2×107 parasites/ml, respectively, p<0.01). As shown in Figure 7A, parasitemia values were slightly higher than those obtained with benznidazole but no significant differences were found between the two compounds.
The ability of cumanin to control infection was also reflected in the significant survival of the treated mice, respect to the control (p = 0.0086). Figure 7B shows the important reduction in mice mortality observed in cumanin-treated mice, where 100% of PBS-treated mice died between days 22 and 26 postinfection, while 66% of cumanin-treated mice survived the T. cruzi infection by the end of the experiment on day 100 (not shown). Again, cumanin-treated mice did not show any significant differences with respect to those treated with the approved drug for this parasitosis.
Hepatic toxicity of the STL was evaluated through the determination of a panel of hepatic-linked enzyme markers. Serum levels of AST, LDH and AST were measured at 7 days post treatment. Table 2 shows that cumanin treated mice exhibited similar levels of the analyzed enzymes to those of PBS-treated control mice, suggesting that cumanin is not hepatotoxic in vivo at the doses used.
In a previous investigation we have reported the isolation of STLs from Argentinean Ambrosia species, some of which have shown significant trypanocidal and leishmanicidal activities [20]–[22]. These promising results prompted us to continue with the search for other antiprotozoal compounds from Ambrosia scabra and its related species, A. elatior.
The organic extract of A. elatior showed significant activity against T. cruzi, being able to inhibit 94% of epimastigote growth at a concentration of 100 µg/ml. When this extract was chromatographed on a Silica gel column, 9 final fractions (F1AE–F9AE) were obtained. Fractions F5AE, F6AE and F7AE displayed the highest trypanocidal activity against the non-infective form of T. cruzi and afforded compound A. The structural elucidation of this substance was based on the analysis of its spectral data and was identified as cumanin, which has been previously reported in Ambrosia artemisiifolia [28] and Ambrosia cumanensis [29].
We recently found that the F5AS fraction of the organic extract of A. scabra was very active against T. cruzi. From F5AS fraction we isolated the STL psilostachyin C, whose anti-epimastigote and anti-trypomastigote activity had been previously reported [22]. In this manuscript we report that further isolation steps from A. scabra fraction F5AS yielded other three compounds (B, C, D). Compounds B and C were identified as the STLs, psilostachyin and cordilin, respectively, and compound D as daucosterol. The identification of all these compounds was performed by analysis of their spectral data. Psilostachyin had been previously isolated from A. tenuifolia by our group [20]; however, this is the first report of its presence in A. scabra, and the first time the isolation of cordilin and daucosterol from A. scabra has been reported. Psilostachyin and cordilin are diasteroisomers that only differ in the spatial configuration of the hydroxyl and methyl groups on C-5 (Figure 1).
T. cruzi is genetically highly diverse [30]. While RA, belonging to TcII-DTUs, is a highly virulent pantropic/reticulotropic strain, K98 (TcI) is a low virulence myotropic strain [31]. In order to analyze the relevance of cumanin, cordilin and daucosterol in different T. cruzi populations, K98 and RA strains were evaluated for their trypanocidal activity. Even though the epimastigote stage of the parasite is non-infective, it is easy to cultivate and therefore, it is useful for a preliminary screening test [32]. Cumanin exerted significant in vitro trypanocidal activity against this parasite form, showing an IC50 value 12 µM and 4 µM for RA and K98 strains. Cordilin was less active with an IC50 26 µM and 44 µM, for the two strain analyzed. Daucosterol displayed even lower trypanocidal activity. However, in the search for new potential trypanocidal candidates, it is necessary to analyze the activity of the compounds on the infective forms of the parasite. Consequently, cumanin, cordilin and daucosterol were tested against trypomastigotes and all the isolated compounds were evaluated against intracellular amastigotes, including psilostachyin that had not been previously evaluated against this parasite form [20]. Cordilin and cumanin showed low activity on trypomastigotes (IC50 = 89 µM and 180 µM, respectively. The two STLs psilostachyin and cumanin inhibited the growth of the intracellular forms of T. cruzi with IC50s values of 21 µM and 8 µM, respectively. Cordilin and daucosterol did not display any activity against amastigotes.
The fact that cumanin is active against amastigotes is of particular interest, since the DNDi organization prioritizes the development of drugs that are useful during the indeterminate and chronic phases of the infection where parasites remain intracellular [33]. The ability of compounds to inhibit the intracellular growth of T. cruzi amastigotes is a more rigorous and relevant test of anti-T. cruzi activity, as it is applied to a stage which is the predominant and replicative form in mammals cells. The impairment of amastigotes replication upon cumanin treatment could lead to a reduction in tripomastigotes release from the cells and the subsequent low parasitemia observed in mice.
Nowadays, the association of compounds could be an interesting strategy for the control of Chagas disease. Thus, psilostachyin and cumanin were tested together on epimastigotes to assess their possible interaction. An additive effect was observed between these two compounds.
Considering the incidence of the Leishmania spp. in South America and the existence of patients co-infected with these parasites and T. cruzi [34]–[36], the effect of the four compounds on L. braziliensis and L. amazonensis promastigotes was also evaluated. Cumanin, psilostachyin and cordilin revealed significant leishmanicidal activiy against both L. amazonenzis and L. brazilensis, while daucosterol leishmanicidal activity was lower.
Trypanocidal activity of STL is highly influenced by stereochemical or structural differences [37]. Schmidt et al. [38] demonstrated that the STL helenalin was more active against T. cruzi and T. b. rhodesiense than its diasteroisomer mexicanin, which differs from the former only in the spatial orientation of an OH group. In the case of the two epimeric STLs isolated from A. scabra, cordilin and psilostachyin, the behavior of the two compounds against non-infective and infective forms of T. cruzi was very different. We have previously reported that psilostachyin is very active against epimastigotes and trypomastigotes showing IC50 values of 4 µM and 3 µM, respectively [20]. Those values compared with the results herein reported for cordilin with values of 26 µM (epimastigotes) and 89 µM (trypomastigotes) signal that psilostachyin is more active than cordilin. In the present study, we have also demonstrated that psilostachyin is more active than its 5-epimer on the intracellular form of the parasite. Thus, these results support the fact that stereochemistry plays an important role in the biological activity.
The administration of cumanin to T. cruzi-infected mice (1 mg/kg/day) produced a significant reduction in parasitemia levels, even lower than that produced by benznidazole, when compared with PBS-treated control mice. An 8-fold reduction in parasitemia levels was observed on day 22 postinfection compared with control. Moreover, cumanin was able to reduce the number of parasites during the whole course of the infection. More importantly, 66% of the mice survived the deadly challenge with trypomastigotes while only 20% of benznidazole-treated mice survived, compared with 100% mortality in the PBS-treated mice group. It should be noted that the dosis of 1 mg/kg/day used was irrespective of cumanin and benznidazole IC50 values.
Interestingly, cumanin is more active on intracellular amastigotes than on free trypomastigotes, which is not the case for psylostachyin, since it is very active in both stages (Figure 5 and reference [20]). This phenomenon could be attributed to the fact that cumanin differently affects both parasite stages somehow interacting in different metabolic pathways that deserve to be further investigated. An alternative hypothesis suggesting that cumanin affects both parasite stages similarly as well as the cells infected by amastigotes, consequently killing the cell host and parasites, cannot be sustained since cumanin showed to be nontoxic for non-infected cells at the concentration used (Table 2).
The determination of in vitro and in vivo toxicity is a very important point in a drug discovery process [17]. Cumanin and cordilin displayed low cytotoxicity on Vero cell line at concentrations of up to 100 µg/ml (data not shown). Moreover, cumanin-treated mice (non-infected) exhibited similar serum levels of hepatic-linked enzyme markers than those in the PBS-treated control mice (Table 2). Thus, by using in vitro and in vivo toxicity assays, we demonstrated the non-toxic effect of cumanin at the doses used. These results are in agreement with those reported by Lastra et al. [39], who found that cumanin produces low cytotoxicity on peritoneal murine macrophages.
As it was mentioned in the introduction section, different types of terpenoid compounds have shown to display antiprotozoal activity [18], [40], and among them, several STLs having an α,β-unsaturated-γ-lactone moiety have been reported to show trypanocidal and leishmanicidal activities [18], [41], [42]. Thus, the STLs psilostachyin and cumanin can be considered interesting lead molecules for the development of drugs for Chagas' disease. Variability in the outcome and morbidity of T. cruzi infection might be associated, at least in part, with the complex population structure of the parasite. Taking this into account, the fact that cumanin shows high activity at different stages and strains of T. cruzi highlights the importance of searching new drugs against Chagas disease. In addition, the leishmanicidal activity shown by these compounds in a preliminary assay could be an interesting fact to consider, since the endemic areas of Chagas disease and leishmaniasis usually overlap in Latin America and co-infected patients have been reported [34]–[36].
Despite the advances in the biology of protozoan parasites like T. cruzi and Leishmania sp., the discovery and development of safer drugs to treat American trypanosomiasis and/or leishmaniasis still constitute a challenge. The results presented in this work demonstrate that terpenoids, particularly STLs, are an interesting group of natural compounds that could become good candidates for antiprotozoal chemotherapy. Further studies involving the evaluation of different targets will be useful to understand the mechanisms of action of the isolated STLs.
|
10.1371/journal.pntd.0001573 | Extensive Genetic Diversity, Unique Population Structure and Evidence of Genetic Exchange in the Sexually Transmitted Parasite Trichomonas vaginalis | Trichomonas vaginalis is the causative agent of human trichomoniasis, the most common non-viral sexually transmitted infection world-wide. Despite its prevalence, little is known about the genetic diversity and population structure of this haploid parasite due to the lack of appropriate tools. The development of a panel of microsatellite makers and SNPs from mining the parasite's genome sequence has paved the way to a global analysis of the genetic structure of the pathogen and association with clinical phenotypes.
Here we utilize a panel of T. vaginalis-specific genetic markers to genotype 235 isolates from Mexico, Chile, India, Australia, Papua New Guinea, Italy, Africa and the United States, including 19 clinical isolates recently collected from 270 women attending New York City sexually transmitted disease clinics. Using population genetic analysis, we show that T. vaginalis is a genetically diverse parasite with a unique population structure consisting of two types present in equal proportions world-wide. Parasites belonging to the two types (type 1 and type 2) differ significantly in the rate at which they harbor the T. vaginalis virus, a dsRNA virus implicated in parasite pathogenesis, and in their sensitivity to the widely-used drug, metronidazole. We also uncover evidence of genetic exchange, indicating a sexual life-cycle of the parasite despite an absence of morphologically-distinct sexual stages.
Our study represents the first robust and comprehensive evaluation of global T. vaginalis genetic diversity and population structure. Our identification of a unique two-type structure, and the clinically relevant phenotypes associated with them, provides a new dimension for understanding T. vaginalis pathogenesis. In addition, our demonstration of the possibility of genetic exchange in the parasite has important implications for genetic research and control of the disease.
| The human parasite Trichomonas vaginalis causes trichomoniasis, the world's most common non-viral sexually transmitted infection. Research on T. vaginalis genetic diversity has been limited by a lack of appropriate genotyping tools. To address this problem, we recently published a panel of T. vaginalis-specific genetic markers; here we use these markers to genotype isolates collected from ten regions around the globe. We detect high levels of genetic diversity, infer a two-type population structure, identify clinically relevant differences between the two types, and uncover evidence of genetic exchange in what was believed to be a clonal organism. Together, these results greatly improve our understanding of the population genetics of T. vaginalis and provide insights into the possibility of genetic exchange in the parasite, with implications for the epidemiology and control of the disease. By taking into account the existence of different types and their unique characteristics, we can improve understanding of the wide range of symptoms that patients manifest and better implement appropriate drug treatment. In addition, by recognizing the possibility of genetic exchange, we are more equipped to address the growing concern of drug resistance and the mechanisms by which it may spread within parasite populations.
| Trichomoniasis is the most common non-viral sexually transmitted infection (STI) world-wide. Of the estimated 174 million new infections each year [1] – making it more prevalent than gonorrhea and chlamydia combined – ∼7 million occur in the United States [2]. Historically, trichomoniasis has been considered a self-clearing female ‘nuisance’ disease [3], but recent studies indicate that without antimicrobial treatment women may maintain chronic infections indefinitely, while men usually resolve infection without treatment [4]–[6]. In women, symptoms include malodorous vaginal discharge, vulval irritation and inflammation, and punctate microhemorrhages on the cervix known as ‘strawberry cervix’ [7]. Males, though often asymptomatic, can present with urethritis, urethral discharge and dysuria [8]. Trichomoniasis has been associated with severe reproductive health sequelae in both sexes, including, pelvic inflammatory disease [9] and adverse pregnancy outcomes [10], [11] in women, and prostatitis, infertility, and an increased incidence of aggressive prostate cancers in men [8],[12]. Perhaps most importantly, trichomoniasis has also been implicated in increasing sexual transmission of HIV up to two-fold [13], [14]. Because of the high prevalence of trichomoniasis, this translates into a significant number of global HIV infections [15].
Trichomonas vaginalis, the causative agent of human trichomoniasis, is a highly motile, aerotolerant, haploid eukaryotic parasite that resides in the urogenital tract and has an apparently simple life-cycle consisting of a trophozoite stage that is transmitted from host to host through sexual intercourse. The parasite itself can harbor a linear, double-stranded RNA virus known as T. vaginalis virus (TVV), which has been reported in approximately 50% of isolates, and may have important implications in virulence [16]. Currently, only the 5-nitroimidazole family of drugs (specifically metronidazole and tinidazole) is approved for the treatment of trichomoniasis; however, drug resistance (documented since this family of drugs was first used to treat the infection [17]) is a major concern, with estimates of up to 10% of infections not responding to treatment in the United States [18].
Current knowledge of T. vaginalis population genetics has been limited by a lack of appropriate tools. Crude genotyping markers such as random amplified polymorphic DNA (RAPDs) and restriction fragment length polymorphisms (RFLPs), have indicated genetic variation among T. vaginalis isolates and have inconclusively detected evidence of population structure [19]–[25]. These methods, however, are highly sensitive to contaminating DNA or to slight variation in conditions, which may influence the interpretation of data collected with these techniques. To address these limitations in existing methods of genetic characterization, we recently developed the first panel of T. vaginalis-specific microsatellite and single nucleotide polymorphisms (SNPs) as robust genetic markers [26]. These markers were utilized to sample the diversity of seven laboratory strains of T. vaginalis collected during the past >55 years and propagated in vitro, in some instances for more than a year. As a result of this work, we identified a significant amount of diversity across most loci in the seven strains, although differences existed between the topology of trees inferred from different types of markers.
However, limitations of that study included (a) a focus on laboratory strains collected many years previously; (b) sampling of a limited number of geographical regions; and (c) the small number of strains characterized. To assay the global diversity and population structure of T. vaginalis, we present here analysis of 235 T. vaginalis isolates from ten world-wide regions in Mexico, Chile, India, Australia, Papua New Guinea, Italy, Africa and the United States, including 19 clinical isolates recently collected from 270 women attending New York City sexually transmitted disease (STD) clinics. We find high genetic diversity within the T. vaginalis parasite, and a two-type population structure that is distributed in near equal frequencies world-wide. In addition, we show that the two types differ in the frequency in which they harbor TVV and in their metronidazole sensitivity. Finally, we present evidence of recent intragenic recombination and speculate on the possibility of a sexual life-cycle of the parasite in the absence of obvious sexual stages. Our findings enhance the understanding of the population genetics and diversity of T. vaginalis and suggest the possibility of genetic exchange in the parasite, which has wide-ranging implications for the epidemiology and control of human trichomoniasis.
Global strains and isolates obtained from collaborators, including references that provide details of their collection, are given in Table S1. All parasites were cultured in modified Diamond's media [27], supplemented with 10% horse serum, penicillin and streptomycin (Invitrogen) and iron solution composed of ferrous ammonium sulfate and sulfosalicylic acid (Fisher Scientific). Minimum lethal concentrations (MLCs) for metronidazole were determined under aerobic conditions by the lab of origin according to previously published protocols [22], unless otherwise indicated. A subset of the samples, the Secor Lab isolates (Table S1) are isolates sent to the CDC for drug resistance testing from patients who had previously failed at least two courses of standard therapy.
Vaginal swabs were collected from 270 women attending eight New York City Department of Health and Mental Hygiene STD clinics in four of the five New York boroughs: Clinic G (N = 29) in the Bronx, Clinic D (N = 37) and Clinic F (N = 29) in Queens, Clinic A (N = 28) and Clinic E (N = 37) in Brooklyn, Clinic B (N = 36), and Clinic C (N = 41), and Clinic H (N = 33) in Manhattan. Approval for specimen collection and study was granted by the institutional review boards of NYU School of Medicine, the Centers for Disease Control and Prevention, and the New York City Department of Health and Mental Hygiene. IRB approval did not require informed consent since the study was deemed exempt. All samples were anonymized. Vaginal swabs collected during a routine pelvic examination, which would otherwise have been discarded, were used to inoculate InPouch TV culture kits [28]. Pouches were retrieved from clinics within three days of inoculation. Specimens were assigned a study ID and, after linkage to a limited set of demographic and clinical data, specimens were stripped of patient identifying information. Cultures were incubated at 37°C and examined for the presence of T. vaginalis by microscopy each day for five to seven days post-inoculation (PI) with vaginal swab. On day seven PI regardless of T. vaginalis diagnosis, 3 mL of culture was used for DNA extraction and the remaining culture media cryopreserved in 10% DMSO.
Diagnostic PCR was also used to diagnose T. vaginalis, using published methods. Two primer sets TVK3/TVK7 [29], [30] (5′-ATTGTCGAACATTGGTCTTACCCTC-3′)/(5′-TCTGTGCCGTCTTCAAGTATGC-3′) and TV16f-2/TV16r-2 [31] (5′-TGAATCAACACGGGGAAAC-3′)/(5′-ACCCTCTAAGGCTCGCAGT-3′) were used. Discrepancies between primer sets were resolved using a third primer set, BTub3f/BTUB_Bkmt [31] (5′-TCCAAAGGTTTCCGATACAGT-3′)/(5′-GTTGTGCCGGACATAATCATG-3′). All diagnostic PCRs were performed at least twice, and samples were considered positive if T. vaginalis was detected by wet mount, in vitro culture and/or by PCR with two different primers.
The UNSET buffer and phenol-chloroform extraction method was used for DNA isolation of all samples as described [26].
Isolates were genotyped at 21 T. vaginalis-specific microsatellite loci in 10 µL volumes as described [26]. Reactions were performed in duplicate and discrepancies were verified with a third reaction. GeneMapper 4.0 (ABI, Foster City, CA) was used to score MS allele sizes. All calls were manually edited to discard data from poorly amplified reactions and to ensure that proper allele calls were assigned. Mixed infections were detected by the presence of multiple alleles at two or more loci. Due to the amplification biases described by Havryliuk et al. (2008) [32] in validating the criteria for distinguishing between minor alleles and stutter peaks (i.e., minor peaks classified as >33% of the size of the major allele [33], [34]), we relied on the reproducibility of minor alleles over three independent rounds of amplification and electrophoretic analysis of labeled PCR products, and the presence of multiple alleles at two or more loci, in order to detect mixed infections. Electrophoretic readouts were also compared between samples to determine stutter patterns, and ambiguous minor alleles were ignored.
To ensure that the association of type 1 with TVV infection was not caused by interaction with the MS locus specific primers and the TVV genome, we performed a BLAST search of all known TVV genomes, including species I–IV, using each forward and reverse primer as a query. We found no more than 75% identity to any single primer, indicating that the TVV genome differs enough from the MS loci examined to prevent any unintended complementation.
Microsatellite allelic richness was estimated using ADZE [35], which implements the rarefaction method for analyzing allelic diversity across populations while correcting for sample size difference. Allelic richness estimates were graphed for each sample size (g) to estimate the sample size necessary to ensure that the majority of non-rare alleles had been detected.
Isolates were grouped according to geographical origin (or status as a laboratory strain) using the following ten categories: Laboratory (N = 5), Western United States i.e. west of the Mississippi River (N = 31), Eastern United States i.e. east of the Mississippi River (N = 51), Mexico (N = 11), Chile (N = 14), Italy (N = 12), Southern Africa (N = 19), Australia (N = 14), Papua New Guinea (PNG) (N = 30), and India (N = 1). Genetic diversity was determined by calculating expected heterozygosity (HE) at each locus, using the formula HE = [n/(n−1)][1−∑ni = 1 p2] where p is the frequency of the ith allele and n is the number of alleles sampled and confirmed with Arlequin3.11 [36]. Allelic richness (a measure of the number of alleles independent of sample size) per locus and sample (Rs) and over samples (Rt) was estimated using Mousadik and Petit's (1996) [37] method in FSTAT 2.9.3.2 [38]. FSTAT estimates the expected number of alleles in a sub-sample of 2n genes, given that 2N genes have been sampled (N≥n), where n is fixed as the smallest number of individuals typed for a locus in a sample. The estimation is performed using the formula Rs = Σni = 1[1−[[(2N−Ni)/2n]/(2N/2n)], where Ni is the number of alleles of type i among the 2N genes. For Rt, the same sub-sample size n is kept, but N becomes the overall sample number of individuals genotyped at the locus under consideration. This program was chosen because allele frequencies are weighted according to sample sizes, important in our study due to the variation in the number of samples from different geographical regions. Arlequin3.5 [36] was used to test for Fst between geographical origins.
The Bayesian clustering program STRUCTURE 2.2 was used to assign isolates to K populations according to allele frequencies at each locus [39]. The program was run 10 times each for six K values (K = 1–6) with a burn-in period of 5×105 iterations followed by 105 iterations. The number of populations was inferred by plotting the log probability of the data [Ln P(D)] for each K value, followed by clusteredness calculations. Clusteredness measures the average relatedness of the individual membership coefficients (Q) and estimates the extent to which individual infections belong to a single cluster, rather than to a combination of clusters [40]. Population differentiation was confirmed using Arlequin 3.5. Two-way hierarchical clustering and inference of a minimum spanning network (MSN) were performed to validate clustering assignments determined using STRUCTURE 2.2. Two-way hierarchical clustering was performed on MS data using JMP Genomics 5.0 (SAS), and missing data points were assigned a unique number (999) to allow for the inclusion of all samples in the analysis. MSNs were inferred from individual MS haplotypes profiles using Network 4.516 [41], software developed to reconstruct all possible least complex phylogenetic trees using a range of data types.
Loci TVAG_005070 (DNA mismatch repair homolog, postmeiotic segregation increased-1, PMS1), TVAG_302400 (MutL homolog 1a, Mlh1a), and TVAG_021420 (coronin, CRN) were PCR amplified, purified, and sequenced as described [26]. Nucleotide sequence data is available in the EMBL, GenBanks and DDBJ data bases under the accession numbers: JN380351–JN380802. Sequences were aligned to the reference sequence in GenBank and the alignments manually edited using Sequencher 4.8 (Gene Codes Corporation, Ann Arbor, MI). SNPs were manually verified and included any single nucleotide change that occurred in any single strain. All three genes were successfully sequenced in 94 isolates.
ModelGenerator v. 0.85 [42] was used to infer phylogenies from single copy gene sequences, with the number of gamma categories set at 10 to identify appropriate nucleotide substitution models for each of the loci. PhyML [43] as part of SeaView v. 4.2.4 [44] was used to infer maximum likelihood (ML) phylogenies reconstructed by applying simultaneous NNI (Nearest Neighbor Interchange) and SPR (Subtree Pruning and Regrafting) moves on five independent random starting trees. Substitution rate categories were set at ten and transition/transversion (Ts/Tv) ratios, invariable sites and across-site rate variation were selected as indicated by ModelGenerator. Support values for the tree were obtained by bootstrapping 1000 replicates. We inferred the evolutionary relationship of type 1 and type 2 isolates by phylogenetic analyses of the concatenated protein sequences of the three single copy genes and their orthologs in Tritrichomonas foetus and Pentatrichomonas hominis. Sequences for T. foetus and P. hominis orthologs were obtained from mining low coverage Roche 454 sequence data of each species, and contigs with high sequence similarity were aligned and manually edited using SeaView v. 4.2.4 [44]. Indel regions were deleted from the alignment, leaving 1147 aa aligned sequence. BioNJ [45], a distance based phylogeny reconstruction method packaged in SeaView v. 4.2.4 was used to infer the phylogeny, using Poisson protein-level distances and 1000 bootstrap replicates.
To detect linkage disequilibrium (LD) between the MS loci we calculated pairwise LD using the exact test for haplotypic data encoded in Arlequin. For single copy gene loci, we utilized the 49 SNPs found in alleles of the 94 isolates, and used the LDheatmaps [46] package in R [47] to plot the standardized measure of linkage disequilibrium between pairs of sites, r2. As this program is designed for diploid organisms, we modified our haploid data by making all SNP genotypes homozygous. LIAN software version 3.5 was used to calculate ISA, a standardized index of association that tests for multilocus linkage disequilibrium, for MS loci. ISA is defined as ISA = (VD/VE−1)(r−1), where (VD) is the variance of the number of alleles shared between all pairs of haplotypes observed in a population (D), (VE) is the variance expected under random association of alleles, and r is the number of loci analyzed. VE is derived from 10,000 simulated data sets in which alleles were randomly reshuffled among haplotypes. For single copy gene loci, we used the software package MultiLocus 1.3b [48]. This program can accommodate haploid sequencing data and implements an algorithm for ISA that is independent of the number of loci analyzed.
TVV infection in each parasite isolate was determined by isolating total RNA from 8–10 mls of late log phase cultures. RNA isolation was performed using Trizol (Invitrogen) according to the manufacturer's instructions. A total of 1 µg of total RNA was electrophoresed on a 1% agarose gel; the presence of rRNA bands on the gel served as a loading control to ensure that RNA from approximately equal number of parasites was examined. Gels were stained with ethidium bromide and isolates were considered positive for TVV if the characteristic ∼4.5 kb dsRNA genome band was detected.
In order to sample the genetic diversity and deduce the population structure of extant T. vaginalis in the local population, we collected a total of 270 vaginal swabs from female patients undergoing a pelvic examination at eight STD clinics in four boroughs of New York City (NYC) during the Summer of 2008 (Table 1). The average patient age was 27.7 years, and the majority self-identified as black non-Hispanic (N = 133, 49%) or Hispanic (N = 83, 31%), with the remainder reporting as white non-Hispanic (N = 17, 6%), Asian non-Hispanic (N = 9, 3%), American Indian (N = 1, 0.4%), Multi-ethnic (N = 3, 1%), or other (N = 10, 4%). Data on ethnicity was unavailable for 14 patients (5%). Wet mount diagnosis was performed in all clinics whenever a laboratory technician was available. Wet mount detected five T. vaginalis infections, while in vitro culture using InPouch TV packs diagnosed 19 T. vaginalis infections, and PCR amplification using three different sets of diagnostic primers detected a total of 26 infections. All culture-positive infections were detected by PCR as well, and all wet mount-positive infections were detected by both culture and PCR diagnosis. Thus wet mount, when performed, detected a mere 36% of the infections detected by PCR, and only 42% of infections detected by InPouch culture, suggesting that this method of diagnosis is highly insensitive and detection and treatment of T. vaginalis would be improved through the use of more sensitive point-of-care tests. We detected T. vaginalis infections in 10% of women attending NYC STD clinics, which is lower than the prevalence found in other STD clinics in the United States, but remains within the published range of 8–47% [8].
To gauge the extent of T. vaginalis genetic diversity within NYC, we used our panel of 21 polymorphic microsatellite (MS) markers [26] to genotype 19 isolates (seven infections detected by PCR could not be revived in culture to produce sufficient quantities of DNA for genotyping). One of the 19 isolates genotyped (NYCE32) had more than two alleles at four MS loci, indicating a mixed infection, and was excluded from further analyses. We found that each of the remaining 18 single infections had a unique haplotype, indicating high genetic diversity of the parasite even within the geographically limited area of NYC. This finding was also reflected in the moderately high average expected heterozygosity (HE = 0.67) and allelic richness estimate for a population size of g = 4 (A = 3.24). An average of 4.29 distinct alleles were identified per locus (Table 2).
To determine if the moderately high genetic diversity exhibited by T. vaginalis isolates in NYC was unique to this geographic region, we extended our studies to include a set of 231 global samples, collected from nine countries: the United States, Mexico, Chile, Italy, South Africa, Mozambique, Australia, Papua New Guinea (PNG), and India, and five standard laboratory strains commonly used in research labs (Table S1). Each sample was genotyped in duplicate for all 21 MS markers, and 216 were successfully genotyped at ≥13 of the loci (Table 2). A total of 23 mixed infections (10.6%) was identified among the 216 isolates, 22 of which were double infections, while a Chilean isolate (ANT1) appeared to be a triple infection (two alleles identified at six loci and three alleles at a seventh locus). These isolates were excluded from further analysis. We found only four pairs of isolates from different geographical regions that shared haplotypes: three pairs were isolated from the Western United States (isolates 886 and 1135; 938 and 907; 1020 and 1025), and one was collected from both the Eastern and Western United States (isolates 1027 and 1162). In contrast, all eleven Indian isolates shared the same haplotype, unfortunately due to cross-contamination during their continuous culture in the same laboratory over many years. For this reason, we collapsed the same 11 genotypes to a single data point. A lack of shared haplotypes exhibited by our world-wide collection of T. vaginalis isolates is not due to incomplete sampling because graphing allelic richness estimates for each sample size revealed that the sampling had captured the majority of non-rare alleles (Figure S1).
A variety of population genetics statistics for the 183 clinical single infections and 5 laboratory strains indicate that the global genetic diversity of T. vaginalis is high and stable from region to region. The mean expected heterozygosity across all MS loci is 0.66±0.197, ranging from 0.04 (MS03) to 0.83 (MS17) respectively, and with an average of 8.52 alleles per locus (minimum 3.0 at MS03 and maximum 29 at MS17; Table S2). The expected heterozygosity is similarly high throughout all regions (Table 2), although statistically significant differences were apparent. For example, the T. vaginalis isolates from Chile, Western and Eastern United States and Australia are more diverse while the Southern Africa, Mexico and PNG isolates comprise a slightly less diverse group. We measured population differentiation between the geographical regions using FST measurements in Arlequin, and found that the Southern African and PNG parasite populations were significantly differentiated from the other global populations. The Mexican population differed from all other populations with the exception of the Italian population. The Chilean population differed from that of the Eastern United States, which was similar to the populations of both the Western United States and Australia (Table S3). Overall, we found that the least diverse groups differed the most from other global populations.
Next, we looked for population structure among the 188 global isolates (Table S1). Using the Bayesian clustering model implemented in STRUCTURE 2.2 [39], the most probable number of clusters (populations) was determined by plotting the log probability of the data [Ln P(D)] for each k value followed by clusteredness scores. K = 2 coincided with a significant dip in the log probability of the data and received the highest clusteredness value (0.95 averaged across 10 independent simulations; Figure 1). Interestingly, the two clusters, which we refer to as ‘type 1’ and ‘type 2’, are present at nearly equal frequencies and are well distributed among all geographical locations as defined in Table S1. Two exceptions to this are isolates from Southern Africa and Mexico, which are significantly biased towards type 1 and type 2, respectively. Independent testing of this population structure was provided by two-way hierarchical clustering, which produced an identical clustering pattern, assigning the same isolates to the same clusters, and provided further evidence for a distinct two-type structure (Figure S2). Minimum spanning networks showed similar population differentiation, although we did not find perfect correlation (Figure S3). No evidence for further sub-population structure was found after repeating the analysis on each type individually.
In addition to their geographical distribution, we investigated the temporal distribution of these two types. Likelihood ratios revealed no significant difference in the frequencies of the two types when isolates were categorized by the year in which they were isolated (Figure S4). To deduce which type is older in evolutionary history, we sequenced three single-copy genes – Coronin (CRN), MutL Homolog 1a (Mlh1a), and postmeiotic segregation increased 1 (PMS1), validated for phylogenetic analyses and described previously [26] – from 94 T. vaginalis isolates. Orthologs of these single-copy genes in two distant relatives of T. vaginalis [49], Tritrichomonas foetus (a trichomonad that infects the bovine urogenital tract) and Pentatrichomonas hominis (a human intestinal trichomonad), were identified, and used as outgroups to construct the phylogenetic tree for the concatenated protein sequences of the three genes. Although support at several nodes is weak, the phylogeny suggests that parasites more similar to type 1 existed before the emergence of parasites characteristic of type 2 (Figure 2a). We also found that type 1 has greater allelic richness than type 2, which further supports its ancestral nature, as the ancestral node would be expected to have accrued greater diversity (Figure 2b).
We sought to identify phenotypic differences between the two types by correlating available clinical data with genotype (Table 3). Although the mean age of women infected with type 1 parasites is 35.0 years compared to 30.9 years for women infected with type 2 parasites, this difference was not statistically significant. We also found no statistically significant difference in the vaginal pH of women infected with the different types, nor did we find a significant difference in the percentage of isolates associated with a positive whiff test (a test used in the diagnosis of trichomoniasis). However, a highly significant difference was found in the minimum lethal concentration (MLC) of metronidazole necessary to kill isolates of the two types, with type 2 isolates demonstrating a mean MLC of 228.4 µg/ml of metronidazole, while type 1 isolates exhibited a mean MLC of 76.6 µg/ml of metronidazole. We also found that infection of T. vaginalis with the T. vaginalis virus (TVV) occurred significantly more frequently in type 1 isolates (112 of 154 isolates tested) than in type 2 (42 of 154 isolates; Table 3). Finally, our data – albeit insufficient for reliable statistics on this point – suggest that infections with type 1 parasites are more likely to be detected by wet-mount (microscopic) diagnosis than are infections with type 2 parasites; easier detection by microscopy might indicate a higher parasite load in type 1 infections.
We used several population genetics tools to address the question of whether genetic exchange occurs in T. vaginalis: (1) Pairwise linkage disequilibrium (LD: a locus to locus comparison to detect cases where specific alleles are found together more frequently than would be expected by chance alone); (2) the standardized index of association (IAS,: a measurement of the variance of the genetic distance between pairs of strains compared to variance in a shuffled matrix [50]); and (3) Maximum Chi-Squared Tests for recombination (Max Χ2: compares the distribution of polymorphic sites along paired sequences with those expected to occur by chance [51]). For parasites with a clonal population structure, i.e., with no genetic exchange, the expectation would be to observe significant LD between loci and significant IAS between MS and single-copy gene alleles.
We measured pairwise LD for each type using both MS loci and single-copy gene SNPs (Figure 3). Analysis of type 2 MS data revealed 42 cases of pairwise LD, while analysis of type 1 data MS revealed 15 (Figure 3a). This difference was confirmed upon calculation of IAS, which is highly significant in type 2 (IAS = 0.0153, p≤1.00×105) but not significant in type 1 (IAS = 0.0006, p = 0.396), suggesting that the MS loci of isolates comprising the latter are in linkage equilibrium, while those in the former are not. We also measured LD within and between the single-copy genes and found minimal LD for both types (Figure 3b). Interestingly, strong LD is restricted and rare between the three genes in type 1, whereas type 2 is characterized by higher LD distributed among all three genes. However, in both cases, it appears that there is little genome-wide linkage, suggesting that the excess LD in type 2 may be due to either a recent bottleneck, a recent loss of recombination, or even to a recent expansion of evolutionarily favorable mutations within the population. These results are also consistent with the IAS measurements calculated using LDheatmap. Breaking the three genes into linkage groups, we find that type 1 sequences have a non-significant IAS (IAS = 0.0296, p = 0.153), while type 2 is marginally non-significant (IAS = 0.0598, p = 0.057), and becomes significant when the classical Index of Association (IA) is calculated (IA = 0.1195, p = 0.046).
We tested for recombination events using Max Χ2 analysis in the bioinformatic program START2 (Table S4). Among the 36 alleles found during sequencing of the single-copy gene CRN from 202 T. vaginalis isolates, we identified one putative recombination event between alleles CRN-25 and CRN-36 (Max Χ2 = 54.6422, p = 0.030). Using a p = 0.05 cutoff, we also identified 89 putative recombination events within 37 unique alleles identified from sequencing PMS1 of 144 T. vaginalis isolates, and 15 putative recombination events among the 41 unique alleles identified through sequencing Mlh1a of 110 T. vaginalis isolates.
Finally, we compared the type assignments inferred by STRUCTURE from MS genotyping with phylogenies constructed from DNA sequences of each of the three single-copy genes (Figure 4). We found that the topologies are similar, each supporting a two-type population structure; however, in a number of cases, isolates from different types had identical SNP haplotypes within one gene but very different haplotypes within the other two genes. This suggests that some of our T. vaginalis isolates are recombinants that were generated through genetic exchange, which appears to occur within types and rarely between types. The DNA phylogenies are also shown in Figure S5, where isolates are color-coded by geographical origin.
The existence of genetically different T. vaginalis ‘types’ has been suggested by several previous studies. Stiles et al. (2000) found ten distinct HSP70 RFLP subtypes using 36 global reference strains and isolates collected from patients in Mississippi [19]. Rojas et al. (2004) utilized RAPD markers to genotype 40 isolates from Cuba and identified four subtypes with dendrograms inferred using UPGMA (unweighted pair group methods analysis) [23]. Meade et al. (2009) employed RFLPs to type 129 U.S. clinical isolates and used phylogenetic methods to identify two major groups composed of five subgroups [20]. Snipes et al. (2000) utilized RAPD polymorphisms and neighbor-joining phylogenetic methods with 63 U.S. strains and identified two groups [22]. Our own work with single-copy genes and microsatellites and a small number of laboratory strains also suggested a two-group genetic structure [26]. This current study, however, is the first to conclusively demonstrate the global distribution of two T. vaginalis types using robust, reproducible and diverse population genetic markers to genotype >230 isolates from nine regions around the globe. Our use of a set of powerful population genetic statistical tools, ranging from cluster analysis to two-way hierarchical clustering, along with tests for recombination, has allowed a rigorous investigation into T. vaginalis genetic diversity, population structure and genetic exchange. This information will be essential for future investigation into the parasite's biology.
We also demonstrate the utility of our panel of microsatellite markers to detect mixed genotype clinical infections. In a recent review, Balmer and Tanner discuss the theoretical and experimental work that suggests that mixed infections have a broad range of clinically relevant effects in a number of human pathogens, including effects on the host immune response, the ability to efficiently prevent and treat infection, and changes to pathogen and disease dynamics caused by intraspecific interactions, many of which can lead to pathogen evolution [52]. The availability of sensitive methods allowing the detection of multiple genotype infections in T. vaginalis research is likely to prove highly significant in understanding clinical trichomoniasis, as the ∼11% prevalence of mixed isolates identified in our study represents a non-trivial number of real T. vaginalis infections.
How did the striking two-type population structure of T. vaginalis arise? We propose three scenarios. We have previously hypothesized that the ancestor of T. vaginalis was an enteric pathogen (as are most trichomonads) that transitioned to the urogenital tract during its evolution [53]. It is possible that two separate colonization events occurred, producing two genetically distinct lineages within the urogenital niche. If genetic exchange in T. vaginalis is a rare event, this could explain the maintenance of the two lineages. Alternatively, it is possible that the two types evolved sympatrically after a single colonization event. The presence of the two types in nearly equal frequencies globally suggests that some form of balancing selection is maintaining both types in natural infections. Potential drivers for this balancing selection, i.e., what causes one type to have an evolutionary advantage over the other under different selective conditions, may become apparent from studies characterizing phenotypic differences between the types. At this point, we have identified type-specific differences in frequency of T. vaginalis virus (TVV) infection and in susceptibility to metronidazole. As yet untested phenotypes that may be important in this context are: (a) differences in the ability to colonize the urogenital tracts of male versus female hosts; (b) a reduction in parasite fitness associated with metronidazole resistance when metronidazole treatment is not a selective force; or (c) differences in growth rates and virulence. Third and finally, the population structure reported here may have evolved when barriers arose that reduced the parasite's ability to undergo genetic exchange, causing gradual genetic isolation. In this respect it is interesting to note that ∼60% of the ∼160 Mb T. vaginalis genome consists of active transposable elements, virus-like repeats and retrotransposons [53], foreign DNA whose parasitism of the genome could have influenced the mechanics of genetic exchange, for example chromosome pairing. Indeed, transposable elements have been postulated to play a significant role in facilitating ectopic recombination in Drosophila melanogaster [54].
In contrast to the near-equal frequencies of the two T. vaginalis types detected in most regions, we found significant bias toward type 1 in Southern African samples and toward type 2 in Mexican samples (Figure 1). The low sample number (N = 11) for the Mexican isolates may explain why the frequencies appear to differ in this region; however, the 23 Southern African sources were comparatively diverse, comprised of asymptomatic women attending an antenatal clinic and symptomatic women attending an STD clinic.
Our finding of a highly significant difference in the frequency of TVV infection between type 1 and type 2 may have important implications for understanding variation in T. vaginalis virulence and disease pathogenesis. TVV has been implicated in affecting the expression of cysteine proteinases and of a highly immunogenic protein family (P250) on the parasite's surface [55]–[57]. In regard to the potential of such double-stranded RNA viruses to influence pathogenicity, it has been recently demonstrated that Leishmania RNA virus-1 controls the severity of mucocutaneous leishmaniasis by inducing Toll-like receptor 3, and ultimately inducing proinflammatory cytokines and chemokines that increases susceptibility to infection [58]. In addition, the greater prevalence of the virus in one type over the other may suggest differences in the functionality of the RNAi machinery that has been identified in the T. vaginalis genome [53].
The interesting observation that type 2 isolates have a significantly higher MLC for metronidazole may also have repercussions for understanding the mechanism(s) of metronidazole resistance in T. vaginalis, which has so far eluded scientists [59]–[61]. Isolates with an in vitro aerobic MLC of greater than or equal to 50 µg/ml are considered resistant to metronidazole [62], suggesting that the difference in median metronidazole MLCs of the two types may be clinically relevant (25 µg/ml vs. 200 µg/ml), and may have influenced the evolution of the species. For example, it is tempting to speculate that type 2 isolates may have diverged from type 1 isolates due to a selective advantage in being able to evade higher levels of metronidazole. This could account for the derived position of type 2 isolates in the T. vaginalis evolutionary tree (Figure 2), and could also explain their relative lack of diversity and genetic recombination, since there has been less time for mutations to accumulate in the more recently-evolved lineage. In addition, through limiting recombination, type 2 isolates may maintain favorable gene combinations such as those for increased metronidazole tolerance. It should be noted, however, that it is unlikely that metronidazole treatment has been adequately widespread to induce such selective evolution.
A major goal of this work was to use population genetics to identify evidence of genetic exchange in T. vaginalis. The parasite divides mitotically in the host, and no gamete form or cell fusion has been observed in vitro. However, circumstantial evidence suggests that T. vaginalis parasites may be capable of infrequent genetic recombination or may have only recently lost this ability. For example, analyses have revealed that closely related isolates share biologically relevant phenotypes, such as metronidazole resistance, but this pattern has no correlation with geographical origin, suggesting a spread of the phenotypes by recombination and the presence of strong selection [25]. In addition, Cui et al. (2010) found reassortment of polymorphic TMAC pseudogenes that cannot be explained by a strictly clonal population structure [63]. More persuasively, analyses of the T. vaginalis genome identified a complete set of conserved meiotic genes, suggesting that the meiotic process remains under, or has only recently been relieved of, conservative selection pressure [53], [64].
Tibayrenc and Ayala (2002) have outlined criteria and tests of clonality relating to eukaryotic parasites [65]. Among the criteria clonal organisms should meet are (1) the presence of over-represented, identical genotypes that are widespread; (2) evidence of linkage disequilibrium; and (3) the absence of segregating or recombinant genotypes. To address the first criterion, our studies found significant genotypic diversity (average HE 0.66) and few shared haplotypes (total two in four isolates) among 188 T. vaginalis global isolates. The second criterion was addressed through analysis of the haplotypes generated using 21 microsatellite markers. Results of these tests indicated that T. vaginalis populations – and in particular type 1 – are in linkage equilibrium, indicative of genomes that have recently undergone recombination. Finally, we have identified recombination events between alleles of three different single-copy genes, providing evidence of recombinant genotypes. Taking these data as a whole, we infer that T. vaginalis does not fit the clonality model but rather appears to have undergone frequent genetic exchange in its recent evolutionary past. In addition, the presence of a complete set of meiosis-specific genes and the frequency (∼11%) at which mixed infections encounter each other in the host, suggest that the parasite continues to be capable of recombination, although the rate at which it occurs and under what conditions in natural populations remain to be determined.
The ability of T. vaginalis parasites to undergo genetic exchange has significant implications for the epidemiology and control of trichomoniasis. The Weismann hypothesis argues that genetic recombination functions to provide variation for natural selection to act upon, giving recombining species an evolutionary advantage in responding to selective pressures [66]. In other words, it provides opportunities for newly emerged, beneficial genes to be exchanged, potentially allowing them to be combined with other favorable genes, which may ultimately allow for the novel gene to become widespread throughout a population. This has obvious implications for such phenotypic traits as drug resistance, where a rare gene favorable to the parasite (and unfavorable to the host) may become widespread, with grave implications for treatment of the host. Not all consequences of genetic recombination are negative, however; should the mechanisms and conditions conducive to meiosis and genetic recombination in T. vaginalis be elucidated, important resources such as genetic crosses and quantitative trait loci (QTL) maps could be developed, significantly advancing our understanding of this neglected parasite.
|
10.1371/journal.pbio.1002231 | Neocortical Rebound Depolarization Enhances Visual Perception | Animals are constantly exposed to the time-varying visual world. Because visual perception is modulated by immediately prior visual experience, visual cortical neurons may register recent visual history into a specific form of offline activity and link it to later visual input. To examine how preceding visual inputs interact with upcoming information at the single neuron level, we designed a simple stimulation protocol in which a brief, orientated flashing stimulus was subsequently coupled to visual stimuli with identical or different features. Using in vivo whole-cell patch-clamp recording and functional two-photon calcium imaging from the primary visual cortex (V1) of awake mice, we discovered that a flash of sinusoidal grating per se induces an early, transient activation as well as a long-delayed reactivation in V1 neurons. This late response, which started hundreds of milliseconds after the flash and persisted for approximately 2 s, was also observed in human V1 electroencephalogram. When another drifting grating stimulus arrived during the late response, the V1 neurons exhibited a sublinear, but apparently increased response, especially to the same grating orientation. In behavioral tests of mice and humans, the flashing stimulation enhanced the detection power of the identically orientated visual stimulation only when the second stimulation was presented during the time window of the late response. Therefore, V1 late responses likely provide a neural basis for admixing temporally separated stimuli and extracting identical features in time-varying visual environments.
| Animals are constantly exposed to a visual world that varies over time. To examine how the visual cortex integrates visual information that is temporally spaced, we monitored neuronal activity of the primary visual cortex (V1) using single- and multicell recording techniques. We discovered that a brief visual stimulus induced an early, transient activation as well as a delayed reactivation of V1 neurons in mice and humans. Notably, this reactivation of visual cortex conveyed information about stimulus orientation: presentation of a second visual stimulus during this reactivation enhanced the V1 response specifically when the orientations of the two stimuli were identical. Behavioral tests in mice and humans revealed that the ability to detect visual stimuli was also enhanced when the second stimulus was presented during the time window of V1 reactivation. Because animals extract visual information from an environment in constant change, the modulation of visual responses through cortical reactivation might be a strategy commonly used in the visual system.
| The primary visual cortex (V1) has been used as an experimental model to study cortical responses to sensory input. V1 receives direct synaptic inputs from the lateral geniculate nucleus (LGN) of the thalamus and provides the output of its computation to higher-order cortical areas [1,2]. This route, commonly known as the feed forward pathway, contributes to the hierarchical neural processing of specific visual features, such as orientation, direction, color, and motion. Classical visual processing models consider V1 as a passive relay station for visual information; that is, V1 encodes instantaneous information by transiently responding to the present stimulus feature. However, recent evidence has demonstrated that V1 activity persists over time [3–7] and even propagates throughout the V1 network [8,9]. This complex activity is likely associated with the representation of reward timing [4,5], iconic memory [10,11], and working memory [12–14]. Indeed, reverberatory neuronal activity within neocortical circuitry has been proposed as a potential mechanism for short-term storage of information [15,16].
How does V1 encode the external world while under a constant flow of visual stimuli? The measurement of cortical dynamics has revealed that V1 response tuning evolves with time [17], during which it may interfere with later V1 information [18]. Indeed, preceding visual stimuli are reported to modulate visual perception after brief stimulus-onset asynchrony (SOA) [19–22]. Therefore, poststimulus V1 activity appears to intermingle with the subsequent visual information, which produces a complex output [23–25].
In this study, we discovered a novel V1 activation pattern in nonanesthetized mice; in virtually all V1 neurons, an oriented flashing light–induced biphasic membrane voltage (Vm) response that consisted of an early, transient depolarization and a late, slow depolarization. The late response exhibited high orientation selectivity, which indicates that V1 maintains the information of a recent stimulus with high fidelity for some time. Flash-induced late response was also observed using electroencephalogram (EEG) recordings in humans, suggesting that a long-delayed V1 reactivation prevails in mammals. To understand the effect of the late response on the upcoming visual input, we paired a flashing stimulus to another visual stimulus with a time lag. Flashes modulated the V1 response to the subsequent input in an orientation-selective manner. The flash-induced selective modulation was also replicated in the psychophysical parameters of mice and humans.
We monitored the spiking activity of V1 layer (L) 2/3 neurons of P35–P44 mice using the cell-attached recording technique (Fig 1A) and applied a brief flashing stimulus (17–50 ms) of a full-field grayscale sinusoidal grating with one of four orientations (0°, 45°, 90°, and 135°) to the eye contralateral to the recording site. As previous reports have demonstrated that L2/3 neurons fire sparsely [26–30], 56.5% of V1 neurons (43 of 76 cells) exhibited a significant increase in their firing rates in response to the grating flashes (defined by a criterion of p < 0.05 versus the baseline firing rates, Z test for comparison of two counts [31]). The responses were classified into two types; the first type of responses was spikes immediately (< 0.3 s) after the stimulus onset (early spiking, Fig 1A top), whereas the second type was spikes with latencies longer than 0.4 s (late spiking, Fig 1A bottom). In the pooled data, the population firing rates exhibited two distinct peaks that corresponded to the first and second types of spikes; for individual responsive neurons, the mean firing rates during the early and late responses were 1.27 ± 0.91 Hz and 0.28 ± 0.19 Hz, respectively (mean ± standard deviation [SD] of 11 and 36 neurons). Late-spiking neurons were numerically dominant (Fig 1B, inset). Thus, we defined the early and late responses as activity that occurred between 0–0.3 s and 0.4–2 s, respectively.
To investigate the subthreshold Vm dynamics that underlie the biphasic spike responses, we conducted whole-cell current-clamp (I = 0) recordings from V1 neurons (S1A and S1B Fig). In the typical neuron shown in Fig 1C, a grating flash reliably induced early and late depolarization responses. Remarkably, we observed similar biphasic Vm responses in all 28 recorded neurons (S1C Fig), irrespective of their firing types, including nonspiking neurons (S1D Fig). The early depolarization was transient and peaked at latencies of < 0.3 s, whereas the late depolarization was more persistent and peaked at approximately 0.4−2.0 s. On average, the peak amplitudes of the early and late depolarizations were 6.7 ± 4.2 and 6.4 ± 4.4 mV (mean ± SD of 28 cells), respectively, and were correlated with each other (S1C Fig left). The area under curves of individual Vm traces during a late period of 0.4–2.0 s (late area) was correlated with their peak amplitudes (S1C Fig middle). Therefore, we quantified both early and late responses using their peak amplitudes in the following analyses. The areas of late responses were not correlated with their peak latencies (S1C Fig right). Thus, the latencies did not affect the magnitude of late responses. This fact also validates our choice of the time window for late Vm responses (0.4–2.0 s).
The fact that late depolarizations occurred in all recorded neurons suggests that late visual responses represent a global phenomenon that involves the entire V1 cortex. To confirm this possibility, we recorded local field potentials (LFPs), which reflect the compound activity of multiple neurons surrounding the tip of a recording electrode [32]. We found that LFPs in V1 L2/3 responded reliably to a grating flash with biphasic negative fluctuations (Fig 2A). The response signal, if any, was less evident in LFPs recorded from the retrosplenial cortex, a more anterior brain region. We also recorded voltage dynamics of the neocortical surface. We loaded the cerebral surface with RH-1692, a voltage-sensitive dye (VSD), and monitored the spatiotemporal patterns of flash-evoked activity [33]. As expected by the LFP data, early cortical VSD responses were observed in V1 (S2 Fig). Then, the VSD signal decreased transiently, producing a transitional period. After approximately 0.4 s, the late VSD responses also arose at V1. Therefore, similar to Vm responses in patch-clamp recordings, the VSD signal in V1 was biphasic. We extended the field potential work to visual responses in humans. We recorded EEG from 10 adult participants and measured visual event-related potentials (ERPs) at O1 and O2, according to the international 10/20 coordinate convention [34]. Human ERPs in response to grating flashes were also biphasic; an early and late negative reflection peaked around 0.15 s and 0.7 s, respectively, after a grating flash (Fig 2B).
Previous studies have also reported a specific form of late, slow activation of the rat V1 [4,5] and the mouse primary somatosensory cortex [35]; however, these responses emerged as a result of sensory reinforcement learning and were not observed in naïve animals. There is also a study that has reported biphasic responses in naïve cat visual cortex [36]; however, the latency and the duration of this late response was much shorter. By contrast, our flash-evoked late V1 responses occurred in naïve animals and had a much longer latency and duration. Therefore, they represent novel V1 dynamics. This discrepancy most likely occurs as a result of the difference in the features of visual stimuli. Indeed, the durations of flashes were critical [7]; we failed to observe evident long-delayed LFP activity at flash durations of more than 200 ms (Fig 2C). Moreover, we used full-field flashes, which might recruit synaptic inputs from both classical and nonclassical visual receptive fields. It should also be noted that flash-induced late response has a much longer duration than the well-known OFF response that has been described in other studies [37].
The amplitudes of both early and late responses increased at higher contrasts of flash gratings (S3 Fig). Thus, it is feasible that the late responses encode the orientation of flashing stimuli [36]. We measured the orientation selectivity, which is a characteristic of V1 neuron responses [38–41]. Grating flashes with various orientations induced different changes in the late spike rates (Fig 3A and S4A Fig). We calculated the orientation selectivity index (OSI) for each late-spiking neuron. On average, the OSIs were 0.37 ± 0.25 (mean ± SD of 36 cells). To evaluate the statistical significance of OSIs, we compared them with the chance distribution obtained from the trial-shuffled surrogate data (Fig 3B). Overall, the OSIs exhibited significantly higher values than chance, which indicates that the late-spiking responses were orientation-selective (p = 3.3 × 10−3. D = 0.29, n = 36 cells, Kolmogorov-Smirnov test). Late subthreshold Vm responses were also significantly orientation-selective (Fig 3C and 3D, p = 2.7 × 10−9, D = 0.66, n = 34 cells, Kolmogorov-Smirnov test). Their OSIs were lower compared with the late spike responses (S4B and S4C Fig, p = 5.0 × 10−3, t19 = 3.17, n = 20 cells, paired t test), consistent with many previous reports about orientation selectivity of Vm responses [42–44].
Because the early responses were also orientation-selective, we focused on the tuning properties of the early and late responses. We computed the correlation coefficients between the early and late Vm tuning curves of each cell and compared the pooled data to the chance-level distribution of the correlation coefficients in their trial-shuffled surrogates. The correlation coefficients were significantly higher compared with chance, which indicates that the early and late Vm responses of each neuron had a similar orientation tuning (Fig 3E, p = 0.014, D = 0.27, n = 34 cells, Kolmogorov-Smirnov test). Moreover, the OSIs of late responses were positively correlated with the OSIs of early responses (Fig 3F, R2 = 0.61, p = 1.2 × 10−4, t17 = 4.94, t test for a correlation coefficient). Note that neither early nor late OSIs depended on firing rates (S4A Fig, p = 0.490, R2 = 0.01). We thus conclude that late responses conveyed selective information of visual stimuli.
We further confirmed flash-induced responses using two-photon calcium imaging. We loaded V1 L2/3 neurons with Fura 2 by pressure-applying its acetoxymethyl ester (AM) derivative (S5A Fig). The amplitude of a spike-elicited calcium elevation (|ΔF/F|) was nearly linear with the number of action potentials involved in the calcium event (S5B Fig). Note that our imaging system was able to resolve two action potentials at an interspike interval of less than 400 ms (S5C Fig), allowing us to classify early and late spiking neurons. We imaged spike-triggered calcium events en masse from 64.6 ± 6.04 neurons per video (mean ± SD of nine videos from nine mice) with a single-cell resolution at five frames per s (S5D Fig). In the example neuron shown in S5E Fig, the amplitudes of the ΔF/F responses evoked by grating flashes exhibited orientation selectivity. Of the 581 neurons, 323 (56%) neurons were responsive to flashes, and the preferred orientations were uniformly distributed (S5F Fig). Because early spiking responses occurred around 0.1–0.2 s after a flash, they would be reflected in a rapid ΔF/F increase in the first video frame (0.2 s) after the stimulus. According to this definition, we estimated that early spiking neurons contributed 10.0% (58 out of a total of 581 cells), consistent with patch-clamp recording data showing that the majority of flash-responsive neurons are of the late-spiking type (Fig 1B inset and S1D Fig). Therefore, we assumed that most ΔF/F responses reflected putatively late spikes. Although they may overlap with the early-spiking component, the orientation tuning properties were approximately congruent between the early and late responses (see Fig 3E), and thus, the ΔF/F response tuning is still thought to reflect the late-spiking tunings. Consistent with this notion, the distribution of OSIs in the ΔF/F responses was similar to the late-spiking responses obtained by patch-clamp recordings (S5G Fig, p = 0.497, D = 0.15, Kolmogorov-Smirnov test) and was higher than that of their surrogate data (p = 2.3×10−6, D = 0.15, n = 323 cells).
Because the late response has a long latency, it may interact with a subsequent visual stimulus. We tested this idea by recording the ΔF/F responses to grating stimuli that moved for 2 s toward one of eight directions (0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°), which were presented alone (Drift-only trials) or 0.5 s after grating flashes (Flash+Drift trials). To minimize photobleaching and phototoxicity, we did not test all possible combinations of the flash orientations and the drifting grating directions; instead, we fixed the grating flash orientation to 0° (vertical orientation; vFlash) and reduced the total imaging period (S6A Fig). We compared the ΔF/F responses between Flash+Drift and Drift-only trials and examined how the preceding vFlash (prime) modulated the ΔF/F responses to subsequent drifting gratings (target). The combinational pattern of a vFlash stimulus and a drifting grating was described as a Δorientation, which represents the orientation difference between vFlash and the drifting gratings and comprised a value of −45°, 0°, 45°, or 90° (= −90°). In Drift-only trials, Δorientation indicates the difference between 0° and the orientations of drifting gratings (i.e., the absolute orientation).
S6B Fig summarizes the data from a representative neuron. For each Δorientation in Drift-only and Flash+Drift trials, we statistically judged whether the neuron responded, i.e., whether the ΔF/F amplitude was significantly higher compared with the baseline ΔF/F fluctuation (p < 0.05, n = 10−18 trials, paired t test). The significant responses are marked by dark red boxes below the tuning plot. Three other examples are shown in S6C Fig. We pooled the data from the 581 neurons (S6D Fig). For each Δorientation, we compared the number of cells that exhibited significant ΔF/F in Drift-only trials to the number of significant cells in Flash+Drift trials. Notably, the number of significantly responsive cells increased at Δorientation = 0°, where the orientations of vFlash and drifting gratings were matched. The number of responsive cells did not increase at the other Δorientations. Thus, two sequential stimuli with the same orientation activated V1 neurons more efficiently compared with stimuli with different orientations. By focusing on individual cells that were activated under the iso-orientation condition, we analyzed their intrinsic orientation preferences. Flash-induced response enhancement was more evident in cells whose preferred orientations were different from the stimulus orientation (S6E Fig). These data indicate that a flash recruited otherwise irresponsive cells (due to their cross orientation preferences) to a subsequent stimulus with the same orientation as the flash.
Previous studies have reported that paired visual stimuli lead to a functional adaptation of neuronal responses to the target [45,46]. In other words, visual cortical neurons decrease their responsiveness to repeated stimuli. Calcium imaging did not allow us to strictly quantify the response amplitude, and we could not determine whether the observed changes are adaptation (desensitization) or priming (sensitization). To quantify the effect of flashes in more details, we returned to patch-clamp recordings of subthreshold Vm responses. In these experiments, the drifting grating orientation was fixed to vertical (0°, 180°; vDrift), and the orientations of the preceding flashes varied across four orientations (0°, 45°, 90°, or 135°) in a pseudorandom order (Fig 4A). First, the SOA was set to be 0.5 s (Fig 4B). We compared the amplitudes of Vm responses to a combination of flash and vDrift stimuli (Flash+vDrift) with those of the responses to vDrift alone (vDrift-only). On average, the absolute amplitudes of Flash+vDrift responses were larger than those to vDrift-only responses (p = 0.012, t51 = 2.60, paired t test); however, for individual neurons, the amplitude relations depended on the amplitudes to responses to Flash alone (Flash-only, Fig 4C). That is, when a neuron exhibited a large depolarization in Flash-only trials (>2 mV), then the depolarization in Flash+vDrift trials was more increased compared to vDrift-only responses. On the other hand, when a neuron exhibited a small depolarization in Flash-only trials (<2 mV), the Flash+vDrift response amplitude was nearly comparable to the vDrift-only response amplitude. To further examine this effect, we employed a new analysis in which we compared Flash+vDrift responses with the linear summation of the Flash-only response and the vDrift-only response (Fig 4D). We found that this augmentation occurred below the value of simple arithmetic summation of two responses. That is, individual responses to Flash-only and vDrift-only stimuli were sublinearly integrated in Flash+vDrift trials (Fig 4D). In our experimental conditions, therefore, a flash facilitated the vDrift responses through a sublinear integration of Vm depolarizations. Notably, their sublinearity differed depending on the orientations of flash gratings and was smaller at Δorientation = 0° than at 90° (Fig 4D). In other words, when two orientations of flash gratings and drifting gratings were matched, the combined responses were less sublinear, thereby exhibiting apparently larger response amplitudes, which is consistent with the flash-induced enhancement in the calcium imaging experiments. This Δorientation-dependent difference was not found at SOAs of 0.05 or 3 s (Fig 4D), suggesting the involvement of the orientation selectivity of flash-induced late responses. We replotted these sublinear behaviors (SOA = 0.5 s) as a function of the difference between their intrinsic orientation preferences and the orientation of the grating stimuli. Flash-induced response sublinearity was the largest in cells whose preferred orientations were identical to the stimulus orientation (Fig 4E). This was also consistent with the results in calcium imaging.
Flash-induced modulation of V1 neuronal activity prompted us to evaluate its behavioral consequences. We first measured the visual performance of mice using a virtual optomotor test, which can assess the visual detection ability of naïve mice without behavioral training [47]. A freely moving mouse was placed on the circular platform surrounded by four computer screens on which vertically orientated gratings moved leftward or rightward for 2 s (Fig 5A; vDrift). As a visuomotor reflex, the mouse turned its head in the same direction as the vDrift movement, a behavior that is called a tracking response. The ratio of trials with the tracking responses to the total trials was calculated as the tracking rate and was used as a quantitative measure of visual function. Under the baseline conditions (i.e., vDrift-only trials), the mean tracking rate was 74 ± 13% (mean ± SD of 10 mice). This ratio increased to 86 ± 10% when vertical flashes were presented 0.5 s before vDrift (Fig 5B, Δorientation = 0°; p = 0.037, t9 = 2.45, paired t test). This increment was not observed when horizontal flashes (Δorientation = 90°) were coupled (Fig 5B; p = 0.92, t9 = 0.10) or when vertical flashes were presented at an SOA of 3 s (Fig 5C; p = 0.69, t10 = 0.41). In mice that received local injection of 10 μM tetrodotoxin into the V1, flash-induced responses in V1 LFP disappeared (S7A Fig). In these mice, the tracking rate for the vDrift-only trials was reduced to 18 ± 16% (n = 4 mice, p = 0.026 versus naïve mice, t3 = 4.13, Student’s t test) and was not increased by vertical flashes (S7B Fig). Thus, flash-induced increases in the tracking rates likely depend on V1 late responses.
Finally, we conducted a psychophysical test in humans. The participants were asked to report the motion directions of 0.25-s drifting gratings (0°, 90°, 180°, or 270°) by flicking a computer mouse toward the same direction within 0.70 s (Fig 5D). In Flash+Drift trials, grating flashings at orientations of 0°, 45°, 90°, or 135° were presented 0.5 s before the drifting gratings. The correct response ratio was approximately 100% and was not modulated by grating flashes with either Δorientation (Fig 5E; p > 0.05, n = 11 humans, n = 486−500 trials each, Student’s t test). However, the latency of the flicking response was significantly shortened at Δorientation = 0 (Fig 5F; Drift-only: 357.3 ± 54.6 ms versus 0°: 347.8 ± 56.0 ms, mean ± SD; P = 0.007, t993 = 2.71). We did not think that this effect was due to illusory motion perception, because the grating phase of a flash stimulus and the first frame of the following drifting stimulus were identical. However, to examine the possible involvement of motion illusion, we presented two successive flashes at an SOA of 0.5 s with various combinations of the grating phases and asked participants to answer the "felt" motion direction (S8 Fig). Each stimulus condition was repeated for 80 times. As a result, the participants were not able to distinguish the motion direction; the responses were approximately 50% (= the chance level). Thus, two consecutive grating stimuli at an SOA of 0.5 s per se did not induce a motion perception.
We discovered that a brief flashing light evokes long-delayed, slow activation of the mouse V1 network. The late response was observed using different techniques, including patch-clamp recording, LFP recording, VSD recording, and EEG recording, which exclude the possibility of our recording artifact. Importantly, the late response actively interacted with subsequent visual input. This novel phenomenon was heretofore overlooked, probably because past studies tended to record visual responses for shorter terms (up to a few hundreds of milliseconds) than our work, and because we used a short flash of full-field gratings, a stimulus pattern that is not very common in vision research. Another reason for the overlook of the late responses may be a consensus that visual responses occur within a few hundred milliseconds after the onset of the visual stimulus, which might have prevented an attempt to record visual responses for seconds.
There are mainly three candidates for the initiation site of the late response. First, the late activation of V1 circuit might be generated through reverberation of the recurrent circuit within the V1. Theoretically, cortical activity is sustained by local reverberation within a recurrent network [15,16]. Anatomically, L2/3 is enriched with horizontal synaptic connections [48,49] and provides the structural basis of a recurrent circuit. Although V1 L2/3 neurons receive synaptic inputs with various orientation preferences [50], the synaptic connection probability is biased toward a similar orientation preference [51,52]. Recent studies have demonstrated that neurons derived from the same precursor cells are more likely connected and share the same orientation preference [53–55]. These observations suggest the existence of fine-scale subnetworks dedicated to process specific information [56]. We determined that the tuning properties were significantly correlated between the early and late responses. Hence, the neuron population activated by a grating flash is preferentially reactivated at the late phase. The visual cortex may filter visual input information through its specifically wired, reverberatory network [57] and may offer a high orientation tuning during the late response. The second possibility is that the V1 rebound activity arose from subcortical regions, including the lateral geniculate thalamus and the superior colliculus (and even the retina). The lateral geniculate thalamus is anatomically eligible for generating rebound activation, because it contains a recurrent network and receives feedback projections from V1 [37,58]. This anatomy might have led to the reliable observation of late response even in the LFP recording. Finally, top-down inputs from higher-order cortices may also have the ability to induce late responses, as recently reported in the hindlimb somatosensory cortex [59]. However, the latency of the late response in the visual cortex was much longer than that observed in the study, suggesting a more complex mechanism than a simple top-down feedback process.
We speculate that reverberatory activity in V1-recurrent circuits admixes with late-coming feed forward V1 activity. Recent studies have demonstrated that costimulation of the thalamocortical and cortical pathways efficiently depolarizes cortical neurons through nonlinear summation [60,61]. Although a single L2/3 neuron receives variously tuned synaptic inputs irrespective of the orientation preference in the cell’s spike output [50], synaptic inputs over dendritic trees are nonrandomly distributed and are often spatially clustered [62–64]. Thus, synaptic inputs from flashing and drifting gratings may be locally converged and may lead to nonlinear dendritic boost [61,65] when two orientations are matched.
At the network level, a grating flash enhanced (or sublinearly integrated) the V1 responses to subsequent drifting gratings in an orientation-selective manner. In these experiments, we used an SOA of 0.5 so that drifting gratings arrived during the period of flash-evoked late responses. Calcium ΔF/F responses to the drifting gratings were enhanced only when their orientations were identical to the preceding flashes. The flash-induced facilitation can be explained by two possibilities. First, the priming effect may facilitate the responses to sequential stimuli [66,67]. However, flash-induced response enhancement is not a normal form of priming because it was not a simple mixture of membrane-potential depolarizations. Flash-induced late response and the response to drifting grating were integrated in a sublinear fashion, but more linearly at Δorientation = 0°, suggestive of the partial existence of priming. It also differed depending on preferred orientations of the neurons. The second possibility is that the facilitation occurred through top-down neural processing [68], especially feature-based attention [69,70]. It is well known that attention modulates the responsiveness of neurons that have receptive fields within the attentional loci [71–73], enhancing task performance on late-coming target stimuli [70,74]. Moreover, it is important to note that feature attention in humans is effective at an SOA of approximately 0.5 s [69], consistent with our findings. Developing a psychophysical method to measure the attentional effect in mice may help verify the second possibility. Focusing on individual neurons and their orientation preferences, a flash recruited neurons with shifted-orientation preferences at the Δorientation = 0° condition. In other words, neurons with cross orientated preferences to the flash orientation were less subject to the sublinearity when the responses were integrated. Consistent with this notion, at Δorientation = 90°, neurons with cross orientated preferences to a flash (i.e., iso-orientated with regard to the orientation of the drifting stimulus) exhibited the minimal sublinear property. Thus, flash-induced late responses might function to recruit neurons that are otherwise irresponsive, leading to stronger activation of the V1.
We found that ongoing visual processing and perception were both affected by the immediately preceding visual information in a feature-specific manner; however, we could not directly show the causal contribution of flash-induced delayed depolarizations per se to subsequent visual perception. Optogenetic prevention of the delayed responses [35] is not applicable to our cases; that is, even if optogenetic manipulation is performed only during the delayed activity period, it inevitably affects both flash-induced delayed responses and drifting grating-evoked activity and cannot isolate the effect of the flash responses on visual perception. Therefore, we need to seek a way to specifically diminish the delayed activity without affecting drifting grating-evoked activity.
In this study, we regarded the featured flashes as a model of the initial visual scenes and aimed to separate the effect of suddenly coming and subsequently continuing visual scenes. Hence, we think that, under natural conditions, the pattern-selective late responses observed here may work to facilitate the responses to the passing object, possibly linking our findings to studies on trans-saccadic integration [75–77]. However, two major concerns remain unresolved. First, the late response occurred to flashes with durations of less than 50 ms, whereas natural saccades usually last about 300 ms. Thus, we cannot rule out the possibility that the late response we found is involved in other visual processes than trans-saccadic integrations. Second, although we obtained the behavioral correlates of flash-induced effects on visual function, flashes recruited neurons that were otherwise irresponsive because of the nonpreferred orientation. Therefore, flashes may increase the overall activity level of V1 and diminish the selective responsiveness of individual neurons. According to this notion, the facilitation of V1 activity would decrease the discrimination acuity of the animal, but at the same time, it could increase the sensitivity per se by lowering the visual detection threshold. This possibility must be clarified using a new behavioral paradigm that can distinguish visual detection from visual discrimination.
Animal experiments were performed with the approval of the animal experiment ethics committee at the University of Tokyo (approval number: 21–6) and according to the University of Tokyo’s guidelines for the care and use of laboratory animals. In human studies, the experimental protocol was approved by the Human Research Ethics Committee of the University of Tokyo (approval number: 24–3) and the Center for Information and Neural Networks (approval number: 1312260010). All participants were provided oral and written informed consents, and they signed the consent forms prior to each experiment.
Postnatal days (P) 28–35 male C57BL/6J mice (Japan SLC, Shizuoka, Japan) were used in the animal experiments as previously described in detail [78,79]. The animals were housed in cages in standard laboratory conditions (a 12-h light/dark cycle, free access to food and water). All efforts were made to minimize the animals' suffering and the number of animals used. The animals were anesthetized with ketamine (50 mg/kg, i.p.) and xylazine (10 mg/kg, i.p.). Anesthesia was confirmed by the lack of paw withdrawal, whisker movement, and eye blink reflexes. The head skin was then removed, and the animal was implanted with a metal head-holding plate. After 2 d of recovery, the head-fixation training on a custom-made stereotaxic fixture was repeated for 1−3 h per d until the implanted animal learned to remain quiet. During and after each session, the animal was rewarded with free access to sucrose-containing water. During the final three sessions, sham experiments were conducted to habituate the animal to the experimental conditions and noise. On the final 2−3 d, the animal was maintained virtually immobile, i.e., quiet but awake, for more than 2 h. After full habituation, the animals were anesthetized with ketamine/xylazine. A craniotomy (1 × 1 mm2), centered at 3.5 mm posterior to the bregma and 2.0 mm ventrolateral to the sagittal suture, was performed, and the dura was surgically removed. The exposed cortical surface was covered with 1.7–2.0% agar at a thickness of 0.5 mm. Throughout the experiments, a heating pad maintained the rectal temperature at 37°C, and 0.2% lidocaine was applied to the surgical region for analgesia. For patch-clamp recordings, the recorded area was confirmed by posthoc imaging of the intracellularly loaded Alexa 594, which was dissolved at 50 μM in patch-clamp solution. For calcium imaging, pressure-injected SR101, which was dissolved at 0.1 mM in Fura 2-containing solution, was imaged posthoc to confirm the recorded area. Recordings were initiated after recovery from anesthesia, which was confirmed by spontaneous whisker movements and touch-induced eye blink reflexes. The total periods of recording were restricted to less than 1 h to minimize stress in the animals.
Visual stimuli were generated in custom-written MATLAB routines (The MathWorks, Natick, MA, USA) with Psychtoolbox extensions. A 17-in TN-LCD monitor (refresh rate = 60 Hz) was placed 30 cm away from the right cornea, so that it covered 38.8° horizontally and 29.6° vertically of the mouse visual field. For flash stimulation, sinusoidal gratings (spatial frequency: 0.16 cpd; temporal frequency: 2 Hz; contrast: 100%) were presented in four evenly spaced orientations (0°, 45°, 90°, and 135°). The flash duration was set to range between 17–50 ms. Measurement using a high-speed CMOS camera (ORCA-Flash2.8, Hamamatsu, imaged at 2,000 Hz) revealed that a flashing light on the TN-LCD monitor decayed with a time constant τ1/2 = 5.5 ms, and thus, the afterglow was virtually ignorable. For each orientation, the gratings were presented at 2–4 spatial phases, and the responses were averaged to remove the effects of spatial phases. Flash stimuli were intervened with a gray screen for intervals of 8–10 s. In each set, stimuli with four orientations were presented in a pseudorandom order, and the set was repeated 10–40 times. For drifting grating stimulation, sinusoidal gratings (spatial frequency: 0.12 cpd; temporal frequency: 2 Hz; contrast: 100%) moved toward eight evenly spaced directions (0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°) for 1.5 s at intervals of 8–10 s for electrophysiology and for 2 s at an interval of 6 s for calcium imaging. A gray screen was shown during the interval period. In each set, drifting stimuli with eight directions were presented in a pseudorandom order, and the set was repeated 10–40 times. In the Flash+Drift trials, each flash stimulus was followed by a drifting grating stimulus at an SOA of 0.5 s. In S6 Fig, the flash stimuli were fixed at the vertical orientation (0°, vFlash), whereas in Fig 4 the drifting gratings were fixed at the vertical orientation (0°, 180°, vDrift) and moved rightward or leftward.
The procedures for in vivo VSD imaging have been previously described in detail [33,80]. The dye RH-1692 (Optical Imaging, New York, NY) [81] was dissolved in 4-(2-hydroxyethyl)-1-piperazineethanesulphonic acid (HEPES)-buffered saline solution (0.6 mg ml−1) and applied to the exposed cortex for 60–90 min, which stained all neocortical layers. Imaging was initiated approximately 30 min after washing the unbound dye. To minimize movement artifacts because of respiration, the brain was covered with 1.5% agarose made in HEPES-buffered saline and sealed with a glass coverslip. For data collection, 12-bit images were captured at 6.67-ms temporal resolution with a charge-coupled device camera (1M60 Pantera, Dalsa, Waterloo, ON) and an EPIX E4DB frame grabber with XCAP 3.1 imaging software (EPIX, Inc., Buffalo Grove, IL). RH-1692 was excited with red LEDs (Luxeon K2, 627-nm center) and excitation filters of 630 ± 15 nm. Images were obtained with a microscope composed of front-to-front video lenses (8.6 × 8.6 mm field of view, 67 μm per pixel). The depth of field of our imaging setup was 1 mm. RH-1692 fluorescence was filtered through a 673-to-703-nm band-pass optical filter (Semrock, New York, NY). Visual responses were averaged from 40–80 trials of stimulus presentations. Responses to flashes were expressed as the percent change in RH-1692 fluorescence relative to the baseline fluorescence intensity (ΔF/F0 × 100%). Gating flashes were applied to the retina at a distance of approximately 10 cm from the cornea contralateral to the recording site to cover the entire optic angle. Stimulation was repeated every 10 s.
The signal was amplified with a MultiClamp 700B, analyzed with pCLAMP10.1 (Molecular Devices, Union City, CA, USA) and digitized at 20 kHz. The data were reduced to 2 kHz and off-line analyzed using custom-written MATLAB routines. Patch-clamp recordings were obtained from L2/3 neurons at depths of 150–350 μm from the V1 surface using borosilicate glass electrodes (3.5–6.5 MΩ) that were pulled with a P-97 puller (Sutter Instruments, Novato, CA, USA). The electrode tips were lowered perpendicularly into the V1 with a DMX-11 electric manipulator (Narishige, Tokyo, Japan) or obliquely (at 30°) with a PatchStar micromanipulator (Scientifica, Uckfield, UK). For cell-attached recordings, pipettes were filled with aCSF. For whole-cell recordings, the intrapipette solution consisted of the following (in mM): 130 K-gluconate, 10 KCl, 10 HEPES, 10 Na2-phosphocreatine, 4 Mg-ATP, 0.3 Na2GTP, 0.05 Alexa-594 hydrazide, and 0.2% biocytin, adjusted to pH 7.3. For morphological reconstruction of the recorded cells, mice were perfused transcardially with 4% paraformaldehyde, and their brains were coronally sectioned at a thickness of 200 μm using a DTK-1500 vibratome (Dosaka, Kyoto, Japan). The sections were incubated with 0.3% H2O2 for 30 min and permeabilized with 0.2% Triton X-100 for 1 h. Then, the sections were processed with ABC reagent at 4°C overnight and developed with 0.0003% H2O2, 0.02% diaminobenzidine, and 10 mM (NH4)2Ni(SO4)2. Experiments in which the series resistance exceeded 70 MΩ or changed by more than 15% during the recording session were discarded. For each neuron, spike responses to a brief inward current were examined, and regular spiking neurons were selected as putative pyramidal cells for the subsequent analyses. LFPs were recorded at a depth of 300 μm from the V1 surface, which corresponded to L2/3, using borosilicate glass pipettes (1−2 MΩ) filled with aCSF. Traces were band-pass filtered between 1 and 250 Hz.
Ten healthy adults (four males and six females, 25.9 ± 5.4 (mean ± SD) years old) with normal or corrected-to-normal vision participated in our EEG experiments. The EEG experiment was conducted in a dark room to explore early and late components of the visually evoked ERPs for brief exposures to high-contrast grating stimulus flashes. Visual stimuli were generated on a computer using Psychophysics MATLAB toolbox [82]. The stimuli were presented using a gamma-corrected [83] LCD display (EIZO FlexScan S2243W, EIZO corporation, Ishikawa, Japan) whose spatial resolutions were 1,920×1,200 pixels, and the refresh rate was 60 Hz. Participants viewed the stimuli at a 55-cm distance from the display. The experiment contained two stimulus conditions (vertical and horizontal gratings), and the EEG signals for each of the stimuli were acquired 200 times (100 for the horizontal grating and 100 for the vertical grating). In each trial, the start of the trial was informed by the change of the color of the central fixation point (from gray to white). After 3–4 s (randomly jittered to exclude participant’s expectation effect on the EEG signals) of the fixation color change, a high-contrast (100% from the gray background) grayscale sinusoidal grating (1.03 cycles per degree) pattern (35.2 × 24.4° in visual angle) was flashed for 50 ms. The background brightness was 17.80 cd/m2, which corresponds roughly to 4.88 lux, and the grating brightness ranged from 0.26 cd/m2 (0.07 lux) to 35.62 cd/m2 (9.77 lux). Then, participants were asked to keep fixating the central fixation for 4 s without blinking as much as possible. After the 4-s fixation period, the central fixation color changed from white to gray to inform the end of a trial. The task start was initiated by a button press by a participant. The participants could take breaks between trials as they liked, and they could proceed the experiments at their own paces. The stimulus presentation order was pseudorandomized for each participant. One EEG session took about 2 h. The human visual ERPs at O1 and O2 (following the international 10/20 coordinate convention) for the two stimulus configurations were collected at 1 kHz (the left earlobe was used as a reference) with a wireless EEG system (Polymate Mini AP108, Miyuki Giken Co., Ltd, Tokyo, Japan) with pasteless dry electrodes (National Institute of Information and Communications Technology, Japan) [84]. Electrode impedances for O1 and O2 were kept below 5 kΩ at the beginning of the measurements. Eye movements and blinks were simultaneously recorded with an electrode put on a left eye lid. The onset of the visual stimulus presentation and the EEG measurements were synchronized using a customized photo-trigger detection system (C6386, Hamamatsu Photonics K.K., Shizuoka, Japan). The recorded EEG and eye blink–related signals were saved on a computer using in-house MATLAB subroutines after each trial through a Bluetooth wireless connection. The ERP time series were analyzed using EEGLAB MATLAB toolbox ([85], http://sccn.ucsd.edu/eeglab/) and in-house subroutines written in MATLAB. The EEG signals were aligned offline so that we could evaluate the time series from −200 ms to 3,000 ms relative to the grating stimulus onset. The raw data were preprocessed offline by a linear trend removal and a band-pass filtering (0.5 to 100 Hz). Additionally, EEG epochs that contained large potentials exceeding the threshold (40 μV) and abnormal spike or drifting components were excluded by EEGLAB’s automatic outlier detection utilities and visual inspections. These noisy epochs were generally derived from eye movements and blinks. The signal amplitudes were recomputed carefully by taking the mean of −200 to 0 ms (relative to the stimulus onset) samples as the baseline for each epoch. The recorded signals from two electrodes were similar and hence averaged for each participant. Finally, the ERPs averaged over 10 participants were given as the final visual event-related time series. The statistical tests to explore whether the signals were higher or lower than the baseline were evaluated by the standard two-tailed t test at each sampling point (p < 0.05 without corrections of multiple comparisons).
The OSI was defined according to the following equation:
OSI=(∑Rθsin2θ)2+(∑Rθcos2θ)2∑Rθ
where Rθ is the mean response amplitude to a grating with direction θ [86]. Note that this equation defines the normalized norm of the averaged vector [86] and may give a value that is different from OSI used in other reports [41]. The similarity of the tuning curves between the early and late responses was evaluated using the correlation coefficient (R) of the amplitudes of the responses:
R=∑(Rθ_early−R¯θ_early)∑(Rθ_late−R¯θ_late)∑(Rθ_early−R¯θ_early)2∑(Rθ_late−R¯θ_late)2
where Rθ_early and Rθ_late are the amplitudes of early and late responses, respectively, to a grating with direction θ. R-θ_early and R-θ_late represent the mean of the response amplitudes Rθ_early and Rθ_late across all eight θs. For each cell, the OSI and R were compared with their chance levels, which were estimated using a conventional random resampling method in which 1,000 surrogates were generated by randomly shuffling all trials irrespective of θ.
The mouse was placed in a stereotaxic frame and then on the stage of an upright microscope (BX61WI; Olympus). Cortical neurons were loaded with Fura 2, a calcium-sensitive fluorescent dye, under online visual guidance with a two-photon laser scanning microscope (FV1000; Olympus). Fura 2 AM was dissolved at 10 mM in DMSO with 10% pluronic acid and diluted at the final concentration of 1 mM in aCSF that contained 0.1 mM SR101. This solution was pressure-injected (50–100 mbar for 10 s) into V1 at a depth of 150–250 μm from the surface through a glass pipette (tip diameter: 10–30 μm). The pipette was carefully withdrawn, and the craniotomized area was sealed with 2% agar and a glass cover slip. After 50–70 min, which enabled the dye loading to the neuronal soma and the washout of extracellular dyes, the Fura-2 fluorescence was two-photon imaged from V1 L2/3 neurons. Neurons and astrocytes were discriminated based on astrocyte-specific staining with SR101 [87]. Fura 2 and SR101 were excited by a mode-locked Ti: sapphire laser at wavelengths of 800 nm and 910 nm, respectively (100 fs pulse width, 80 MHz pulse frequency; Maitai HP; Spectra Physics) [88]. Fluorescent light was corrected by a water-immersion objective lens (20×, numerical aperture 0.95; Olympus). Videos were taken from a 320×320-μm area at five frames per s using FV10-ASW software (version 3.0; Olympus). Neurons that exhibited significant visual responses above the baseline (p < 0.05, paired t test) in any recording session were selected for analysis.
The apparatus was located in a dark, soundproofed room. The room temperature was maintained at 25°C during the experiment. A virtual cylinder comprising a vertical sinusoidal grating (0.17 cpd, 10%–40% contrast) was displayed in three-dimensional coordinate space on four 24-in monitors (refresh rate: 60 Hz) that were arranged in a quadrangle arena. The images on the monitors were extended by two mirrors on the top and bottom of the arena. A platform (a white acrylic disc; ϕ = 6.0 cm) was positioned 13.5 cm above the bottom mirror. In each experiment, a single male P28–35 C57BL/6J mouse was placed on the platform and was allowed to move freely. The behavior of the mouse was monitored through a camera (Logicool HD Webcam C615; Logitech, Tokyo, Japan) that was attached over a small hole of the top mirror. Vertical gratings that drifted leftward or rightward (temporal frequency: 0.5 Hz) were presented simultaneously on all four screens for 2 s with a random interval between 2–4 s. From the animal’s point of view, the virtual cylinder appeared to rotate around the platform at an angular velocity of 5° per s). The mice normally tracked the grating with reflexive head movements in concert with the rotation direction. The drifting directions were randomly alternated, and the rotations were repeated 120 times in one session that took approximately 10 min. In some trials, either a vertical or horizontal grating (0.17 cycles per degree, 100% contrast) was flashed 0.5 or 3 s before a drifting grating. Animals were habituated to the system prior to the first behavioral test by experiencing at least one full session. When the mice slipped or jumped down from the platform during the test, they were manually returned to the platform, and the test was resumed. If the animal’s head tracked a cylinder rotation, the trial was counted as a “success.” Manual counting was checked by two independent trained researchers who were blind to the experimental conditions. Through computer-generated order randomization of the stimulation conditions, the experimenters were also blind to the treatment. The trials in which a mouse was grooming or made large movements were excluded from the analyses (invalid trials). The success rate was calculated as a ratio of the successful trials to the total valid trials. Tetrodotoxin was dissolved at 10 μM in aCSF and directly applied to the cortical surface 15 min prior to the behavioral sessions. The exposed cortices were covered with the craniotomized bone segments and mounted with dental cement. The effects of tetrodotoxin were confirmed by flash-induced LFP responses in V1 L2/3.
Eleven healthy right-handed individuals (three females) with normal or corrected-to-normal vision participated. The ages ranged from 22 to 42 years, with 26.5 ± 5.1 years (mean ± SD). The participants performed tasks using a computer mouse with their right hands. A 24-in monitor was placed at a distance of 0.5 m from the participants’ eyes in a dark, pseudosoundproofed room. The participants were instructed to report the motion direction of drifting gratings presented on the screen. A 2 × 2 cm2 open square was displayed at the center of the screen against a gray background (60 cd/m2, 5 lux). Each trial was initiated when a participant clicked the computer mouse on the square. Then, the square was filled in black, and after a random time interval between 1–3 s, a sinusoidal drifting grating (spatial frequency: 0.12 cpd; temporal frequency: 1 Hz; contrast: 40%) was presented for 0.25 s in one of four movement directions (0°, 90°, 180°, and 270°). A 50-ms beep tone was presented 0.5 s before a drifting grating stimulus. In some trials, a 50-ms grating flash (spatial frequency: 0.12 cpd; contrast: 100%) was displayed simultaneously with the tone. A full gray screen was displayed during all interstimulus intervals. After each stimulus, the participants were asked to move the mouse cursor in the same direction as the grating motion as rapidly as possible. When the mouse cursor traversed the edge of the square, the square became blank, which cued the trial completion. Incorrect motion reports or failures to respond within 600 ms (misses) from stimulus onset were considered errors and were indicated to the participants through a 200-ms peep tone. Each participant performed 160–244 trials per session.
|
10.1371/journal.pcbi.1006933 | An integrative transcriptome analysis framework for drug efficacy and similarity reveals drug-specific signatures of anti-TNF treatment in a mouse model of inflammatory polyarthritis | Anti-TNF agents have been in the first line of treatment of various inflammatory diseases such as Rheumatoid Arthritis and Crohn’s Disease, with a number of different biologics being currently in use. A detailed analysis of their effect at transcriptome level has nevertheless been lacking. We herein present a concise analysis of an extended transcriptomics profiling of four different anti-TNF biologics upon treatment of the established hTNFTg (Tg197) mouse model of spontaneous inflammatory polyarthritis. We implement a series of computational analyses that include clustering of differentially expressed genes, functional analysis and random forest classification. Taking advantage of our detailed sample structure, we devise metrics of treatment efficiency that take into account changes in gene expression compared to both the healthy and the diseased state. Our results suggest considerable variability in the capacity of different biologics to modulate gene expression that can be attributed to treatment-specific functional pathways and differential preferences to restore over- or under-expressed genes. Early intervention appears to manage inflammation in a more efficient way but is accompanied by increased effects on a number of genes that are seemingly unrelated to the disease. Administration at an early stage is also lacking in capacity to restore healthy expression levels of under-expressed genes. We record quantifiable differences among anti-TNF biologics in their efficiency to modulate over-expressed genes related to immune and inflammatory pathways. More importantly, we find a subset of the tested substances to have quantitative advantages in addressing deregulation of under-expressed genes involved in pathways related to known RA comorbidities. Our study shows the potential of transcriptomic analyses to identify comprehensive and distinct treatment-specific gene signatures combining disease-related and unrelated genes and proposes a generalized framework for the assessment of drug efficacy, the search of biosimilars and the evaluation of the efficacy of TNF small molecule inhibitors.
| A number of anti-TNF drugs are being used in the treatment of inflammatory autoimmune diseases, such as Rheumatoid Arthritis and Crohn’s Disease. Despite their wide use there has been, to date, no detailed analysis of their effect on the affected tissues at a transcriptome level. In this work we applied four different anti-TNF drugs on an established mouse model of inflammatory polyarthritis and collected a large number of independent biological replicates from the synovial tissue of healthy, diseased and treated animals. We then applied a series of bioinformatics analyses in order to define the sets of genes, biological pathways and functions that are affected in the diseased animals and modulated by each of the different treatments. Our dataset allowed us to focus on previously overlooked aspects of gene regulation. We found that the majority of differentially expressed genes in disease are under-expressed and that they are also associated with functions related to Rheumatoid Arthritis comorbidities such as cardiovascular disease. We were also able to define gene and pathway subsets that are not changed in the disease but are, nonetheless, altered under various treatments and to use these subsets in drug classification and assessment. Through the application of machine learning approaches we created quantitative efficiency profiles for the tested drugs, which showed some to be more efficiently addressing changes in the inflammatory pathways, while others being quantitatively superior in restoring gene expression changes associated to disease comorbidities. We thus, propose a concise computational pipeline that may be used in the assessment of drug efficacy and biosimilarity and which may form the basis of evaluation protocols for small molecule TNF inhibitors.
| In an era of unprecedented accumulation of biomedical data, our understanding of the mechanisms of the development of complex diseases is greatly enabled by the performance of high-throughput experiments and their subsequent analyses at various levels that range from single genes to biological pathways, modules and networks [1, 2]. Genome-wide transcriptomic profiling has been instrumental in providing accurate representations of the expression programs in homeostasis and disease as well as before and after pharmaceutical interventions [3–7]. Among various pathological conditions, inflammatory diseases such as Rheumatoid Arthritis (RA) present great challenges in the understanding of the process through which an initial trigger may lead to generalized and highly variable changes at molecular, cellular and eventually organ level [8]. In this respect, animal models have proven invaluable in the detailed study of these intricate mechanisms and have been the choice of preference in many studies due to particular advantages such as accessibility of material, robustness and standardization [9,10].
Recently, an attempt to assess the differential properties of different substances used in the treatment of RA was performed at whole blood transcriptome level of human patients [11], producing different gene signatures that reflected the mechanistic differences of the tested biologics (an anti-TNF, an anti IL6-R and an inhibitor of T-cell maturation). Among the various therapeutic agents used in the treatment of RA, anti-TNF antibodies have been the primary line of defense since TNF was shown to be a major driver of the disease [12]. In this context, a concise analysis of the effect of the different anti-TNF biologics at transcriptome level has been lacking. The need for investigating the variability in differential patterns of anti-TNF agents [13] has been supported by findings at various levels that include the stratification of RA subtypes [14], cell-type dependent responses [15] and variability among cells of the same type, namely fibroblast-like synoviocytes [16,17].
In this work, we present a computational analysis [18–24] based on a standardized and highly robust protocol involving the administration of four different anti-TNF agents in the hTNFTg (Tg197) humanized mouse model of RA, that has been essential in proving the central role of TNF in the arthritis pathology and its validity as a major therapeutic target [12]. The large number of analyzed profiles and the incorporation of both untreated and healthy samples in our study, enables us to go beyond a simple recording of differentially expressed genes and enriched pathways, which are the usual output of such analyses. Instead, we present a framework for the assessment of drug similarity at various functional levels through the implementation of a combination of gene clustering and state of the art classification methods. Treatment profiles were analyzed at two levels, focusing first on genes that are associated to the disease and then on those that are altered by the treatment even if unchanged in transgenic (diseased) animals. In this way we are able to pinpoint a number of functional attributes that are treatment-specific and accordingly devise measures of profile similarity between treatments and the healthy state. Our results suggest subtle, yet significant differences between the different biologics as well as between different intervention timing (early or late).
Our work provides a general framework for the comparison of treatment-specific transcriptomes that may assist in a) the detailed profiling of the gene and functional modules addressed (or left unaffected) by a given treatment b) the multi-level assessment of the treatment’s efficiency in restoring gene expression levels and c) similarity searches between different pharmaceutical agents. Implementation of our analyses may thus lead to interesting applications in the search of biosimilarity and specify suggestions in the evaluation of small molecule inhibitors.
The hTNFTg (Tg197) human TNF transgenic model develops chronic inflammatory polyarthritis with 100% incidence and with clinical manifestations and histological findings very similar to those of the human disease [12] becoming evident as early as 3 weeks of age with signs progressively worsening as animals age. Animals were treated in a therapeutic regimen from week 6 of age when pathology is already established or in a prophylactic regimen from week 3 of age when pathology is at an early stage. Groups of mice of the same age (gender balanced) received either saline or one of the following anti-TNF agents: a) the chimeric monoclonal antibody infliximab (Remicade, Janssen Biotech), b) the pegylated antigen binding fragment (Fab) certolizumab pegol (Cimzia, UCB) c) he fully human monoclonal antibody adalimumab (Humira, Abbvie) and d) the fusion protein TNFR2/IGg1Fc etanercept (Enbrel, Pfizer). The first three were administered at 10 mg/kg intraperitoneally twice weekly, while etanercept was administered subcutaneously at 30 mg/kg thrice weekly. Apart from a standard therapeutic stage administration, infliximab was also administered with the same dosage but starting at an earlier stage (3 weeks). During a treatment period of 6 weeks mice were regularly monitored and scored for the progress of the disease pathology and for their overall health status. All therapeutic treatments showed very similar disease progression patterns with an overall in vivo arthritic score dropping steadily from the onset of the intervention, suggesting similar patterns of remission of the hTNFTg arthritis pathology (Fig 1A).
Sets of Differentially Expressed (DE) genes were defined on the basis of standard thresholds i.e. a |log2FC|≥1 combined with an adjusted p-value≤0.05, against either wild-type (WT) or transgenic (TG) samples. The cutoffs used, corresponded to values lying beyond two standard deviations (+/-2σ) in a symmetrical log2FC distribution and resulted in the selection of <5% of the total genes. We then combined DE genes in various lists depending on the initial comparison undertaken and focused our analyses on the following DE gene subsets:
For the initial part of our analysis we focused on the first subset of the 867 TG/WT DE genes, which was used as a disease-associated gene set that allowed us to monitor the functions and pathways that are mainly affected in the diseased animals regardless of pharmaceutical intervention. The more extensive sets of 1064 and 1338 DE genes were used in further analyses of treatment-specific properties, since they included genes whose expression was modulated by the treatment even if they remained unchanged in the diseased animals. They thus represented more inclusive lists comprising genes that are not directly related to the diseased state.
Expression patterns of the 867 disease-associated genes in treated and untreated hTNFTg against wild-type samples are shown in Fig 1A. One can see that the DE genes are roughly equally divided in over- (404) and under-expressed genes (463) in the diseased versus wild-type state. The slight predominance of under-expressed genes is a distinguishing feature of our analysis compared to previous attempts on human samples, where the lack of normal controls often shifts the balance to over-expressed genes in ratios that vary from 3:1 to 6:1 [25–27]. The large number of under-expressed genes points towards a number of functions that are associated to the diseased state but are not necessarily linked to known, implicated immune and inflammatory pathways.
Focusing on particular treatment profiles, we found that the prophylactic intervention with infliximab was the one that most readily restored expression to wild-type levels. This was expected due to the time of intervention at 3 weeks of age and prior to full disease development. On the other hand, in this particular intervention, gene expression changes were in many cases over-compensated, leading to expression levels being reversed, that is genes over-expressed in diseased mice, being under-expressed in treated samples and vice versa.
Among the therapeutic interventions, infliximab appears to be the treatment with more direct effects on over-expressed genes, while adalimumab, etanercept and certolizumab pegol present great similarity in terms of expression levels, showing increased restoring potential for a large set of under-expressed genes. Differences between infliximab and the rest of the treatments were primarily quantitative, an observation that is in accordance with a mild advantage of infliximab in the clinical data (see Fig 1A). For specific subsets of genes, restoration to normal (wild-type) levels was more extensive with infliximab compared to the other three substances.
Based on the expression patterns of treated and untreated mice, we clustered the 867 disease-associated genes in 8 distinct clusters, a number that was indicated by the application of a Silhouette consistency test (see Materials and methods). Of the 8 clusters, 5 contained genes that were under-expressed in disease (463 genes), while 3 included a total number of 404 over-expressed genes, as may be seen in the coloured side bar of Fig 1B. A more detailed inspection of the differential expression values for each cluster from Fig 1B reveals different patterns of response for the different treatments.
Among the three over-expressed gene clusters, the first (light blue) shows all treatments sufficient to bring expression down to wild-type levels. The second cluster (dark green) contains genes for which prophylactic intervention drives a partial reversal of gene expression levels, while therapeutic treatment is more or less comparable regardless of the administered biologic, with a slight but quantifiable advantage for infliximab. The third cluster (yellow) comprises the subset of the most over-expressed disease-associated genes (compare with Disease state in Fig 1B) which are, expectedly, partially restored to normal levels by therapeutic interventions but are, interestingly, strongly reversed in the prophylactic one (see also S1 Fig). This reversal may indicate that the subset of the most over-expressed genes is probably not present, or at least not fully formed before later stages of the disease. Thus, it contains genes, which are reversed when the treatment is administered early but not fully addressed at a later stage, an observation which may bear significant implications on the dynamics of the disease.
Under-expressed genes are clustered in five groups of variable size. Going from top to bottom (Fig 1B), for two of the five (Fig 1B, shown in orange and red) there is a quantitative difference in gene expression restoring potential between adalimumab, etanercept and certolizumab pegol against infliximab. What is also worth mentioning, is that for these clusters infliximab shows limited capacity to modulate gene expression at prophylactic intervention. This may be indicative of different expression dynamics between over- and under-expressed genes, with the first being deregulated earlier and thus addressed at a prophylactic stage while the latter being associated with a more delayed onset. A third under-expressed gene cluster (dark blue) comprises 61 genes for which gene expression is partially restored with all treatments, while a fourth, shown in brown at the bottom of Fig 1B shows partial resetting of gene expression levels with therapeutic interventions and a reversal (from under- to over-expressed) in the case of the prophylactic infliximab treatment. Lastly, a small cluster of only 12 genes (Fig 1B pink), contains the genes with the most acute under-expression patterns, which seem to be preferentially addressed by infliximab at prophylactic intervention but not by the rest of the treatments. The small number of these genes (12) does not allow for an enrichment analysis but in general they correspond to genes associated with collagen and myofibrils (Col10a1, Glt25d2, Xirp2, Myl2 and Clec3a) innate immunity (C7, Cytl1, Vsig4), the phospholpases Pla1a and Pla2g2a, the serine peptidase Htra4 and the developmental protein Hhip (heghehog interacting protein).
In order to gain insight on the functional characteristics of the underlying genes we performed a functional enrichment analysis individually for each cluster. Functional enrichments were calculated with the implementation of gProfiler [20] as described in Methods and a summary of the most enriched terms at the levels of Gene Ontology, KEGG pathways and Transcription Factor targets is shown in Fig 2A and 2B.
Two main observations stem from the enrichments of over- and under-expressed gene clusters respectively. On one hand, terms related to the inflammatory and immune responses, including cytokine and chemokine signaling as well as infectious pathways and the associated transcription factors are prominent in all three over-expressed gene clusters (Fig 2A). On the other hand, there is a characteristic over-representation of functional terms related to muscle and heart functions and associated diseases, as well as metabolic and developmental pathways among the under-expressed genes (Fig 2B).
Heart failure and related cardiovascular diseases are known comorbidities of Rheumatoid Arthritis patients [28–30], while we recently showed this to be the case for the animal model under study [31] It is thus interesting to see that associated functions are linked with the under-expression of certain genes, which, as noted above, have been largely overlooked in human studies. In this respect, the aforementioned differential capacity of the analyzed anti-TNF agents to effectively restore under-expressed genes, could lead to more detailed insights on treatment-specific efficiency regarding indirect and secondary effects of the disease.
As a next step we wanted to assess the level of potential for gene expression resetting that may be achieved by each treatment. For this we employed a distance-from-WT calculation (see Materials and methods) for all disease-associated genes and for each of the gene expression clusters defined above. Mean absolute log2FC values were calculated for each cluster and for each treatment and the distance from WT was then used as an indication of treatment efficiency. The results are summarized in Fig 1C, where distances from WT are shown in the form of a clustered heatmap.
Overall mean distances (all genes, black) show small differences between treatments with infliximab lying closer to the healthy state, regardless of time of intervention. When looking into separate clusters one can observe that clusters associated with over-expressed genes, and with immune and inflammatory responses, respond better to infliximab at prophylactic stage. For the same over-expressed clusters, infliximab seems to have a small quantitative advantage at the therapeutic stage, over the rest of the treatments. On the other hand, adalimumab in particular but also etanercept and certolizumab pegol, are more readily restoring gene expression levels of under-expressed gene clusters, an effect that is especially strong for two of the largest clusters (orange and red) which are functionally associated with heart-related functions and diseases (see Fig 2B).
This similarity of the three agents is also reflected in the treatment clustering, in which they are placed together in a group that is separate from inflixιmab. It thus seems that one particular property discriminating infliximab from the rest of the other three anti-TNF agents is related to its increased efficiency in modulating the expression of inflammatory and immune-related genes that are up-regulated in disease. On the other hand, adalimumab, etanercept and certolizumab pegol show greater capacity in restoring gene expression levels for under-expressed genes associated with RA comorbidities.
When assessing pharmaceutical interventions, one of the main aspects we need to address is general changes that are not directly associated with the condition under treatment. In the context of our transcriptomic analyses this meant looking into the genes, whose expression changed between treatment and wild-type profiles but which were not altered in the diseased state, as well as into genes that were changed between treatment and diseased samples in general. In this sense, we may divide differentially expressed genes in three groups:
Analysis of the corresponding genes belonging to each category resulted in a total of 1338 genes that were unequally distributed among treatments (S2 Fig). At a first-level quantitative assessment based on the number of genes, infliximab performs better in terms of restoration to wild-type levels, at both prophylactic and therapeutic interventions, with the numbers of non-restored genes being only 3 and 25 (out of a total of 867) respectively, compared to significantly higher numbers of non-restored genes for adalimumab (92), etanercept (106) and certolizumab pegol (175). When assessing the treatments through their distance from WT for this extended set of genes, it is infliximab at therapeutic intervention which now shows the highest efficiency (mean distance = 0.47) followed by the same substance when administered at prophylactic stage (mean distance = 0.53). Among the three remaining biologics differences are becoming clearer with adalimumab performing considerably better (mean distance = 0.61 compared to etanercept’s 0.72 and certolizumab’s 0.78) (S3 and S4 Figs). It may be noted here that, variability in human patients’ response to adalimumab treatment has been attributed to differential levels of cytokine expression in the synovium [32], which may reflect the quantitative effect compared to infliximab that we are observing.
At this first, purely quantitative level, our results are indicative of a more direct response at gene expression level for infliximab (particularly at prophylactic intervention) and adalimumab compared to etanercept and certolizumab pegol. When one looks at the altered genes, a time effect appears to become important, with infliximab at prophylactic intervention showing a greater number of genes (170) which is comparative to certolizumab (175) and etanercept (217), the latter being the therapeutic treatment with the highest number of non-disease associated modulated genes. Indeed, the time of intervention appears to be crucial since the same substance (infliximab) when administered at a later stage (6w) has practically no altered genes, suggesting maximal specificity. Overall, infliximab and adalimumab at a therapeutic stage of intervention show the greater number of restored genes combined with a small number of altered genes and thus appear, at this level, to reflect the most effective gene expression modulation patterns.
In order to gain more detailed insight on how the response to treatment may be compared to the diseased state, we plotted gene expression values for the three gene categories (Restored: green, not Restored: red, Altered: blue) in two dimensional scatterplots shown in Fig 3A. Here, the expression of each gene in disease is shown on the horizontal axis and the corresponding value for the treated samples is shown on the vertical one. In this respect, a good overall response will have points distributed on a horizontal line around 0 (no change compared to wild-type), while points that fall away from the horizontal baseline will represent genes that are either non-responsive to treatment (red) or genes that are altered (blue). The diagonal dashed line corresponds to identical values in treatment and disease, therefore linear trendlines, based on the total number of genes, are representative of the overall response profile of the treatment.
Slopes greatly deviate from 1 (diagonal), with infliximab again showing the smallest slope values, closer to the desired horizontal. Infliximab at prophylactic intervention shows the smallest slope value, but this is mostly due to the small deviation of the altered genes. A close inspection of Fig 2A shows that altered genes (blue) for infliximab at prophylactic intervention fall on the opposite side of the trendline, which is indicative of gene expression reversals. One may observe this in Fig 3B, where (on the right) the gene expression values for the genes altered in this particular treatment show a strong reversal of expression levels against the diseased animals. Thus, a significant number of genes seem to be affected in a very acute way when intervention takes place early, as may be seen with a direct comparison to the infliximab therapeutic treatment (Fig 3B, right).
We next performed a detailed analysis of partial overlaps of not restored (Fig 3C) and altered genes (Fig 3D). Genes not restored to normal levels do not appear to be treatment-specific, as may be seen in extensive overlaps among all treatments (Fig 3C, layer of black dots spanning most conditions). Certolizumab pegol and etanercept in particular, share the highest number of not restored genes (88 genes in total, 20 of which are specific between them). Infliximab has 25 genes whose levels are not restored, 21 of which are shared with adalimumab. As may be seen in Fig 3A, these genes (red dots) are primarily over-expressed in disease and even though their deregulation is ameliorated upon treatment they remain quantitatively active, above the thresholds that qualify them as differentially expressed. This quantitative effect may explain the functional enrichments of not restored genes shown in the table accompanying Fig 3C. With the exception of infliximab, whose 25 not restored genes are enriched in functions related to the extracellular matrix and its degradation, the other three treatments show enrichments in immune, inflammatory and infectious pathways (see also S5 Fig). Even though this seems, at first, counter-intuitive it is supported by the results in Fig 3A and 3B where one can see that there is a quantitative lag in restoring the expression of disease-associated genes. Thus, even though treatments may show similar macroscopic effects (Fig 1A), they have quantifiable differences at the molecular level (S4 Fig), an observation that highlights the importance of -omics approaches in the analysis of detailed treatment profiles.
An identical gene overlap and functional analysis was carried out for altered genes (Fig 3D). In the case of altered genes, effects appear to be largely treatment-specific, with most of the altered genes belonging to particular treatments (compare layers of dots in 3C and 3D). As noted earlier, the most striking observation is the significant number of altered genes in the case of infliximab prophylactic intervention (170). Of these, 164 are exclusive to the particular treatment, which is indicative that they are largely associated to the age and the different dynamics of gene expression of younger mice that have not yet manifested the disease in its full proportions. These differences are also reflected in the functional analysis where terms related to development, cell motility and fatty acid and lipid metabolism are the ones primarily enriched. What is particularly interesting for altered genes of the infliximab prophylactic treatment, is their expression levels, which show a strong reversal compared to the disease state (Fig 3B, right). This is yet another indication of the strong effects of the prophylactic intervention, effects that extend beyond the disease-associated gene set.
Altered genes of adalimumab, etanercept and certolizumab are also largely treatment-specific (Fig 3B, right). Etanercept and certolizumab share a significant percentage due to the overall large numbers of altered genes (217 and 175 respectively). At the functional level, however, differences are also apparent, with certolizumab’s genes being enriched in structural proteins and oxidative metabolism, etanercept’s being very strongly associated with skin functions and differentiation and adalimumab having only a handful of enriched pathways related to structural proteins.
Together, the analysis of not restored and altered genes reveals differences at both qualitative (gene lists and functions) and quantitative (extent or reversal of gene expression levels). Therefore, accurate representations of treatment regimens at transcriptomic level require the analysis of such inclusive gene sets.
We set out to build a classification model that would preferably discriminate between the treatments and the diseased samples, while at the same time, provide us with insight on the most important genes and pathways. We chose a Random Forest approach since it combines both such characteristics and implemented it on the total number of profiles, including wild-type and transgenic [24]. The best out of 1000 models (see Materials and methods) gave a perfect discrimination between wild-type and transgenic profiles (which was expected given their considerable differences) but also led to a significant improvement in the classification of the different treatments.
The results of the Random Forest classification are visualized in Fig 4 first as PCA plots based on the total of 1338 differentially expressed genes (Fig 4A) as opposed to using only the top 100 most important genes based on the Random Forest classification model (Fig 4B). (Notice that given the large number of analyzed features, PCA is used here as an indicative representation of the capacity of the Random Forest model to discriminate between the different conditions). A common characteristic of both PCA analyses is related to a greater variability of transgenic profiles, compared to wild-type and treated samples. Such variability is widely reported in humans and is considered a hallmark of complex diseases [3], however, an even greater number of samples will be required in order to conclude that this variability may be accurately reflected and quantified through our approach. In terms of classification efficiency (Fig 4C) there is a near perfect classification of wild-type and diseased samples, as expected given that the genes under consideration are predominantly differentially expressed between them. Among the various treatments infliximab at both stages of intervention clusters closely to adalimumab. On the other hand, etanercept and certolizumab pegol are classified together in a separate subgroup.
When looking closer at the genes predicted by the model to act as the most important classifiers we find a number of genes involved in functions related to the cell adhesion and the extracellular matrix (Fig 4D). The lack of the dominant infectious, immune and inflammatory pathways is expected since the scope of the classification is to define the genes that would better discriminate between all profiles, including the 5 different treatment regimens. Thus, the genes that achieve the best classification are mostly revealing of inter-treatment differences as may be seen in their expression profiles. When looking into gene expression levels of the top 100 most important genes (Fig 5A) one can see a predominance of genes that are under-expressed in diseased animals since it is among them that the greater variability among treatment becomes manifest.
Fig 4A and 4B suggest that the main premise of the random forest classifier is a distinction between the diseased and wild-type states. As this is likely to be obscuring a better discrimination between treatments, we employed an identical classification approach on a restricted set of samples that included only the treated animals. The results of this classification (S6 Fig) show a marked differentiation of infliximab profiles compared to the three other substances, which now show partial but not complete overlaps. The most important genes in the classification are again primarily associated with functions related to the extracellular matrix, development and cell adhesion (S7 and S8 Figs). Inspection of the gene expression levels of the top 300 most important genes shows that classification is based predominantly on two properties: a) the tendency of infliximab at prophylactic intervention to drive a reversal of over-expressed genes and b) the generally limited capacity of certolizumab pegol to restore over-expressed genes (S8 Fig). Together this first level Random Forest analysis suggests that even though significant differences between treatments can be identified, the most important features are somehow lacking descriptive power. That is the genes that best classify the samples were only partially reflecting the system under study. In order to further dissect these differences, we next employed a Random Forest analysis at functional level.
We applied an identical Random Forest approach at the functional level, starting from a set of differentially expressed genes, mapping them to the corresponding functional categories and then using the mean expression value as input feature for the model (see Materials and methods). By applying the same strategy as the one described above, we were able to define the most important GO, KEGG Pathway and Transcription Factor (TF) categories in the discrimination of treatment profiles. The results from the best model are summarized in Fig 5B where the top 30 most important KEGG Pathways are shown for the best model applied on the full dataset comprising all samples (including healthy and diseased states). When looking at functional level, both pathways (Fig 5B) and transcription factor features (S9 Fig), accurately reflect the system under study with major inflammatory and immune disease pathways and related transcription factors being among the most important predictive characteristics of the model. This indicates that functional enrichment analyses may confer additional, if not superior insights in the system under study, compared to the more detailed inspection of particular genes.
When classifying treatments in the absence of healthy and diseased profiles, thus having removed strictly disease-associated pathways, a first interesting observation is the role of metabolic pathways, which appear to be dominant both at the level of KEGG Pathways (Fig 5C) and Transcription Factor enrichments (S9 Fig). Nitrogen metabolism in particular, including nucleobase and aminoacid rank among the top most important pathways, while, at the level of TF, immune-response transcription factors such as Nfkb have given their place to metabolic, growth and developmental regulators such as Pparg, Atf1 and Hoxc8 (S9 Fig).
In an attempt to capture a more complete picture of treatment efficiency we calculated mean distances from both wild-type and transgenic samples for each set of profiles. In principle, profiles need to be far from the transgenic but close to the wild-type state, but in the actual data a variety of intermediate profiles may be observed. In order to monitor the range of these responses we have devised a simple two-dimensional approach, that aims to capture, describe and visualize differences between treatment profiles at functional level. The approach is generic which means that it can be readily applied to any gene categorization (see Materials and methods for details) but is, herein, restricted to the level of already discussed GO, KEGG Pathway and TF functional categories.
We went on to interpret and visualize these enrichments in the following way: For any given functional category we have calculated a mean distance of expression of each treatment (taking into account only DE genes) against both wild-type and transgenic samples. Then, for a selected list of such functional categories, we plot the distances from both profiles (wild-type and transgenic) in the form of density scatterplots as those shown in Fig 6A. These represent the density of DE genes belonging to the selected functional categories in two dimensions, in which the vertical and horizontal displacements from 0 correspond to the distances from diseased and healthy samples respectively. In this sense, efficient responses are represented by clouds lying around the vertical axis x = 0, with minimal horizontal (low adverse effects) and maximal vertical (high response to disease) values (also see Materials and methods).
Fig 6A shows the combined density plots for the top 30 most important KEGG pathways described by the best model in the Random Forest classification. Infliximab at prophylactic intervention stands out with the most preferable pattern, represented as a cloud along the vertical axis and with a small displacement towards negative values for the distance from wild-type. Differences between the therapeutic interventions may be qualitatively observed in the shape of the density plot. Starting from infiximab and down to certolizumab pegol, the contours flatten out along the horizontal axis pointing to increasing levels of distance from wild-type. All therapeutic interventions share similar patterns with vertical displacements (distance from transgenic) systematically greater than the horizontal ones (distance from wild-type) but with considerable distances from both diseased and healthy states, represented as diagonally diffuse clouds.
This approach may be used for different functional categorizations in order to capture particular characteristics of the response, as may be seen in the case of GO terms or Transcription Factors (S10 and S11 Figs). Again, starting from infliximab the density plots show a gradually increasing displacement along the x-axis, which is representative of insufficient restoration of gene expression to healthy, wild-type levels. Some interesting features of this analysis are related to the persistence of the retinoic acid receptors (S11 Fig) whose targets appear to be invariably under-expressed in transgenic but over-compensated in the treatments. Differences between treatments may be primarily attributed to metabolic regulators such as Pparg and Ppara as well as Atf1, and Atf2, which have been reported as being selectively under-expressed in RA synovial extracts even when compared to osteoarthritis (OA) controls [33].
Two-dimensional density plots may provide a helpful framework for the visualization of the treatment profiles but are not easily quantifiable. Using, the 2D profiles as starting point, we devised a simple measure of response efficiency that aims to capture a combination of a treatment’s potential in a) restoring expression levels that are changed in disease while b) leaving wild-type, healthy levels unaltered. We calculated such an efficiency measure as the log10-transformed ratio of the absolute transgenic over the absolute wild-type distance (see Materials and methods for details), for various subsets of functional categories.
The results for the most important KEGG pathways are shown in Fig 6B, where one may observe particular tendencies of the different responses in great detail. The overall greater efficiency of the prophylactic intervention is again apparent, however there is also a greater consistency in the pattern of infliximab at therapeutic intervention with very few negative scores (which correspond to high frequency of gene expression reversals). Perhaps the most interesting observation from the efficiency analysis comes from the bottom of Fig 6B where one can see increased positive scores for adalimumab, etanercept and, to a lesser extent, certolizumab pegol for a number of functions related to heart disease as well as two metabolic signaling pathways (Adipocytokine and Insulin signaling), for which infliximab shows lower efficiency at both stages of intervention. These results are further supported by more extensive analyses for broader functional sets (Fig 6C) as well as at the level of transcription factor enrichments (S12 Fig). In fact, a recurring observation from various points of our analysis points towards a general pattern for adalimumab and etanercept targeting pathways associated with known RA comorbidities more effectively than infliximab, compared to a less pronounced response against key inflammatory processes which are more readily addressed by the latter (Fig 7).
The development of anti-TNF therapies has been a milestone in the treatment of rheumatoid arthritis. Currently there are 5 different biologics (infliximab, adalimumab, etanercept, golimumab, certolizumab pegol) in the market while novel biologics or biosimilars have also been developed or are under development. Although all of them target the same molecule, they are different in their molecular structures which range from a fusion protein (etanercept), to human (adalimumab) and chimeric (infliximab) mAbs and a PEGylated Fab fragment (certolizumab pegol). Such structural differences may translate to the differences these agents exhibit in antibody-depended cell-mediated cytotoxicity, complement-depended cytotoxicity, capacity to induce apoptosis, ability to neutralize membrane bound TNF or differential inhibition of TNFR1 and TNFR2 signaling [34, 35]. These properties, together with the specific pharmacokinetic and pharmacodynamic profile of each anti-TNF biologic, contribute in shaping their clinical performance profile including their efficacy and safety parameters such as their immunogenicity, as well as the extent to which each specific biologic may affect comorbid pathologies, allergic responses and host defense mechanisms [34, 36–38]. However, comparisons of the different agents exist only in the clinical setting that does not allow head to head or molecular comparisons. With this study we address the need for such comparisons at the mechanistic/molecular level by using an established arthritis model widely used for the preclinical evaluation of anti-TNF therapeutics [12].
To this end we have developed a generalized framework to address differential gene expression in transcriptomic profiles obtained under disease and treatment with 4 different biologics. In the context of RA, we analyze an extended dataset, which consists of a large number of biological replicates and includes both healthy and untreated animals. This enables us to define gene expression changes against both states and thus achieve considerable insight into the way each treatment modulates expression levels. The detail with which the profiles are analyzed allowed us to reveal, previously unreported characteristics related to gene under-expression and off-target responses as well as to pinpoint particular functional attributes that appear, to a great extent, to be treatment-specific. Thus, on one hand, we bring forward hitherto unreported properties of anti-TNF biologics in the context of an established animal model of inflammatory polyarthritis, while, on the other, we describe a concise set of computational analyses for transcriptomic analyses of drug interventions. From the computational analysis point of view, we show that transcriptomics may capture significant differences between anti-TNF treatments that remain largely unobserved at a macroscopic level. Thus, while small differences may be recorded in clinical readouts such as the disease activity scores (which are by definition coarse and subject to noise), more pronounced variability at both quantitative and qualitative levels is revealed through a detailed dissection of gene expression and functional enrichment data, like the one we present in this work. From a computational analysis viewpoint, one particularly interesting aspect is the greater descriptive capacity achieved from the analysis of broader functional terms compared to the “granularity” of genes. As suggested by the last part of our analyses classification at pathway levels provides a more accurate representations of the system under study. These observations are of particular interest when it comes to considering approaches of alignment between animal models and the human condition, as they are suggestive that data integration at hierarchically higher functional levels provides more accurate representations of the conditions under study. Similar systems approaches in human samples, albeit at a smaller scale have shown the increased descriptive power of pathways and modules instead of simple gene signatures [39, 40].
When focusing into the system under study, the response of inflamed synovial tissue to anti-TNF intervention, the results presented in this work point out to a set of functions and terms as immune response, cell communication, cell cycle and signaling already reported for human patients treated with anti-TNF biologics [32, 41] as well as a number of largely overlooked features regarding anti-TNF therapeutic approaches with potentially important implications for the human condition. A first point has to do with the importance of under-expressed genes. Most published works in both humans and animal models focus on over-expressed genes as they are clearly enriched in inflammatory and immune response pathways, targeted by anti-TNF agents. Nevertheless, our data show that more than half of the differentially expressed genes have decreased expression levels compared to healthy controls. These are interestingly enriched in functions that are not directly related to inflammation, such as those associated with heart and muscle development and related diseases as well as secondary metabolic pathways. Such functions are not expected to be directly addressed by anti-TNF action. We find, however, that they are differentially modulated by the substances analyzed, an observation that points to interesting treatment-specific characteristics.
Another interesting aspect regarding under-expressed genes is that they represent a subset of genes and functions for which a prophylactic intervention fails to efficiently restore expression levels. This may be indicative of a time-dependent effect under which an initial wave of over-expression of inflammation genes is followed by a late onset under-expression, which thus escapes the early intervention. The time of intervention is also shown to be important through our analyses of altered genes. Early, prophylactic intervention schemes are unlikely to form part of treatment regimens in humans but are, nonetheless, valuable in the context of animal models as they reveal certain aspects of disease development, that cannot be monitored otherwise. The reduced potential of early intervention to address under-expression may be related to a very different gene expression programme that is characterized by low inflammation. High levels of inflammation in human synovial tissue have been shown to be positive predictors of anti-TNF response [42] and so this could partially explain the shortcomings of the prophylactic treatment.
Last but not least, a number of observations in our work are potentially interesting in the alignment of the mouse with the human condition. The first is the previously overlooked prominence of under-expression of genes related to known RA commorbidities [31] discussed above. A second point is related to the level at which animal models and human patient samples are to be compared. As shown from our random forest analysis, a representative description of the diseased state is better achieved through functional enrichment analysis at pathway level instead of aiming at the definition of gene lists and signatures. In the past we have shown this superiority of pathway level interpretation for the alignment of mouse to human data at both transcriptome and methylome levels [43]. A third point is related to the well-established variability of human patients in terms of both disease severity and response to therapy [13, 16–17]. Even though the number of samples in our study is limited, such variability is replicated with disease samples showing broader expression patterns compared to wild-type controls, as is evident in both the PCA analyses and the distributions of expression values. Perhaps the most interesting implication of our study may be that the significant quantitative and qualitative differences we detect between different anti-TNF agents are underlying the well-studied variations in patient response.
In all, our approach involving head to head comparisons of different anti-TNF biologics aligns to the limited human data and enables us to capture subtle, or more profound differences between anti-TNF agents and to quantify them through an innovative scheme of efficiency scores. We herein show that transcriptomic analyses represent a valuable means for the study of disease mechanisms and the intricate modes of action of specific treatments. The developed computational pipeline may be easily modified and extended to accommodate comparative analyses of drug similarity or small molecule inhibitor efficacy by quantitatively highlighting treatment-dependent discriminatory characteristics. Moreover, such pipelines might be a tool to support the preferential use of a particular agent or class of agents in specific clinical pathology niches thus driving personalized medicine approaches.
WT and human TNF transgenic mice (Tg197) [12] were bred and maintained in a mixed CBA×C57BL/6J genetic background in the animal facilities of Biomedcode Hellas S.A. under specific pathogen-free conditions. Animals were housed in standard plastic cages with wood chip bedding. The animal facility was under an inverted 12:12-h light/dark cycle at a constant temperature of 22 ± 2 °C and relative humidity of approximately 60%. Food pellets and filtered water were provided ad libitum. Experiments were approved by the BSRC Al. Fleming Institutional Committee of Protocol Evaluation in conjunction with the Veterinary Service Management of the Hellenic Republic Prefecture of Attika according to all current European and national legislation and were performed in accordance with relevant guidelines and regulations.
Animals were treated in a therapeutic regimen from week 6 of age or in a prophylactic regimen from week 3 of age. Groups of mice of the same age (gender balanced) received either saline, infliximab (Remicade, Janssen Biotech), certolizumab pegol (Cimzia, UCB) or adalimumab (Humira, Abbvie) administered at 10 mg/kg intraperitoneally twice weekly or etanercept (Enbrel, Pfizer) administered subcutaneously at 30 mg/kg thrice weekly. At the end of the treatment period all mice were sacrificed and hind limbs were flash frozen.
Total RNA was isolated using Trizol reagent from frozen tissues of wild-type (healthy), huTNFTg (diseased) and huTNFTg mice that were treated with one of the 4 different agents in therapeutic or prophylactic regimen as shown in S1 Table. All therapeutic interventions were carried out in 10 biological replicates, while infliximab prophylactic was carried out in 3. Wild-type and transgenic mice were analyzed in replicates of 10 and 13 respectively for a total sum of 66 profiles.
All samples were hybridized on the Affymetrix GeneChip Mouse Gene 2.0 ST array. Data analysis was performed using Transcriptome Analysis Console (TAC 4.0) Software, Applied Biosystems. CEL files were quantile-normalized with RMA. Log-transformed expression measurements were then converted to gene space by calculating mean probeset values referring to the same genes. This was done in order to minimize the complexity of alternative transcript abundance, which we considered, at this level, to be minimal. Only values from genes that were measured in all samples were finally included in the dataset, which consisted of 18704 common measured genes in all 66 samples (S1 Data).
Differential expression analysis is usually calculated against a single background condition that corresponds to a baseline control. We took advantage of our experimental setup and calculated differential expression in the treated samples against both wild-type samples, to quantify gene expression changes versus the native, healthy state, and untreated huTNFTg transgenic samples, to assess changes relative the diseased state. This is a considerable advantage of our approach compared to human studies where the lack of healthy controls often undermines a series of comparisons. Differential expression was calculated as log2Fold-Change (log2FC) values from a one-way ANOVA followed by Dunnett’s test for multiple comparisons using either Wild-type (WT) or Transgenic (TG) samples as control condition. The differential expression analysis was implemented in R and is included in S1 Code. Standard thresholds of |log2FC|≥1 and an adjusted p-value≤0.05 were applied for the definition of differentially expressed genes.
A combination of clustering methods was employed in the clustering of genes and functional enrichments, as well as in the clustering of treatment efficiency scores (see below). Genes and conditions were clustered with agglomerative hierarchical clustering using Ward’s minimum variance criterion [18]. The optimal number of clusters was defined in all cases, on the basis of a Silhouette consistency analysis [19].
Functional analysis was performed with the use of gProfileR [20], through its R package implementation and separately at the levels of GO terms (BP: Biological Process, MF: Molecular Function, CC: Cellular Component), KEGG pathways and Transcription Factor targets. Transcriptional regulator targets were incorporated through the repositories of RNEA [21].
Profile similarity/distance was assessed in the form of the mean absolute difference of gene expression as log2FC. Thus, for a given number of N genes the distance between two profiles (P1, P2) is defined as:
d(P1,P2)=∑i=1N|log2FC(P1)gi−log2FC(P2)gi|N
(1)
Profile distance from wild-type samples was used as a proxy for treatment efficiency, calculated for the entire set of differentially expressed genes, as well as for various subsets defined through the clustering approaches described above.
Comparison of multiple gene lists leads to complex intersection patterns. We used UpSetR [22], an R implementation of the UpSet technique [23] to analyze the intersection between various gene lists. UpSetR produces visualizations of complex set intersections in a matrix-based layout, while also providing information on the original set sizes.
We employed a Random Forest (RF) [24] classification strategy to define subsets of genes and functional categories that best discriminate healthy from diseased profiles as well between various treatments. Random Forests were implemented through the randomForest R package. Starting from the complete dataset we used a 70/30% split for training and test sets respectively and built 1000 RF models with 500 trees each, using 10 variables at each split. From the 1000 RF models we chose the one with the lowest out-of-bag error rate and obtained the variables with the greatest importance on the basis of the higher Mean Gini Decrease (MGD). Arbitrary thresholds for the top most important predictors were applied depending on the downstream analyses (e.g. 30 or 50 for representation reasons, 100 for Principal Component Analysis).
Random Forest classification was implemented in two different datasets. One with all samples, that included wild-type and untreated transgenic samples and one that excluded them, focusing only on the treatment profiles. We used the first to observe broader differences between healthy and diseased states and the second to provide us with a more detailed view of the treatment-specific characteristics.
Two-dimensional analysis of gene expression was calculated as a combination of differential expression values against wild-type (healthy) and transgenic (diseased) mice profiles. The two-dimensional approach was aggregated at functional level. For each functional category we obtained the DE genes that belonged to that category and then, for each treatment, we calculated the mean log2FC value for this subset of genes versus both wild-type and transgenic samples. This pair of values was then used in two-dimensional representations of treatment profiles and in the calculation of treatment efficiency scores. The two-dimensional density plots represent the landscape of response of a given treatment for a certain subset of genes or functions. They are formed through the aggregation of pairs of distances (from wild-type and transgenic) for functional categories specified by the experimenter or derived from previous analyses (such as the classification schemes described above). Each treatment landscape is thus visualized as a contour cloud, the shape and size of which is representative of its efficiency. Displacement along the transgenic (vertical) axis corresponds to desirable distances from the diseased state, while displacement from the wild-type (horizontal) axis is typical of undesirable effects that place the treatment at a distance from the healthy condition.
The relative amplitudes of this data cloud may be quantified in the form of efficiency scores. These are calculated as the log10-based values of the ratios of mean absolute gene expression values of DE against transgenic (TG) over wild-type (WT) profiles, according to the following formula:
E(P,f)=log10|log2FC(Tg)gi¯||log2FC(Wt)gi¯|
(2)
where E(P,f) is the efficiency score of profile (treatment) P for the functional category f, which contains N differentially expressed genes. In (2), gi corresponds to a set of N genes belonging to a given category, over which the mean absolute log-fold-change is calculated. High efficiency scores are thus obtained for functional groups with genes having large absolute changes when compared against transgenic, and low when compared to wild-type profiles.
All analyses were performed in the R environment with the combination of custom scripts and available libraries. Annotated code is provided in a R Mardown file as S1 Code and the processed data files, required for the full replication of our analysis are provided in one compressed folder as S1 Data.
Animal experiments were approved by the Veterinary Service Management of the Hellenic Republic Prefecture of eastern Attika (Approval license Protocol No. 2478, 17/01/2011).
All processed data generated or analysed during this study are included in this published article. Processed files are provided in a single zipped folder (S1 Data). The code for the analysis is also provided as a R Markdown file (S1 Code).
|
10.1371/journal.ppat.1006972 | A fine-tuned vector-parasite dialogue in tsetse's cardia determines peritrophic matrix integrity and trypanosome transmission success | Arthropod vectors have multiple physical and immunological barriers that impede the development and transmission of parasites to new vertebrate hosts. These barriers include the peritrophic matrix (PM), a chitinous barrier that separates the blood bolus from the midgut epithelia and modulates vector-pathogens interactions. In tsetse flies, a sleeve-like PM is continuously produced by the cardia organ located at the fore- and midgut junction. African trypanosomes, Trypanosoma brucei, must bypass the PM twice; first to colonize the midgut and secondly to reach the salivary glands (SG), to complete their transmission cycle in tsetse. However, not all flies with midgut infections develop mammalian transmissible SG infections—the reasons for which are unclear. Here, we used transcriptomics, microscopy and functional genomics analyses to understand the factors that regulate parasite migration from midgut to SG. In flies with midgut infections only, parasites fail to cross the PM as they are eliminated from the cardia by reactive oxygen intermediates (ROIs)—albeit at the expense of collateral cytotoxic damage to the cardia. In flies with midgut and SG infections, expression of genes encoding components of the PM is reduced in the cardia, and structural integrity of the PM barrier is compromised. Under these circumstances trypanosomes traverse through the newly secreted and compromised PM. The process of PM attrition that enables the parasites to re-enter into the midgut lumen is apparently mediated by components of the parasites residing in the cardia. Thus, a fine-tuned dialogue between tsetse and trypanosomes at the cardia determines the outcome of PM integrity and trypanosome transmission success.
| Insects are responsible for transmission of parasites that cause deadly diseases in humans and animals. Understanding the key factors that enhance or interfere with parasite transmission processes can result in new control strategies. Here, we report that a proportion of tsetse flies with African trypanosome infections in their midgut can prevent parasites from migrating to the salivary glands, albeit at the expense of collateral damage. In a subset of flies with gut infections, the parasites manipulate the integrity of a midgut barrier, called the peritrophic matrix, and reach the salivary glands for transmission to the next mammal. Either targeting parasite manipulative processes or enhancing peritrophic matrix integrity could reduce parasite transmission.
| Insects are essential vectors for the transmission of microbes that cause devastating diseases in humans and livestock. Many of these diseases lack effective vaccines and drugs for control in mammalian hosts. Hence, reduction of insect populations, as well as approaches that reduce the transmission efficiency of pathogens by insect vectors, are explored for disease control. Tsetse flies transmit African trypanosomes, which are the causative agents of human and animal African trypanosomiases. These diseases can be fatal if left untreated and inflict significant socio-economic hardship across a wide swath of sub-Saharan Africa [1, 2]. The phenomenon of antigenic variation the parasite displays in its mammalian host has prevented the development of vaccines, and easily administered and affordable drugs are unavailable. However, tsetse population reduction can significantly curb disease, especially during times of endemicity [3, 4]. In addition, strategies that reduce parasite transmission efficiency by the tsetse vector can prevent disease emergence. A more complete understanding of parasite-vector dynamics is essential for the development of such control methods.
For transmission to new vertebrate hosts, vector-borne parasites have to first successfully colonize their respective vectors. This requires that parasites circumvent several physical and immune barriers as they progress through their development in the vector. One prominent barrier they face in the midgut is the peritrophic matrix (PM), which is a chitinous, proteinaceous structure that separates the epithelia from the blood meal [5–7]. In Anopheline mosquitoes the presence of the PM benefits the vector by regulating the commensal gut microbiota and preventing pathogens from invading the hemocoel [8]. In tsetse and sand flies, the PM plays a crucial role as an infection barrier by blocking parasite development and colonization [9, 10]. The presence of the PM can also be exploited by microbes to promote their survival in the gut lumen. The agent of Lyme disease, Borrelia burgdorferi, binds to the tick vector’s gut and exploits the PM for protection from the harmful effects of the blood-filled gut lumen [11]. Unlike vectors that produce a type I PM in response to blood feeding, tsetse’s sleeve-like type II PM is constitutively produced by the cardia organ, which is located at the junction of the fore- and midgut. Upon entering the gut lumen, long-slender bloodstream form (BSF) trypanosomes (Trypanosoma brucei) are lysed while short-slender BSFs differentiate to midgut-adapted procyclic forms (PCF) [12]. During these lysis and differentiation processes, BSF parasites shed their dense coat composed of Variant Surface Glycoproteins (VSGs) into the midgut environment [12]. These molecules are then internalized by cells in the cardia, where they transiently inhibit the production of a structurally robust PM. This process promotes infection establishment by enabling trypanosomes to traverse the PM barrier and invade the midgut ectoperitrophic space (ES) [9]. After entering the ES, trypanosomes face strong epithelial immune responses, which hinder parasite gut colonization success. Detection of PCF parasites in the ES induces the production of trypanocidal antimicrobial peptides [13, 14], reactive oxygen intermediates (ROIs) [15], PGRP-LB [16] and tsetse-EP protein [17]. A combination of these immune effectors eliminates trypanosomes from the majority of flies, leaving only a small percentage of flies infected with PCF parasites in their midgut. The PCF parasites move to tsetse’s cardia where they differentiate into long and short epimastigote forms. These cells then cross the PM for the second time to enter back into the fly’s gut lumen and migrate through the foregut into the salivary glands (SG) for further differentiation into mammalian infective metacyclic forms [18, 19]. Interestingly, the SG infection process, which is necessary for disease transmission, succeeds in only a subset of flies with midgut infections [20]. Even though midgut trypanosomes fail to colonize tsetse’s SG in a subset of flies, parasites persist in the midgut for the remainder of the fly’s adult life. The physiological barriers that prevent SG colonization in the subset of midgut-only infected flies remain unknown.
In this study, we investigated the molecular and cellular mechanisms that prevent parasites from colonizing the SG in a subset of flies with successful midgut infections. Our results show a robust host oxidative stress response reduces parasite survival in the cardia. While preventing parasites from further development, this immune response is costly for tsetse’s cellular integrity and results in extensive damage to cardia tissues. In contrast, less cellular damage is observed in the cardia of flies with midgut parasites that give rise to SG infections. Our results indicate that the ability of the parasites to successfully bypass the PM barrier in the cardia is essential for the establishment of SG infections. We discuss the molecular interactions that regulate this complex and dynamic vector-parasite relationship in the cardia organ, an essential regulator of disease transmission.
Tsetse display strong resistance to infection with trypanosomes. By 3–6 days post acquisition (dpa), parasites that have entered into the ES of the midgut are eliminated by induced vector immune responses from the majority of flies. When newly emerged Glossina morsitans adults (termed teneral) are provided with an infectious bloodmeal in their first feeding, midgut infection success is typically around 30–40% [21, 22]. However, in mature adults that have had at least one prior normal bloodmeal, the infection rate is lower, with only 1–5% of flies housing midgut infections [23, 24]. PCF parasites in susceptible flies replicate in the ES and move forward to the cardia where they differentiate into long and short epimastigote forms. About 6–10 dpa, parasites in the cardia re-enter the gut lumen and migrate through the foregut into the SG, where they differentiate to mammalian infective metacyclic forms within 20–30 days [18, 19].
To generate a suitable sample size of infected flies for down-stream experiments, we provided three groups of teneral G. m. morsitans adult females independent blood meals containing BSF trypanosomes (Trypanosoma brucei brucei RUMP 503) and obtained midgut infection rates of about 30% when microscopically analyzed 40 dpa (Fig 1A). When we analyzed the SG infection status of these gut infected flies, we detected SG infections in about 65% of individuals (Fig 1B). We chose the 40-day time point to accurately score midgut and SG infections, as in our experimental system and insectary environment SG infection status becomes accurately verifiable by microscopy at 30 dpa at the earliest. Hence, two forms of fly infections exist: non-permissive infections where parasites are restricted exclusively to the gut (hereafter designated ‘inf+/-’), and permissive infections where parasites are present in the gut and SGs (hereafter designated ‘inf+/+’) (Fig 1C).
We next investigated whether parasites residing in the non-permissive (inf+/-) gut infections suffer a developmental bottleneck that result in the selection of trypanosomes that are incapable of progressing towards metacyclic infections in the SG. We challenged two groups of teneral adults per os with Trypanosoma brucei brucei obtained from midguts of either inf+/+ or inf+/- flies. We observed a similar proportion of inf+/+ and inf+/- phenotypes regardless of the parasite population (inf+/+ or inf+/-) provided for the initial infection (Fig 1D). This indicates that trypanosomes in the inf+/- gut population are still developmentally competent, and can complete their cyclic development to SG metacyclics. Thus, we hypothesized that the cardia physiological environment may determine the developmental course of trypanosome infection dynamics.
Our prior studies on the role of the PM during the initial parasite colonization event showed that release of VSG from the ingested BSF parasites as they differentiate to PCF cells interferes with gene expression in the cardia resulting in loss of PM integrity early in the infection process. We thus checked to ensure that parasite re-entry into the gut lumen 6–10 dpa was not due to the residual effect of BSF-shed VSG on PM integrity. To do so we analyzed the expression of PM-associated genes (pro1, pro2 and pro3) in cardia three and six days after supplementing flies with a bloodmeal containing purified VSG. Our results show that expression of the PM associated genes is significantly reduced at the day-three time point, but that their expression fully recovers by the day-six time point. These findings indicate that parasite re-entry into the gut lumen in the cardia is unlikely affected by loss of PM integrity that results from the initial VSG effects (Fig 1E).
To investigate the molecular aspects of the infection barriers preventing SG colonization that subsequently limit parasite transmission in the inf+/- group, we used the infection scheme described above and pooled infected cardia into inf+/- and inf+/+ groups (n = 3 independent biological replicates per group, with ten cardia per replicate). For comparison, we similarly obtained dissected cardia from age-matched normal controls (called non-inf; n = 3 independent biological replicates per group, with ten cardia per replicate). We next applied high-throughput RNA-sequencing (RNA-seq) to profile gene expression in the three groups of cardia. We obtained on average > 23M reads for each of the nine libraries, with 77.8% (non-inf), 75.4% (inf+/-) and 64.5% (inf+/+) of the total reads mapping to Glossina morsitans morsitans transcriptome (S1 Fig). The trypanosome reads corresponded to about 4.3% in non-permissive (inf+/-) dataset and about 15% in the permissive (inf+/+) dataset (Fig 2A). To estimate relative parasite densities in the two cardia infection states, we measured the expression of the trypanosome housekeeping gene gapdh in inf+/- and inf+/+ cardia by quantitative real time PCR (qRT-PCR) and normalized the values using tsetse gapdh. We noted significantly higher parasite gene expression values in the inf+/+ cardia compared to the inf+/- cardia samples (Student t-test, p = 0.0028; Fig 2B). We also confirmed that inf+/+ cardia had higher parasite density by microscopically counting trypanosome numbers in the dissected cardia organs using a hemocytometer (S2 Fig). Thus, the difference in the representative parasite transcriptome reads in the two infected groups of cardia is due to an increase in the number of trypanosomes residing in the inf+/+ cardia rather than an increase in parasite transcriptional activity. Interestingly, we noted no difference in the number of trypanosomes present in inf+/- and inf+/+ midguts (S2 Fig). Hence, it appears that parasite density either decreases in the cardia, or fewer parasites colonize the organ despite the fact that inf+/- and inf+/+ flies maintain similar parasite densities in their midguts.
To detect the presence of parasites in the foregut, and to understand the different parasite developmental stages that could be present in the inf+/- and inf+/+ cardia, we investigated parasites by microscopy from these tissues at 40 dpa. We did not observe parasites in the foregut of inf+/- infections, confirming that they are restricted from further development in the cardia in this group of flies. In the inf+/+ state however, we noted presence of many parasites in the foregut in all examined flies. We also looked at the relative presence of the different parasite developmental forms (short and long-epimastigote and trypomastigotes) populating the two different cardia phenotypes by examining parasite morphology and the localization of the nucleus and kinetoplast, as previously described [18, 19]. We observed that the majority of the parasites present in both cardia infection states on day 40 were trypomastigotes with fewer epimastigotes, and that no significant differences between the two cardia infection states was noted (S1 Dataset).
Tsetse’s cardia is composed of several different cell-types with potentially varying functions (schematically shown in Figs 2C and S3) [25–29]. These include an invagination of cells originating from the foregut, which are enclosed within an annular pad of columnar epithelial cells originating from the midgut. The cells occupying this pad secrete vacuoles that deliver components of the PM [27, 29]. The cardia organ is surrounded by muscles that form a sphincter around the foregut, which likely regulates blood flow during the feeding process. Additionally, large lipid-containing cells are localized under a layer of muscle below the sphincter. The function of these cells remains unclear. Microscopy analysis of infected cardia supported our previous molecular findings, as we observed fewer parasites in the cardia of inf+/- (Fig 2D and 2E) when compared to inf+/+ flies (Fig 2F and 2G). Parasites from the inf+/- cardia were restricted to the ES, whereas parasites were observed in both the ES and the lumen of inf+/+ cardia. Hence, the parasite populations resident in inf+/+ cardia had translocated from ES to the lumen, while parasites in inf+/- cardia failed to bypass the PM barrier. These data suggest that the cardia physiological environment may influence the parasite infection phenotype and transmission potential.
For succesful transmission to the next mammalian host, trypanosomes that reside in the ES of the midgut must traverse the PM barrier a second time to re-enter into the gut lumen, move forward through the foregut and mouthparts and colonize the SGs. Traversing the PM a second time is thought to occur near the cardia region [25, 29–31] due to newly synthesized PM likely providing a less robust barrier than in the midgut region. We investigated whether the functional integrity of the PM in the two different infection states varied in the cardia organ. We mined the non-inf cardia transcriptome dataset (S2 Dataset) and identified 14 transcripts associated with PM structure and function [6, 9, 32], which cumulatively accounted for 35.7% of the total number of reads based on CPM value (Fig 3A). The same set of genes represented 26.5% and 34.5% of the inf+/+ and inf+/- transcriptome data sets, respectively (Fig 3A). Thus, PM-associated transcripts are less abundant in the inf+/+ cardia relative to inf+/- and control cardia. We next evaluated the expression profile of PM-associated transcripts and identified those that are differentially expressed (DE) with a fold-change of ≥1.5 in at least one infection state compared to the control (non-inf) (Fig 3B). We observed a significant reduction in cardia transcripts encoding the major PM-associated proventriculin genes (pro2, pro3) in the cardia inf+/+, but not the cardia inf+/- dataset. Both Pro2 and Pro3 are proteinaceous components of the PM [6]. Interestingly, the expression of chitinase was induced in both inf+/- and inf+/+ datasets. Because Chitinase activity can degrade the chitin backbone of the PM, increased levels of its expression would enhance the ability of the parasites to bypass this barrier. Overall, the inf+/+ cardia expression profile we observed here is similar to the profile noted in the cardia 72 hours post BSF parasite acquisition early in the infection process [9]. Results from that study demonstrated that reduced expression of genes that encode prominent PM associated proteins compromised PM integrity, thus increasing the midgut parasite infection prevalence [9]. Loss of PM integrity in the inf+/+ state could similarly enhance the ability of parasites to traverse the PM to re-enter the gut lumen and invade the SGs.
We hypothesized that PM integrity is a prominent factor in the ability of trypanosomes to traverse the barrier in the cardia and continue their migration to the SGs. We addressed this hypothesis by experimentally compromising the structural integrity of the PM in flies that harbored established gut parasite infections. We modified a dsRNA feeding procedure that targets tsetse chitin synthase (dsRNA-cs), which effectively inhibits the production of a structurally robust PM [7]. We challenged flies with BSF trypanosomes as teneral adults and then administered blood meals containing dsRNA-cs on day 6, 8 and 10 post parasite acquisition. This is the time interval when we expect the parasites colonizing the ES of the midgut to bypass the PM barrier in the cardia to re-enter into the lumen [19, 33, 34]. Control groups similarly received dsRNA targeting green fluorescent protein (dsRNA-gfp). Decreased expression of chitin synthase in the experimental dsRNA-cs group relative to the control dsRNA-gfp group was confirmed by qPCR analysis (S4 Fig). Twenty days post dsRNA treatment, midguts and SGs were microscopically dissected and the SG infection status scored. We detected SG infections in 68% of dsRNA-cs treatment group compared to 47% in dsRNA-gfp control group (Fig 3C). Thus, the PM compromised group of flies showed a significant increase in inf+/+ phenotype relative to the control group (GLM, Wald-test, p = 0.0154). These findings suggest that compromising the PM structure later in the infection process increases the proportion of gut infected flies that give rise to mature SG infections (inf+/+). Thus, tsetse’s PM acts as a barrier for parasite translocation from the ES to the lumen of the midgut, an essential step for successful SG colonization.
We sought to determine if components of inf+/+ parasites infecting the cardia could manipulate cardia physiology to bypass the PM. For this, we used a modified version of a host microbial survival assay that was successfully used to evaluate PM structural integrity [7, 9, 35]. In this assay, tsetse with an intact PM fail to immunologically detect the presence of the entomopathogenic Serratia marcescens in the gut lumen. The bacteria thus proliferate uncontrolled in this environment, translocate into the hemocoel and kill the fly [7]. Conversely, when PM structure is compromised, the fly’s immune system can detect the presence of Serratia early during the infection process and express robust antimicrobial immunity that limits pathogen replication and increases host survival [7]. We provided mature adults blood meals supplemented with both entomopathogenic Serratia and heat-treated inf+/+ cardia extracts, while two age-matched control groups received either both Serratia and cardia extracts prepared from flies that had cleared their midgut infections (designated rec-/- for "recovered") or only Serratia (control). We found that survival of flies that received inf+/+ extracts was significantly higher than either of the two control groups (Fig 3D). These findings suggest that cardia inf+/+ extracts contain molecule(s) that negatively influence tsetse‘s PM integrity, thereby enabling these flies to more rapidly detect Serratia and express heightened immune responses to overcome this pathogen. Our transcriptional investigation indicated that PM associated gene expression decreased in the inf+/+ state but not in the inf+/- state (Fig 3B). In the survival assay we described above, we fed flies with cardia extracts from inf+/- containing heat-killed parasites at an equivalent quantity as the one used in the inf+/+ state. When supplemented with the cardia+/- extracts, the survival of flies was decreased to the same level as the two controls, suggesting that extract from cardia inf+/- did not compromise PM integrity (Fig 3D). Collectively, these findings confirm that the parasites in cardia inf+/- differ in their ability to interfere with PM integrity when compared with those in the cardia inf+/+ state. This suggests that parasites in inf+/+ cardia display a different molecular dialogue with tsetse vector tissues.
To understand the cardia-trypanosome interactions, we investigated the parasite populations in the inf+/+ state by transmission electron microscopy (TEM) analysis. We observed that trypanosomes aggregate in the annular cleft formed between the foregut and the midgut parts of the cardia where PM components are synthesized (Figs 4A and S5). Tsetse's PM is composed of three layers; a thin layer that is electron-dense when observed with TEM, a thick layer that is electron-lucent when observed with TEM and a third layer that is not distinguishable when observed with TEM [36]. The newly synthesized PM in the annular cleft is formed by secretions from the annular pad of epithelial cells [27, 29], hence lacking the typical electron-dense and electron-lucent layers observed in the fully formed PM in the midgut (Fig 4B and 4C). From the six inf+/+ cardia analyzed by TEM, we observed trypanosomes embedded in the newly secreted PM as well as present in the lumen. In fact, we had shown above that the expression of putative PM-components decreased in cardia inf+/+ (Fig 3B), and the structural integrity of the PM is compromised based on the Serratia detection assay (Fig 3D). Thus, the structurally weakened PM could enable the trypanosomes to bypass this barrier in inf+/+ flies. Our EM observations (all six inf+/+ cardia) also showed that parasites assemble into compact masses (similar to the previously reported "cyst-like" bodies [33]) in between the layers of the PM (Figs 4C and S6). In three of the six infected inf+/+ cardia analyzed, we noted that the electron-dense layer of the PM restricting a cyst-like body appeared disrupted, which could enable the entrapped parasites to escape the barrier (S7 Fig).
The parasite aggregates we observed in the cardia near the site of PM secretion could represent a social behavior that influences cardia-trypanosome interactions and ultimately parasite transmission success. In vitro, trypanosomes are capable of displaying a similar social behavior termed ‘social motility’ (SoMo) [37]. In this situation early-stage PCF parasites (similar to the forms that colonize the fly midgut) cluster and migrate together on semi-agar plates [38]. In the tsetse vector, phases where trypanosomes group in clusters and move in synchrony have been observed during the infection process independent of the developmental stage of the parasite [34]. Furthermore, parasites co-localize in the cardia near the cells that produce the PM [34], similar to our EM observations. By forming aggregates, trypanosomes could enhance their ability to resist adverse host immune responses and/or escape the ES by crossing through the newly secreted layers of the PM. In addition, the parasites can also actively compromise the PM integrity at this site, as suggested by the PM integrity assay (Fig 3D), but the parasite components that interfere with host functions as such remain to be determined. We also observed extracellular vesicles associated with trypanosomes in TEM images, which could potentially carry molecules that interact with host cells or PM structure (Fig 4).
To understand the parasite-PM interactions in the cardia inf+/- state, we similarly investigated the parasite populations residing in the cardia inf+/- samples by TEM analysis. We observed that parasites in cardia inf+/- are not present in the lumen and are thus unable to escape the ES where the newly synthesized PM is secreted (Fig 5). We noted that high densities of parasites are either lining along the PM secreting cells (Fig 5A and 5B) or are embedded in the PM secretions (Fig 5C and 5D). Trypanosomes observed in this region also presented multiple vacuolation and nuclear condensation, which are indicative of cell death processes in these parasites. Contrary to the cardia inf+/+ transcriptome data, the expression of the majority of PM-associated genes in cardia inf+/- are not significantly decreased (Fig 3B). Moreover, the Serratia assay we applied by co-feeding flies cardia inf+/- extracts indicated no compromise of PM integrity as this group of flies did not survive the bacterial infection (Fig 3D). Thus, it appears that parasites in the cardia inf+/- are restricted by the PM to remain in the ES even at its point of secretion. Also, while cyst-like bodies were frequent in the cardia inf+/+, only a few cyst-like bodies could be observed in cardia inf+/-. Finally, the presence of many physiologically unfit trypanosomes indicates that the inf+/- state represents a hostile environment for the parasite, restricting its survival and transmission (Fig 5C and 5D).
To understand the factors that can successfully inhibit parasite survival in the cardia inf +/-, we examined the inf+/- and inf+/+ cardia datasets relative to the control non-inf state for differential vector responses. We found that 25% (2093) of the total transcripts identifed were differentially expressed (DE). Of the DE transcripts, 31% (646) were shared between the two cardia infection phenotypes, while 36% (756) and 33% (691) were unique to the inf +/- and inf+/+ infection phenotype, respectively (S8 Fig). Of the shared DE transcripts, 89% (576) were similarly regulated between the inf+/+ and inf+/- states while 11% (70) were uniquely regulated in the two infection states. For putative functional significance, we selected transcripts presenting a fold-change of ≥2 between any comparison and a mean CPM value of ≥50 in at least one of the three cardia states. We identified 576 transcripts that were modified in the presence of trypanosomes independent of the cardia infection phenotype, hence representing the core response of the cardia against the parasite infection (S8 Fig). Among these core responses were three antimicrobial peptide (AMP) encoding genes, including two cecropins with fold-changes of >200 (GMOY011562) and >280 (GMOY011563), and attacin D, with a fold change of >27. Production of AMPs by midgut epithelia is among the major trypanocidal responses, and the fact that both cardia inf+/+ and cardia inf+/- expressed these genes at the same level indicates that the ability of inf+/- flies to restrict trypanosomes in the ES is unlikely driven by an AMP-related immune response. We next investigated DE transcripts unique to the two infection phenotypes (S8 Fig). Two putative immunity products, Immune responsive product FB49 and serpin 1, were expressed 223 and 74 times higher, respectively, in inf+/+ compared to inf +/- cardia. Both of these products are induced upon microbial challenge in the tsetse [13, 39]. Additionally, ferritin transcript abundance was >2 times higher in inf+/+ compared to inf +/- cardia. In the subset of transcripts specifically more abundant in the cardia inf+/-, we noted two transcripts encoding proteins involved in the circadian clock, Takeout and Circadian clock-controlled protein, which were 600 and 2 times more abundant relative to cardia non-inf, respectively. Also, transcripts encoding Kazal-type 1 protein, a protease inhibitor, and Lysozyme were more abundant in cardia inf+/-. Given that no single immune-related gene product could explain the cardia inf+/- ability to restrict trypanosomes in the ES, we chose to further evaluate the cardia cellular physiology under the inf+/+ and inf+/- infection states.
To obtain a global snapshot of cardia functions that could influence parasite infection outcomes, the DE cardia transcripts between control (non-inf) and either cardia inf+/- or cardia inf+/+ datasets were subjected to Gene Ontology (GO) analysis (using Blast2GO) (S3 Dataset). We noted 88 GO terms that were significantly down-regulated preferentially in the inf+/- state, while only 15 GO terms were significantly down-regulated in the inf+/+ state. The 88 GO terms detected in the inf+/- dataset included 5, 11 and 67 terms associated with mitochondria, muscles and energy metabolism, respectively.
To understand the physiological implications of the inf+/- infection phenotype in the cardia, we investigated the transcriptional response of the organ as well as the ultrastructural integrity of the mitochondria and muscle tissue. Gene expression patterns indicate that mitochondrial functions are significantly down-regulated in the inf+/- cardia relative to the inf+/+ state (Fig 6A). More specifically, the putative proteins associated with energy metabolism, including the cytochrome c complex, the NADH-ubiquinone oxidoreductase and the ATP-synthase that function at the organelle’s inner membrane, were strongly reduced. Loss of mitochondrial integrity was further demonstrated by microscopic analysis of cardia muscle cells (Figs 6B–6D and S9) and fat-containing cells (Fig 6E–6G). In the cardia inf+/- phenotype, TEM observations showed mitochondrial degradation around myofibrils associated with muscle cells (Fig 6C), while few such patterns were noted in the control cardia (Fig 6B) and cardia inf+/+ (Fig 6D). The mitochondria within the lipid containing cells of both inf+/- and inf+/+ presented a disruption in the organization of their cristae, suggesting a disruption of the inner membrane (Fig 6F and 6G), in support of transcriptomic level findings (Fig 6A).
In addition to putative mitochondrial proteins, we found that the expression of genes encoding structural proteins responsible for muscle contraction, such as myosin and troponin, is also significantly reduced upon infection, particularly in the cardia inf+/- state (Fig 7A). Electron microscopy analysis also revealed a disorganization of the Z band of sarcomeres in muscle tissue surrounding the midgut epithelia in inf+/- cardia, but not in the control and inf+/+ cardia (Fig 7B–7D). Extensive loss of muscle integrity was noted along the midgut epithelia in the inf+/- state. In addition, dilatation of the sarcoplasmic reticulum, muscle mitochondria swelling and vacuolation were observed, suggesting compromised muscle functions associated with this infection phenotype (S9 Fig). The detrimental effects of trypanosome infection on cardia structure and function are more apparent in the inf+/- compared to inf+/+ state, despite the higher number of parasites present during the latter phenotype.
Mitochondria produce reactive oxygen intermediates (ROIs) [40], which in excess can damage the organelle and surrounding cellular structures [41, 42]. The structural damage we observed in mitochondria, muscle tissue and fat cells of inf+/- cardia is symptomatic of oxidative stress [43]. Additionally, our TEM observations demonstrate that parasites in inf+/- cardia exhibit cell-death patterns such as vacuolation and swelling (Fig 8A and 8B), while parasites in inf+/+ cardia appear structurally intact (Figs 2F and 4). Because ROIs modulate trypanosome infection outcomes in tsetse [15, 44], we hypothesized that ROIs may be responsible for controlling trypanosomes in inf+/- cardia and for producing an oxidative environment that concurrently results in tissue damage. We observed a significant increase of peroxide concentrations in both inf+/- (406nM; TukeyHSD posthoc test, p<0.0001) and inf+/+ (167nM; TukeyHSD posthoc test, p = 0.0008) cardia relative to the control cardia (19 nM), with peroxide levels significantly higher in the inf+/- state (TukeyHSD posthoc test, p<0.0001) (Fig 8C). When we experimentally decreased oxidative stress levels in infected flies by supplementing their blood meal with the anti-oxidant cysteine (10μM) (Fig 8D), 85% of midgut infected flies developed SG infections, while only 45% of midgut infected flies had SG infections in the absence of the antioxidant (GLM, Wald-test p<0.001). Our results indicate that the significantly higher levels of ROIs produced in the inf+/- cardia may restrict parasite infections at this crucial junction, while lower levels of ROIs present in the inf+/+ cardia may regulate the parasite density without impeding infection maintenance.
Homeostasis of redox balance is one of the most critical factors affecting host survival during continuous host-microbe interaction in the gastrointestinal tract [45]. In the mosquito Anopheles gambiae, increased mortality is observed when ROIs are produced in response to Plasmodium berghei infections [46]. A similar trade-off expressed in the inf+/- cardia may restrict parasite infections while causing collateral damage to essential physiologies. Conversely, strong anti-parasite responses that compromise essential physiologies are absent in the cardia of the inf+/+ group, thus allowing the parasites to continue their journey to colonize the SG and successfully transmit to a new host. Additionally, flies with SG parasite infections also suffer from longer feeding times due to suppressed anti-coagulation activity in the SG, which may further help with parasite transmission in this group of flies [47].
Trypanosome transmission by tsetse reflects a tug-of-war that begins with parasite colonization of the midgut and ends when parasites are transmitted to the next vertebrate via saliva. Initially during the infection process, BSF trypanosome products manipulate tsetse vector physiology to bypass the gut PM to colonize the midgut ES [9]. Our results show that to successfully colonize the SG, trypanosomes may again manipulate tsetse physiology to escape the midgut ES for access to the foregut, and subsequently to the SG. To re-enter the lumen, it is hypothesized that trypanosomes cross the PM in the cardia where newly synthesized PM is less structurally organized and hence can provide an easy bypass [25, 29–31]. Here, we provide evidence in support of this hypothesis by showing that in flies where trypanosomes successfully colonize the SG, the parasites are accumulating in the region where the PM is newly secreted, and are observed both embedded in the PM secretions and free in the lumen (summarized in Fig 9). To facilitate their passage, components of trypanosomes in the cardia can apparently manipulate PM integrity by influencing the expression of PM-associated genes through molecular interference, the mechanisms of which remain to be studied. Alternatively, trypanosome-produced molecules may directly reduce the integrity of the PM as a barrier.
The presence of trypanosomes in the cardia triggers immune responses which include the production of ROIs. In flies where midgut infections fail to reach the SG (inf+/-), increased levels of peroxide produced in the cardia may restrict parasite survival and prevent them from further development in the fly. Given that the inf+/- phenotype is costly and leads to collateral damage in the cardia tissues of infected flies, it is possible that flies may be able to sustain this phenotype under laboratory conditions where resources are abundant for a minimal effort. It remains to be seen if the inf+/- phenotype could sustain in natural populations in the field. Because in field infection surveillance studies estimating the time of initial parasite infection acquisition is not possible, concluding the cardia infection status in natural populations is difficult. It may however be possible to initiate parasite infection experiments using field-caught teneral flies to partially evaluate the potential colony-bias that could arise under insectary conditions using fly lines that have been kept in captivity for many years.
Trypanosome colonization of tsetse’s SG could represent a trade-off where vector tolerance to parasites leads to minimal self-inflicted collateral damage. Interestingly, different tsetse species may have evolved varying strategies to defend against parasitism. For instance, under similar laboratory conditions and using the same parasite strain for infection, Glossina pallidipes heavily defends against the initial infection, as the occurrence of the inf+/- phenotype in this species is rare despite similar resistance to SG transmission [48]. On the other hand, the closely related species G. morsitans, which we investigate here, has developed a different strategy to combat against parasite transmission [48]. Glossina morsitans presents a less efficient defense against the initial parasite infection in the midgut compared to G. pallidipes, but can similarly control the parasite transmission by restricting SG infections in midgut infected flies. Investigating the causes leading to this drift in strategies could lead to the development of new control strategies based on enhancing the immune defenses of the vector against parasites. Our work highlights the central role tsetse’s PM plays in parasite-vector interactions and infection outcome. This work opens up the possibility for exploiting this matrix as a target for vector control strategies to enhance its barrier function to block parasite transmission.
This work was carried out in strict accordance with the recommendations in the Office of Laboratory Animal Welfare at the National Institutes of Health and the Yale University Institu- tional Animal Care and Use Committee. The experimental protocol was reviewed and approved by the Yale University Institutional Animal Care and Use Committee (Protocol 2014–07266 renewed on May 2017).
Glossina morsitans morsitans were maintained in Yale’s insectary at 24°C with 50–55% relative humidity. All flies received defibrinated bovine blood (Hemostat Laboratories) every 48 hours through an artificial membrane feeding system. Only female flies were used in this study.
Bloodstream form Trypanosoma brucei brucei (RUMP 503) were expanded in rats.
Flies were infected by supplementing the first blood meal of newly eclosed flies (teneral) with 5x106 parasites/ml. Where mentioned, cysteine (10μM) was added to the infective blood meal to increase the infection prevalence [15].
For survival assays, Serratia marcescens strain Db11 was grown overnight in LB medium. Prior to supplementation with Serratia, the blood was inactivated by heat treatment at 56°C for 1 hour as described in [7].
At day 40 post parasite challenge, all flies were dissected 48 hours after their last blood meal, and midgut and salivary glands (SG) were microscopically examined for infection status. Flies were classified as inf+/+ when infection was positive in both the midgut and the SG, as inf+/- when infection was positive in the midgut but negative in the SG. Cardia from inf+/+ and inf+/- flies were dissected and immediately placed in ice-cold TRIzol (Invitrogen). For each infected group, inf+/+ and inf+/-, 10 cardia were pooled into one biological replicate and three independent biological replicates were obtained and stored at -80°C prior to RNA extraction. Similarly, three independent biological replicates containing 10 cardia from age-matched flies that had only received normal blood meals (non-inf) were prepared. Total RNA was extracted from the nine biological replicates using the Direct-zol RNA Minipreps kit (Zymo Research) following the manufacturer instructions, then subjected to DNase treatment using the Ambion TURBO DNA-free kit AM1907 (Thermo Fisher Scientific). RNA quality was analyzed using the Agilent 2100 Bioanalyzer RNA Nano chip. mRNA libraries were prepared using the NEBNext Ultra RNA Library Prep Kit for Illumina (New England BioLabs) following the manufacturer recommendations. The nine libraries were sequenced (single-end) at the Yale Center for Genome Analysis (YCGA) using the HiSeq2500 system (Illumina). Read files have been deposited in the NCBI BioProject database (ID# PRJNA358388).
Using CLC Genomics Workbench 8 (Qiagen), transcriptome reads were first trimmed and filtered to remove ambiguous nucleotides and low-quality sequences. The remaining reads were mapped to Glossina morsitans morsitans reference transcriptome GmorY1.5 (VectorBase.org). Reads aligning uniquely to Glossina transcripts were used to calculate differential gene expression using EdgeR package in R software [49].
Significance was determined using EdgeR exact test for the negative binomial distribution, corrected with a False Discovery Rate (FDR) at P<0.05.
Identified genes were functionally annotated by BlastX, with an E-value cut-off of 1e-10 and bit score of 200, and peptide data available from D. melanogaster database (FlyBase.org). Blast2GO was utilized to identify specific gene ontology (GO) terms that were enriched between treatments based on a Fisher’s Exact Test [50].
Cardia tissues from three non-inf, five inf+/- and six inf+/+ 40 day-old flies were dissected in 4% paraformaldehyde (PFA) and placed in 2.5% gluteraldehyde and 2% PFA in 0.1M sodium cacodylate buffer pH7.4 for 1 hour. Observed infected cardia were obtained from two different groups of flies independently infected with trypanosomes (n1 = 3 and n2 = 2 for inf+/-; n1 = 3 and n2 = 3 for inf+/+). Tissues were processed at the Yale Center for Cellular and Molecular Imaging (CCMI). Tissues were fixed in 1% osmium tetroxide, rinsed in 0.1M sodium cacodylate buffer and blocked and stained in 2% aqueous uranyl acetate for 1 hour. Subsequently, tissues were rinsed and dehydrated in a series in ethanol followed by embedment in resin infiltration Embed 812 (Electron Microscopy Sciences) and then stored overnight at 60°C. Hardened blocks were cut in sections at 60nm thickness using a Leica UltraCut UC7. The resulting sections were collected on formvar/carbon coated grids and contrast-stained in 2% uranyl acetate and lead citrate. Five grids including two sections prepared from each different insects were observed using a FEI Tecnai Biotwin transmission electron microscope at 80Kv. Images were taken using a Morada CCD camera piloted with the iTEM (Olympus) software. Contrasts of the pictures were adjusted using Photoshop CC 2018 (Adobe).
At day 40 post parasite challenge, flies were dissected 72 hours after their last blood meal, and midgut and salivary glands (SG) were microscopically examined for infection status. Cardia were dissected, pooled by 5 in ice-cold TRIzol (Invitrogen) in function of their infection status (inf+/+ or inf+/-), and then flash-frozen in liquid nitrogen. RNA was extracted using the Direct-zol RNA MiniPrep (Zymo Research) following the manufacturer instructions, then subjected to DNase treatment using the Ambion TURBO DNA-free kit AM1907 (Thermo Fisher Scientific). 100ng of RNA was utilized to prepare cDNA using the iScript cDNA synthesis kit (Bio-Rad) following the manufacturer instructions. qPCR analysis was performed using SYBR Green supermix (Bio-Rad) and a Bio-Rad C1000 thermal cycler. Quantitative measurements were performed in duplicate for all samples. We used ATTCACGCTTTGGTTTGACC (forward) and GCATCCGCGTCATTCATAA (reverse) as primers to amplify trypanosome gapdh. We used CTGATTTCGTTGGTGATACT (forward) and CCAAATTCGTTGTCGTACCA (reverse) as primers to amplify tsetse gapdh. Relative density of parasite was inferred by normalizing trypanosome gapdh expression by tsetse gapdh expression. Statistical comparison of relative densities was performed on Prism 7 (GraphPad software) using a Student t-test.
Direct counting of parasites was operated by dissecting the cardia and the whole remaining midgut from flies prepared similarly than above. Individual tissues were homogenized in PSG buffer (8 replicates for each tissue). Homogenate was then fixed in an equal volume of 4% PFA for 30 min. The solution was then centrifuged 15 min at 110g, the supernatant was discarded and the pellets containing the trypanosomes from cardia and midguts were suspended in 100μl and 2,500μl PSG buffer, respectively. Trypanosomes from the total solution were counted using a hemocytometer. Statistical comparison of numbers was performed on Prism 7 (GraphPad software) using a Mann-Whitney rank test.
At day 40 post parasite challenge, flies were dissected 72 hours after their last blood meal, and midgut and salivary glands (SG) were microscopically examined for infection status. Around 40 inf+/+ and inf+/- were independently pooled together, and then roughly homogenized in 500μl of PSG buffer (PBS+2% glucose). Each homogenate was centrifuged 10min at 30g to precipitate midgut debris, and then each supernatant containing parasites was transferred to a new tube to be centrifuged 15min at 110g to precipitate the parasites. Supernatants were then discarded and each pellet containing midgut procyclic trypanosomes either from inf+/+ or inf+/- flies was suspended in 500μl PSG. Parasites were counted using a hemocytometer.
Newly emerged adult females were provided a blood diet including 10μM Cysteine and supplemented with 5×106 of procyclic trypanosomes from either inf+/+ or inf+/- flies prepared as described above. All flies were subsequently maintained on normal blood thereafter every 48 h. Four independent experiments were done for each type of trypanosomes. Midgut and salivary gland infections in each group were scored microscopically two weeks later. Precise sample sizes and count data are indicated in S1 Dataset.
Statistical analysis was carried out using the R software for macOS (version 3.3.2). A generalized linear model (GLM) was generated using binomial distribution with a logit transformation of the data. The binary infection status (inf+/+ or inf+/-) was analyzed as a function of the origin of the procyclic trypanosomes (inf+/+ or inf+/-) and the experiment it belongs to. The best statistical model was searched using a backward stepwise procedure from full additive model (i.e. parasite origin+experiment#) testing the main effect of each categorical explanatory factor. Using the retained model, we performed a Wald test on the individual regression parameters to test their statistical difference. Precise statistical results are indicated in S1 Dataset.
Cardia from inf+/+ and inf+/- flies were dissected 40 dpa. Ten organs were pooled and gently homogenized in 100μL PBS and parasite numbers were evaluated using a hemocytometer. As cardia inf+/- contain less trypanosomes than cardia inf+/+, homogenates from cardia inf+/+ were diluted to the density of parasites present in cardia inf+/-. Equal numbers of parasites were then fixed in 2% Paraformaldehyde (PFA) PBS by adding an equal volume of 4% PFA PBS to the cardia inf+/+ and inf+/- homogenates. Parasites were then centrifuged for 10min at 500g and the resulting pellet was resuspended and washed in PBS. Samples were then centrifuged for 10min at 500g and the resulting pellet was resuspended in 200μL distilled water. 50μL of parasite-containing solution was deposited on poly-lysine coated slides and air dried. Slides were permeabilized for 10min in 0.1% Triton X-100 PBS, and then washed in PBS 5min and in distilled water 5min. Fluorescent DNA staining was then applied by covering the slides with a solution of DAPI in distilled water (1μg/mL) for 20 min in the dark. Slides were subsequently washed in distilled water two times for 5 min before being air dried in the dark. Microscopic observations were realized using a Zeiss AxioVision microscope (Zeiss). Detailed counts are indicated in S1 Dataset.
Soluble VSG (sVSG) was prepared as described in [9]. Eight-day old adult flies received a blood meal containing purified sVSG (1μg/ml), or bovine serum albumin (BSA) (1μg/ml) as a control. To assess the effect of sVSG on gene expression at three days, cardia organs were microscopically dissected at 72h post treatment. To assess the effect of sVSG on gene expression at six days, remaining flies that were not dissected at three days were given a second normal blood meal, and the cardia organs were microscopically dissected at 72h post second feeding. Five biological replicates for each treatment and each time point were generated. Five dissected cardia were pooled for each replicate and their RNA was extracted. 100ng RNA was used to generate cDNA. Quantitative real-time PCR (qRT-PCR) was used to evaluate the expression of the PM-associated genes proventriculin-1, -2 and -3 as described in [9]. Normalization was performed to the internal control of gadph mRNA for each sample. Pairwise comparisons for each time point of the genes relative expression between sVSG and BSA treated flies was carried out with the Prism 7 software (GraphPad software) using a Student t-test. Precise statistical results are indicated in S1 Dataset.
Green fluorescent protein (gfp) and chitin synthase (cs) gene specific dsRNAs were prepared as described in [7]. Newly emerged adult females were provided with a trypanosome supplemented blood diet that also included 10μM Cysteine. All flies were subsequently maintained on normal blood thereafter every 48 h. After 6 days (at the 3rd blood meal), flies were divided into two treatment groups: first group received dsRNA-cs and the second group control dsRNA-gfp. The dsRNAs were administered to each group in 3 consecutive blood meals containing 3mg dsRNA/20μl blood (the approximate volume a tsetse fly imbibes each time it feeds). Four independent experiments using the same pool of dsRNA were generated for each treatment. Midgut and salivary gland infections in each group were scored microscopically three weeks later. Precise sample sizes and count data are indicated in S1 Dataset. Statistical analysis on the infection outcomes following the antioxidant feeding was carried out using the R software for macOS (version 3.3.2). A generalized linear model (GLM) was generated using binomial distribution with a logit transformation of the data. The binary infection status (inf+/+ or inf+/-) was analyzed as a function of the dsRNA treatment (dsRNA-gfp or dsRNA-cs) and the experiment it belongs to. The best statistical model was searched using a backward stepwise procedure from full additive model (i.e. dsRNA treatment+experiment#) testing the main effect of each categorical explanatory factors. Using the retained model, we performed a Wald test on the individual regression parameters to test their statistical difference. Precise statistical results are indicated in S1 Dataset.
Quantitative real-time PCR (qRT-PCR) was used to validate the effectiveness of our RNAi procedure as described in [7]. For each treatment of each experiment, we dissected the cardia of five randomly selected flies 72h after their third dsRNA-supplemented blood meal. The five dissected cardia were pooled together and their RNA was extracted. 100ng RNA was used to generate cDNA. RNA extractions from experiment #3 failed, but as the same dsRNA pools were used for all experiments and considering the consistency of the knockdown we observed, we decided to maintain experiment #3 in our counting results.
To assess the PM integrity, we applied a host survival assay following per os treatment of each group with Serratia marcescens as described in [7, 9]. We provided to three groups of 8 day-old flies (in their 4th blood meal) either cardia extracts obtained from challenged flies that cleared the trypanosomes and are subsequently recovered from initial infection (rec-/-), or a cardia extract from inf+/- flies, or a cardia extract from inf+/+ flies. We included a fourth group of 8-day old flies that received an untreated blood meal.
Cardia extract was obtained by dissecting, in PBS, the cardia from 40 days-old infected as described above. Approximately fifty cardia from either rec-/-, inf+/- or inf+/+ flies were pooled together, and then gently homogenized. Parasites were counted from the homogenates of inf+/- and inf+/+ using a hemocytometer. The three cardia homogenates were then heated at 100°C for 10 minutes. inf+/- and inf+/+ extracts were provided to reach a concentration of 5×105 parasites per ml of blood. As inf+/- cardia contain fewer parasites than inf+/+ cardia, the volume of the inf+/+ extract provided was adjusted by dilution in PSG buffer to be equal to inf+/- volume. Rec-/- extract was provided at an equal volume than infected extracts to ensure the presence of a similar quantity of extract molecules coming from the cardia in these groups. 48 hours after the flies received blood meal supplemented with the different extracts, all flies were provided a blood meal supplemented with 1,000 CFU/ml of S. marcescens strain Db11. Thereafter, flies were maintained on normal blood every other day, while their mortality was recorded every day for 30 days. Precise counting data are indicated in S1 Dataset.
Statistical analysis was carried out using the R software for macOS (version 3.3.2). We used an accelerated failure time model (Weibull distribution) where survival was analyzed as a function of the extract received (survreg() function of "survival" package). Pairwise tests were generated using Tukey contrasts on the survival model (glht() function of "multcomp" package). Precise statistical results are indicated in S1 Dataset.
Newly emerged adult females were provided with a trypanosome-supplemented blood diet that also included 10μM Cysteine. All flies were subsequently maintained on normal blood thereafter every 48 h. After 10 days (at the 5th blood meal), flies were divided into two treatment groups: first group received the anti-oxidant Cysteine (10μM) and the second group was fed normally as a control. Cysteine was administered each blood meal until dissection. Four independent experiments were done for each treatment. Midgut and salivary gland infections in each group were scored microscopically three weeks later. Precise sample sizes and count data are indicated in S1 Dataset.
Statistical analysis was carried out using the R software for macOS (version 3.3.2). A generalized linear model (GLM) was generated using binomial distribution with a logit transformation of the data. The binary infection status (inf+/+ or inf+/-) was analyzed as a function of the treatment (control or cysteine) and the experiment it belongs to. The best statistical model was searched using a backward stepwise procedure from full additive model (i.e. antioxidant treatment+experiment#) testing the main effect of each categorical explanatory factors. Using the retained model, we performed a Wald test on the individual regression parameters to test their statistical difference. Precise statistical results are indicated in S1 Dataset.
ROS were quantified using the Amplex Red Hydrogen Peroxide/Peroxidase Assay Kit (ThermoFisher Scientific), following the manufacturer recommendations. 40 days post parasite challenge, flies were dissected 72 hours after their last blood meal, and midgut and salivary glands (SG) were microscopically examined for infection status. For each infection phenotype (i.e. inf+/+ or inf+/-), 3 replicates containing each 10 cardia tissues pooled and homogenized in 80μl of ice-cold Amplex Red Kit 1X Reaction Buffer were generated. Three replicates of age-matched non-infected cardia tissues were conceived in the same manner. 50μl of assay reaction mix was added to 50μl of the supernatant of each samples, and then incubated 60 minutes at RT. Fluorescence units were determined using a BioTek Synergy HT plate reader (530nm excitation; 590nm emission). Peroxide concentrations were determined using the BioTek Gen5 software calculation inferred from a standard curve (precise results are indicated in S1 Dataset). Statistical analysis was performed on Prism 7 (GraphPad software) using a one-way ANOVA where ROS concentration was analyzed as a function of the infection status. Pairwise comparisons were carried out using a TukeyHSD posthoc test.
|
10.1371/journal.ppat.1006079 | Vaccinia Virus Immunomodulator A46: A Lipid and Protein-Binding Scaffold for Sequestering Host TIR-Domain Proteins | Vaccinia virus interferes with early events of the activation pathway of the transcriptional factor NF-kB by binding to numerous host TIR-domain containing adaptor proteins. We have previously determined the X-ray structure of the A46 C-terminal domain; however, the structure and function of the A46 N-terminal domain and its relationship to the C-terminal domain have remained unclear. Here, we biophysically characterize residues 1–83 of the N-terminal domain of A46 and present the X-ray structure at 1.55 Å. Crystallographic phases were obtained by a recently developed ab initio method entitled ARCIMBOLDO_BORGES that employs tertiary structure libraries extracted from the Protein Data Bank; data analysis revealed an all β-sheet structure. This is the first such structure solved by this method which should be applicable to any protein composed entirely of β-sheets. The A46(1–83) structure itself is a β-sandwich containing a co-purified molecule of myristic acid inside a hydrophobic pocket and represents a previously unknown lipid-binding fold. Mass spectrometry analysis confirmed the presence of long-chain fatty acids in both N-terminal and full-length A46; mutation of the hydrophobic pocket reduced the lipid content. Using a combination of high resolution X-ray structures of the N- and C-terminal domains and SAXS analysis of full-length protein A46(1–240), we present here a structural model of A46 in a tetrameric assembly. Integrating affinity measurements and structural data, we propose how A46 simultaneously interferes with several TIR-domain containing proteins to inhibit NF-κB activation and postulate that A46 employs a bipartite binding arrangement to sequester the host immune adaptors TRAM and MyD88.
| Viruses possess mechanisms to interfere with the host immune system to enhance their replication. Vaccinia virus, the viral vaccine used to eradicate smallpox, synthesizes many such proteins. The vaccinia virus protein A46 is one of a series of proteins preventing expression of host proteins that induce an anti-viral state. A46 acts early to inhibit anti-viral state induction by specifically binding to certain host adapter proteins such as MyD88 and TRAM. Here, we extend our knowledge of the A46 structure by determining the structure of the protein's N-terminal domain to be an unusual lipid binding fold. In addition, the full-length A46 molecule has a novel quaternary structure that can both bind proteins and lipids, indicating that A46 uses a variety of interactions to sequester host proteins, thus impairing the activation of the anti-viral state and improving the efficiency of viral replication.
| Viral infection depends not only on the rate and precision of viral reproduction, but also requires a simultaneously efficient inhibition of host immune responses. Viruses have evolved varied strategies to interfere with immune responses of the host, including production of secreted molecules that mimic innate immune receptors, molecules that trap cytokines as well as the shut-off of the cellular transcription and translation machinery [1, 2]. Vaccinia virus (VACV), the virus used to eradicate smallpox, has been extensively studied as a model of virus-host interaction because of its plethora of anti-immune strategies and its large arsenal of immunomodulator tools [3]. Further interest in VACV stems from its role as a vaccine vector against important infectious diseases and its potential role against cancer [4, 5].
Amongst approximately 200 genes in the VACV genome, only half encodes for the viral replication machinery; many of the remaining gene products have roles as extra- and intracellular modulators of the host immunity [6]. The VACV intracellular immunomodulators form a family of Bcl-2-like (B-cell lymphoma 2 like) proteins with low sequence identity but high structural similarity to the eukaryotic Bcl-2 protein family [7]. Eukaryotic Bcl-2 proteins present a diverse group of pro- and anti-apoptotic regulators that share α-helical BH domains [3, 8]. To date, 11 Bcl-2-like proteins encoded by VACV have been identified. Those such as A46, A49, A52, B14, N1, K7 and F1 have an experimentally confirmed Bcl-2 fold [9–16]; others such as C1, C6, C16/B22 and N2 are predicted to have such a fold [10, 17, 18].
NF-κB is a transcriptional factor that responds to the stimulation of Toll-like-receptors (TLRs) and Interleukin-like-receptors (IL-1R) by inducing expression of effector molecules. In the uninfected cell, inactive NF-κB is located in the cytoplasm as a precursor or in a complex with its inhibitor (IκB). Upon stimulation of TLRs by pathogens, a signaling cascade is initiated through the recruitment of adaptor proteins (e.g. MyD88, MAL/TIRAP, TRIF, TRAM) by the cytoplasmic domains of TLRs, consequent stepwise activation of IRAK2-IRAK6-IRAK4 kinases followed by activation of TRAF6 ubiquitin ligase and activation of the IKK (IκB kinase) complex. Finally, the release of the active form of NF-κB results from processing of the precursors or degradation of IκB. Nuclear migration of the free NF-κB permits expression of a range of cytokines allowing the development of both innate and adaptive immune responses [19]. VACV Bcl-2-like immunomodulators disrupt NF-κB activation pathways at different stages by targeting various components [3, 7]. The A46 protein acts close to the plasma membrane by binding numerous TIR-domain containing adaptor proteins such as MyD88, MAL/TIRAP, TRAM and TRIF as well as TLR4 to prohibit further signal propagation [20].
We recently determined the structure of the Bcl-2 domain of A46 comprising residues 87–229 [9]. However, structural information on the N-terminal domain (residues 1–86), its position relative to the Bcl-2-like domain and a plausible function were lacking. Here, we report the crystal structure of the A46 N-terminal domain comprising residues 1 to 76 and demonstrate that this domain binds fatty acids. Further, small-angle X-ray scattering (SAXS) was employed to derive a structural model of full-length of A46(1–240). Using a SAXS-derived model of A46 together with biochemical data, we postulate a mechanism explaining the biological function of this unusual VACV immunomodulator protein.
Members of the VACV Bcl-2-like family whose structure has been determined mainly comprise a single Bcl-2-like domain with an N- or a C-terminal extension (ranging from 5 to 80 amino acids) or both (Fig 1A). At present, structural information is only available for the Bcl-2-like domains and a short unstructured N-terminal region of F1L [21] but not for the rest of extensions. However, the N-terminal extension of A46 spanning residues 1–80 was predicted by PSIPRED (31) to comprise exclusively β-strands. Previous studies using limited proteolysis on the full-length A46 protein confirmed the presence of a structured N-terminal domain in the first 80 residues, suggesting that it would be amenable to crystallography (Fig 1A) [9].
To examine the structure and function of the N-terminal domain of A46, we designed two constructs for expression in E. coli. Both protein expression constructs contained the first methionine of the full-length protein and comprised 73 or 83 A46 residues, as constructs with fewer than 73 residues were either insoluble when His6-tagged or could not be removed from the MBP expression tag. Both variants contained an additional four amino acids (MAQQ, Fig 1B) to improve solubility as observed with full-length A46 [9]. Thus, both fusion proteins had the following structure: His6-TRX-TEVsite-MAQQ-A46(1-73/83). The average yield of both proteins was approximately 2.5 mg highly purified protein per L of bacterial culture. However, as we only obtained diffraction quality crystals with A46(1–83), we performed all subsequent work with this variant (Fig 1C).
We first examined the ability of A46(1–83) to bind the TIR domains of its proposed cellular binding partners such as MyD88 and MAL. Using microscale thermophoresis, we previously demonstrated that the C-terminal domain of A46 binds in the low micromolar range to these TIR domains; the KD values were slightly lower than those observed with the full-length protein (Table 1) [9]. In contrast, the N-terminal A46(1–83) binds to TIR/MyD88 but not to TIR/MAL. The KD value was 8.8 μM, compared to that of 0.52 μM for A46(1–229). We also examined the binding of the TIR/TRAM domain, another proposed A46 in vivo binding partner, to A46 [20, 24]. The interaction of full-length A46 and its C-terminal domain with TIR/TRAM shows KD values of 2.39 μM and 3.62 μM, respectively. However, under the conditions used, A46(1–83) did not bind to TIR/TRAM (Table 1).
Given the binding of A46(1–83) to MyD88, we next examined whether this fragment was sufficient to prevent IL-1β induction of NF-κB-mediated transcription using similar cell-based assays to those described previously [9]. Plasmid amounts were adjusted so that approximately the same amounts of each A46 variant were expressed; the total amount of transfected DNA (500 ng) was kept constant by the addition of empty pCAGGS vector. Unlike full-length A46 and the truncated variant A46(87–229), the N-terminal domain exhibits no appreciable inhibition of the IL-1β driven induction of NF-κB (Fig 2; see figure legend for statistics). Thus, binding of A46 residues 1–83 is insufficient to independently fulfil an immunomodulatory role.
We initiated structural studies of the functional form of the N-terminal domain of A46(1–83) by setting up crystallization trials with commercial screens. Small single crystals of around 20 μm in size were observed after 1 week of incubation at 22°C. They failed, however, to grow larger; nevertheless, several datasets were collected using the beam line for high throughput macromolecular data-collection MASSIF at ESRF, rendering the highest resolution between 1.8 and 2.3 Å. With no known close homologue in the PDB database, we were unable to solve the phase problem by molecular replacement. Thus, we labelled the protein with selenomethionine; diffraction quality crystals grew in the conditions used for the native protein. Data sets for SAD were collected using the MASSIF beamline up to 1.55 Å resolution. However, we were unable to phase the structure using the anomalous signals, most likely because all three methionines in the protein lie in the very N- and C-termini of the A46(1–83) construct and, consequently, are located in flexible regions. Finally, the phases were obtained by ARCIMBOLDO_BORGES [25] crystallographic software. The program exploits tertiary structure libraries extracted from the Protein Data Bank for ab initio phasing. A library of 7650 superimposed polyalanine models, representing 925300 variations on the fold of three stranded antiparallel β-sheets totalling 20 amino acids, was used as fragment hypotheses. This three strands arrangement is most frequently found in β-sheets. Computations were executed on the Gordon supercomputer at the San Diego Supercomputer Center in California. A partial solution was obtained upon location with PHASER [26] of 4 models extracted from the unrelated PDB structures 2QLG, 2GSK, 2EFU, 4DCB as indicated by SHELXE [27] trace correlation coefficients above 40%. The root mean square deviation (rmsd) of the solving library models against the final structure was in the range of 0.35 Å (model from 4DCB) and 0.61 Å (model from 2EFU).
A46(1–83) crystallized with two molecules in the asymmetric unit; electron density for a bound ligand, later identified as myristic acid, was found inside one of the molecules. The two A46 molecules comprise two β-sheets arranged head to head as an extended β-sandwich (Fig 3A, Table 2). A tetramer is formed over a crystallographic twofold axis continuing the β-sandwich with the second dimer rotated approximately 90° relative to the first (S1A and S1B Fig). The PISA server [28] estimates both association interfaces to be present in solution, burying 1276 and 941 Å2.
The A and B independent subunits show marked differences, with a Cα rmsd of 1.3 Å for the 45 common β-strand residues (Fig 3B); the tetramer can be described as an A/B/B/A arrangement. Subunit A has 7 β-strands whereas subunit B presents only 6, lacking the most C-terminal one (Fig 3A). No electron density is seen for either residues 77–83 in subunit A or 67–83 in subunit B, suggesting that these regions may constitute a flexible linker between N- and C-terminal domains in the full-length molecule.
A striking feature of the external β1-β7 face of the A subunit is a partially hydrophobic tunnel, spanning the whole subunit A and reaching into subunit B (Fig 4A). A length of 22 Å, an average radius of 2.5–3 Å and an overall cavity volume of 1150 Å3 (S2 Fig) were calculated with the software MOLE 2.0 [30, 31]. The tunnel is occupied by an extended well-defined electron density, reminiscent of a myristic acid molecule (Fig 4A). The omit electron density map for the ligand is presented in S3 Fig. Mass spectrometry and gas chromatography (GC) analysis of the lipids extracted from purified protein identified the fatty acids C14:0, C16:0 and C16:1 in complex with A46(1–83) (Fig 4B). Repetition of the experiment with a separate A46(1–83) preparation revealed the same three fatty acids but in different ratios, indicating that the relative amounts may be preparation dependent. However, in all preparations so far examined, the C14:0 fatty acid was highly enriched compared to its overall representation in E.coli cells (Fig 4E).
Further, the lipid extraction and identification by mass spectrometry was also done with three independently purified samples of the full-length A46 as well as two purified samples containing the C-terminal domain of A46 (87–229). For the full-length A46, we also identified the three co-purified fatty acids, C14:0, C16:0 and C16:1; in contrast, purified A46(87–229) lacked any complexed lipids (Fig 4C and 4D). Hence, only samples containing the N-terminal domain of A46, either purified independently or as a part of the full-length protein, are capable of binding fatty acids.
Subunit B, being partially penetrated by the fatty acid, cannot therefore lodge a second molecule. The cavity, present in subunit A, is collapsed in subunit B, bringing both β-sheets 3.5Å nearer (Fig 3B). Tyr37 adopts a dual conformation in the two subunits, suggesting a gate-keeper role as it folds back in subunit A to make room for the myristic acid (Fig 3B). The side chain of the preceding His36, pointing to the outside and located in the loop displaying highest differences between both subunits, also has two conformations. A single loop at each side of the sandwich joins both sheets, allowing the displayed flexibility. One loop (β1 to β2) is unchanged; the other (β4 to β5), containing four charged residues DRDK, differs between the subunits, altering its hydrogen bond pattern (Fig 3B). Together with His36, these electrostatic interactions may provide a lever for myristic acid binding. The absence of bound fatty acid as well as the lack of the β7 strand in subunit B results in a quite different interaction interface to that in subunit A, allowing association of two B subunits, with β1 occupying the position vacated by β7, and thus the assembly of the symmetric tetramer (S1 Fig).
We examined the lipid-binding properties of A46(1–83) by structure-based site-directed mutagenesis. We introduced the single mutations F3D, H36L, Y37A, Y37W, I72A into the expression plasmid for A46(1–83) and successfully expressed and purified protein from all variants. Analysis of their lipid content showed that all variants contained C14:0, C16:0 and C16:1 fatty acids. Furthermore, only the variant Y37A had a wild-type amount of lipids; all of the others had less bound lipid than the wild-type, with the variant I72A having the lowest value of 29% (S1 Table). To investigate whether the level of bound lipids influence the function of A46, the I72A mutant of full-length A46 was examined in a NF-κB transcriptional assay in TLR4-expressing HEK293 cells. The A46 I72A mutant was reproducibly expressed at higher levels, both in mammalian cells (Fig 5A) and bacteria. The A46 I72A variant could achieve similar levels of inhibition of NF-κB mediated signalling as the wild-type (Fig 5B); however, this level could only be reached when a 2-to-3-fold excess of A46 I72A was expressed compared to the wild-type variant (Fig 5A and 5B). Thus, the lower lipid binding capacity of A46 I72A impairs its ability to inhibit TLR4 signalling.
In the light of the crystal structure, we analysed the oligomeric state of A46(1–83) using SAXS (Table 3). The theoretical scattering curve of the A46(1–83) tetramer in crystals presents a good fit to experimental data with Chi2 (Crysol [32]) of 0.66 (Fig 6A) versus a very poor fit for the possible dimer found in the asymmetric unit with Crysol Chi2 of 13.52 (Fig 6A).
How are the N-terminal and C-terminal domains of A46 oriented relative to one another? To address this question, we performed SAXS experiments on full-length A46(1–240) (Fig 6B, Table 3). The envelope is shown in Fig 6C, together with the fitting of the N- and C-terminal structures. This arrangement agrees with the tetrameric nature of the A46(1–240) and with proteinase digestion of the linker leading to the production of two domains with almost all proteinases tested [9].
We have determined the first structure of a structured N-terminal extension of a VACV Bcl-2-like immunomodulator; additionally, we also show that it is complexed with myristic acid. The A46(1–83) domain crystallized, forming regular continuous strands in a simple β-sandwich structure with few disordered residues (Fig 3A; S1 Fig). Nevertheless, the solution of the X-ray structure was complex. Due to the crystals' small size, automatic beam focussing at the MASSIF beam line was essential. Additionally, selenium anomalous signals could not be used because of the position of the methionine residues. Molecular replacement also failed due to lack of a known protein structure to be used as search model. However, the regular crystal packing allowed high-resolution data sets to be obtained that were initially processed at 1.55 Å resolution. This high resolution data, together with the short length of the protein, allowed the phases to be solved using ab initio methods [25]. In this method, which has been used successfully for numerous α-helical structures [34], fragments of known structures are employed as small search models. In our study, phases could be solved by a protein fragment of three β-strands that resembles part of the structure of A46(1–83), revealing two molecules in the asymmetric unit.
Refinement of the structure allowed the determination of electron density for residues 1–76 of subunit A and 1–66 in subunit B. The electron density showed clearly that both subunit A and subunit B were comprised entirely of β-sheets, confirming previous bioinformatic predictions that the N-terminus of A46 has a β-sheet arrangement. Unexpectedly, in the subunit B, the C-terminal strand β7 is disordered and not visible in the electron density. We propose that this difference allows A46(1–83) to form tetramers via the subunit B interfaces whilst interacting with ligands through the subunit A interfaces.
A further wholly unexpected feature of A46(1–83) is a partially hydrophobic cavity which spans the entire subunit A and part of subunit B. The cavity is open on the side of the A interface and accommodates long chain fatty acids that were co-purified from the E.coli cell lysate. In the X-ray structure of A46(1–83), clear electron density for C14:0 myristic acid was found (Fig 4A), with the hydrophobic tail buried in the cavity whereas the carboxyl group is open to the solvent. Such an orientation leads us to hypothesize that the cavity might serve as a specific binding pocket for myristoylated binding partners. To this end, TRAM is the only binding partner of A46 known to be myristoylated; myristoylation is indeed essential for its innate immune function, providing correct location of TRAM to the membranes [35]. Binding of A46 to the myristate of TRAM would prevent the insertion of TRAM into the membrane and thus circumvent intracellular signalling. An acceptable alternative hypothesis would be that the bound fatty acids induce asymmetry of the A46(1–83) dimer, as they block a polymerization interface equivalent to B/B and prevent binding of a second fatty acid copy in subunit B. In such manner, using the same primary sequence, a dimer of heterodimers is formed that allows utilization of different interfaces for distinct functions such as tetramerization (interface B with 6 β-strands only) or binding of cellular targets (interface A with 7 β-strands).
The I72A mutant of A46, which binds lower amounts of fatty acids, indeed showed a reduced ability to inhibit TLR4-stimulated NF-κB-driven transcription compared to the wild-type protein (Fig 5B). The inhibitory level of the wild-type A46 was achieved by the I72A mutant when higher quantities of the mutant protein were expressed. This is not unexpected, as the C-terminal domain alone (A46(87–229)) can bind TIR/TRAM with KD of 3.6 μM (Table 1). Presumably, at higher concentrations, the C-terminus of the A46 I72A mutant can compensate for the loss of binding of the lipid-containing N-terminal domain. Pertinently, we have shown that the C-terminus of A46 alone is capable of efficiently inhibiting MyD88-mediated NF-κB activation when IL-1β stimulation system is used [9]. TLR4-stimulated activation of the NF-κB transcription factor involves both TRAM and MyD88-dependent cascades [36]; taken together, our data suggest that the N-terminal domain of A46 may play a more appreciable role in the inhibition of the TRAM pathway than the MyD88 pathway.
To find similarities of A46(1–83) to other known folds, we searched the PDB database with PDBeFold [37], using subunit A of A46(1–83) as search query. The highest match corresponded to the nuclear movement protein from E. cuniculi GB-M1 (PDBID 2O30, chain B). Six secondary structure elements were aligned involving 57 residues at an rmsd of 3.39 Å for 45 Cα; however, the mutual orientation of both sheets is markedly different and the CS domain seen in NudC does not show oligomerisation. Therefore, we searched for similar local folds of the same connectivity using the same core of 45 residues with the program BORGES [25]. The closest match for the strands of ligand bound subunit A was extracted from 2XN2, with 3.09 Å rmsd, whereas for subunit B, a fold extracted from 2OQE gave 2.42 Å. No instance could be identified of an equivalent fold showing the same structural change upon ligand binding; nevertheless, a survey of the hits revealed recurring instances of carbohydrate binding proteins, proteins forming pores and participating in the proper insertion of periplasmic proteins into membranes. These include proteins such as YidC (PDBID 3BLC), located in the periplasmic space of E.coli that could, theoretically, bind lipidated proteins. The geometry of the local fold described by the 6 sheets is also close to a part found in pore-forming hemolysins and leucodines. Indeed, the S-F heterodimer in the latter ones achieves asymmetry through the association of two components of very different sequence but very close geometry, with up to one C-terminal strand present in only one of the copies [38].
The structure of the A46(1–83) protein illuminates the oligomerization state of both the N-terminal extension and the full-length protein. Previous data had indicated the presence of a tetramer in solution for the full-length A46 and a dimer for the Bcl-2-like C-terminal domain [9]. The structure of the N-terminal extension shows a tetramer formed by the association over a crystallographic twofold axis of the two copies present in the asymmetric unit (S1 Fig), evaluated to be persistent under physiological conditions. SAXS analysis confirmed that that A46(1–83) is tetrameric in solution (Fig 6A).
For full-length A46 in solution, structural information on the separated N- and C-terminal domains allowed interpretation of the envelope generated by SAXS. The full-length molecule has an elongated shape, with the N- and C- domains linked by a flexible, proteolytically sensitive linker that allows movement of the two domains relative to each other (Fig 6C). Rigid body fitting of the structures in the envelope of the full-length A46 using CORAL software [39] indicated a movement of 90 degrees between the two domains.
What are the implications of this structural data for the function of A46 in inhibiting signalling through the TRAM and MyD88 linked pathways? We note that both the N-terminal and C-terminal domains of A46 can bind the TIR domains of MyD88 (Fig 7A), although the binding of the N-terminal domain is tenfold lower and the expression of this domain alone does not inhibit IL-1 induced NF-κB mediated signalling in cells (Table 1, Fig 2). However, we suggest that this bipartite binding enables A46 to generate a chain around the TIR domain of MyD88 that would prevent the association of its death domain to assemble the Myddosome, an important structure in the development of the inflammatory response [40]. Additionally, we propose the binding of the myristate post-translational modification of TRAM by the N-terminal domain of A46, with the remainder of the TIR domain of TRAM being bound by the C-terminal Bcl-2 domain of A46 (Fig 7A). For the TIR domain of MAL, an interaction was only observed with the C-terminal domain of A46; the binding site on A46 for the TIR domain of TRIF has not yet been determined. We propose here that the interaction is only with the C-terminal domain (Fig 7A).
The above model assumes binding of only one single TIR domain to the A46 tetramer. However, as depicted in Fig 7B, each tetramer can theoretically present four binding sites for TIR domains. We speculate therefore that A46 could form complexes with multiple binding partners. Indeed, it can even be envisaged that one molecule of A46 could bind one molecule each of MyD88, MAL, TRAM and TRIF (Fig 7B, right side). Thus, even with low initial concentrations of A46, this arrangement would serve to strongly inhibit the inflammatory response by keeping MyD88 death domains apart, preventing proper cellular localization of TRAM and sequestering the other signalling and adaptor molecules. Future experimentation will show the accuracy of these predictions.
The cloning of the plasmid containing full-length sequence of the A46R gene from the VACV Western Reserve strain plasmids (NCBI Gene ID:3707702) as well as of those encoding TIR domains of mammalian MAL and murine MyD88 was described previously [9]. The N-terminal portion of A46 was amplified from the plasmid containing full-length A46 [9] at different length using following primers for the indicated fragments: F: 5’-CGCAAGCCATGGCACAGCAAATGGCGTTTGATATATC-3’ and R: 5’-GCCCGGATCCTTAACT ATACTTATTATACAAGTAAGTC-3’ for the fragment A46(1–90); F: 5’-CGCAAGCCA TGGCACAGCAAATGGCGTTTGATATATC-3’ and R: 5’-GCCCGGATCCTTAAGTCATACTAA CCGGCGTATTAAC-3’ for the fragment A46(1–83); and F: 5’-CGCAAGCCATGGCACAGCA AATGGCGTTTGATATATC-3’ and R: 5’-GCCCGGATCCTTAACCAATATTAGTTTCCTCTG-3’ for the fragment A46(1–73). Obtained fragments were digested with NcoI and BamHI restriction enzymes and ligated into the pET-TRX1 containing HIS6-TEV-thioredoxin as an expression tag.
To generate variants of A46(1–83) to examine their lipid-binding properties, we performed PCR mutagenesis using the pTRX-A46(1–83) plasmid as a template and the following primers: F3D, F: 5’-GCAAATGG CGGATGATATATCAG -3’ and R: 5’- CTGATATATCAT CCGCCATTTGC -3’; H36L F: 5’- GTTAATGATACACTCTACACTGTCG -3’ and R: 5’- CGACAGTGTAGAGTGTATCATTAAC -3’; Y37A, F: 5’-GATACACACGCCACTGTCGA-3’ and R: 5’-TCGACAGTGGCGTGTGTATC-3’; Y37W, F: 5’-GATACAC ACTGGACTGTCGAATTTG -3’ and R: 5’-CAAATTCGACAGTCCAGTGTGTATC-3’; I72A, F: 5’- GAAACTAATGCTGGTTGCGCGG -3’ and R: 5’- CCGCGCAACCAGCATT AGTTTC -3’. The generated PCR products were digested with DpnI and subsequently transformed in E.coli TOP10 competent cells.
For expression in mammalian cells, the plasmids coding for the full-length A46 and C-terminal portion of A46 with the respective tags were cloned previously [9]. For the cloning of the N-terminal domain of A46(1–83), the gene was obtained by amplification from the plasmid carrying the full-length A46 with the primers F: 5’-GCCCGAATTCCGAGAATGGAGCAGAAACTCA TCTCTGAAGAGGATCTGGCGTTTGATATATC-3’ and R1: 5’-CCGCTCGAGTTACTTA TCGTCGTCATCCTTGTAATCAGTCATACTAACCGGCG-3’ or R2: 5’- CCGCTCGAGTTA AGTCATACTAACCGGCG-3’ to yield myc-A46(1–83)-FLAG or myc-A46(1–83), respectively. The amplified DNA fragments were digested with XhoI and EcoRI restriction enzymes and ligated into the pCAGGS vector [41].
The plasmid encoding a GST-fusion of the TIR domain of human TRAM (amino acid residues 66–235) was a kind gift from Dr. H. Tochio [42].
Expression and purification of full-length A46, TIR/MyD88 and TIR/MAL were performed as described previously [9].
E. coli BL21 (DE3) competent cells were transformed with the plasmids coding for the variants of the N-terminal domain of A46. The expression was performed in 2 liters of LB medium containing kanamycin (50 mg/liter). The cells were grown at 37°C until the mid-log phase (A600 = 0.6). Expression was induced with 0.25 mM isopropyl 1-thio-β-D-galactopyranoside at 23°C. After 4 hours, cells were harvested and resuspended in 20 mM Tris-HCl, pH 8.5, 100 mM NaCl, 25 mM imidazole, 5% glycerol and 10 mM β-mercaptoethanol. An EmulsiFlex C3 homogenizer (Avestin) was used for cell lysis. The soluble phase was cleared from insoluble material by centrifugation at 18000 rpm for 30 min. Recombinant proteins were bound to Ni-NTA agarose (5 Prime) charged with 300 mM NiCl2 and pre-equilibrated with lysis buffer. Resin was washed with five column volumes of lysis buffer and proteins of interest were eluted in three column volumes of 20 mM Tris-HCl, pH 8.5, 300 mM NaCl, 200 mM imidazole and 15 mM β-mercaptoethanol. Recombinant TEV protease was added to release A46 domains by proteolysis during overnight dialysis against 20 mM Tris-HCl, pH 8.5, 150 mM NaCl, 10 mM imidazole and 15 mM β-mercaptoethanol. The protein of interest was separated from the protease and the tag by four passages through Ni-NTA resin pre-equilibrated with the dialysing buffer. The resulting protein solution was dialysed against 20 mM Tris-HCl, pH 8.5 and 2 mM DTT for 2 hours. SEC with a HiLoad 16/60 Superdex 75 (GE Healthcare) was performed as final purification step in 20 mM Tris-HCl, pH 8.5 and 10 mM DTT. The concentration of the protein of interest was measured by NanoDrop ND-1000 (Thermo Scientific). The accuracy of NanoDrop measurements was confirmed by additional measurement of the concentration of two samples from independent purifications using BCA Protein Assay Reducing Agent Compatible kit (Thermo Scientific) as described by the manufacturer.
Microscale thermophoresis protein-protein interaction studies were performed on the Monolith NT.115 (Nanotemper Technologies, Munich) using fluorescently labeled proteins as described [43, 44]. For the TIR/MAL, TIR/MyD88 and TIR/TRAM protein labeling, the standard labeling kit for the fluorescent dye Alexa Fluor 647 from Nanotemper was used. Solutions of unlabelled A46(1–229), A46(87–229) and A46(1–83) were serially diluted from 150–450 μM to 8–20 nM in the presence of 30–70 nM of one of the labeled TIR/MAL, TIR/MyD88 or TIR/TRAM proteins. Measurements were performed at 25°C in 20 mM TrisHCl pH 8.5, 100 mM NaCl, 5% glycerol, 1 mM TCEP, 1 mM EDTA, 0.05% Tween 20 using 50% LED power and 60% or 80% IR-laser power. Data analysis was performed with Nanotemper analysis software, v.1.2.101.
Crystals of A46(1–83) and A46(1–73) were initially obtained at protein concentration of 6.75 and 3.5 mg/ml, respectively, in 20 mM TrisHCl pH 8.5 and 10 mM DTT in multiple buffer formulations of the PACT Premier crystallization screen (Molecular Dimensions, Suffolk, UK) using the sitting-drop vapour diffusion technique and a nanodrop-dispensing robot (Phoenix RE; Rigaku Europe, Kent, United Kingdom). We obtained crystals of both protein constructs; however, for A46(1–73) the crystals were not amenable for diffraction experiments. For A46(1–83), the largest crystals grown in 100 mM HEPES 7.0, 20% PEG6000 and 0.2 M of one of following salts NaCl, LiCl or NH4Cl were mounted in the loop and flash-cooled in liquid nitrogen. Crystals with selenium methionine labeled A46(1–83) were obtained in the same buffer formulations. The diffraction data set was collected at 100K at the peak of Se at λ = 0.979 Å at the beamline MASSIF-1 ID30A-1 at the European Synchrotron Radiation Facility (Grenoble, France) to 1.55 Å resolution and processed using the XDS package [45]. Crystals belonged to the space group C2 (a = 65.79 Å b = 59.5 Å c = 47.26 Å).
The structure was solved by ARCIMBOLDO_BORGES ab initio phasing software [25] combining fragment search with Phaser [26] and density modification with SHELXE [46] on the supercomputer Gordon at the SDSC.
Autobuilding was carried out using the program AutoBuild from the Phenix package [47]. The structure was refined using the program Phenix Refine [48] and manual adjustments with the software Coot [49]. Stereo-chemistry and structure quality were checked using the program MolProbity [50]. Data collection and refinement statistics are reported in Table 2.
The coordinates of the A46(1–83) X-ray structure have been deposited in the Protein Data Bank (PDB) database, accession number 5EZU. The experimental SAXS data and derived models of the either full-length A46 or its N-terminal domain have been deposited in small angle scattering biological data bank (SASBDB) with the deposition codes SASDBL7 and SASDBK7.
SAXS experiments for the A46(1–83) and full-length A46(1–229) were performed at 0.9918 Å wavelength ESRF at BioSAXS beamline BM29 coupled to the Superdex 200 10/300 exclusion column (Grenoble, France) and equipped with PILATUS 1M detector at 2.867 m distance from the sample, 0.04 < q < 0.5 Å-1 (q = 4π sin θ/λ, 2θ is the scattering angle). The data were collected using protein concentrations of 15.5 and 4.4 mg/ml for the A46(1–83) and A46(1–240), respectively. The samples were in a buffer containing 20 mM Tris-HCl pH 8.5, 10mM DTT and the measurements were performed at 20°C. The data were processed and analyzed using the ATSAS program package [51]. The radius of gyration Rg and forward scattering I(0) were calculated by Guinier approximation. The maximum particle dimension Dmax and P(r) function were evaluated using the program GNOM [52]. To demonstrate the absence of concentration dependent aggregation and interparticle interference in the both SAXS experiments, we inspected Rg over the elution peaks and performed our analysis only on a selection of frames in which Rg was most stable (S4 Fig). Overall, such stability of Rg over the range of concentrations observed in the SEC elution indicates that there were no concentration-dependent effects or interparticle interference. The data collection and structural parameter from SAXS analysis are summarized in Table 3. The ab initio models were derived using DAMMIF [53]. 40 individual models were created for each run, which were then overlaid and averaged using DAMAVER. For the oligomeric state assessment, the theoretical scattering from either theoretical dimer or tetramer using the high-resolution structure (5EZU) was performed. First, the residues missing in the crystal structure were added by CORAL modelling; later the theoretical scattering curves were generated using CRYSOL and compared to the SAXS experimental data for A46(1–83). To obtain a pseudo-atomic model of the full-length A46, CORAL [39] software was used with the structures for A46(1–83) (5EZU) connected by dummy residue linkers to A46(87–229) (4LQK); the C-terminal domain A46(87–229) was extended by 16–19 dummy residues to imitate the full length of the A46 protein.
Cell culture and reporter gene assays were performed as reported previously [9]. The following expression plasmids were used: the full-length myc-A46(1–240)-FLAG (amounts 200, 150 ng), the N-terminal domain A46(1–83)-FLAG (100, 50 ng), the C-terminal domain A46(87–229)-FLAG (400, 300 ng). The amount of DNA per well was kept constant at 500 ng by supplementation with pCAGGS empty vector.
Human embryonic kidney cells 293 stably transfected with TLR4 or MD2 were kind gifts from Dr. Sylvia Knapp. HEK293-TLR4 and HEK293-MD2 were maintained in DMEM supplemented with 10% fetal calf serum, 1% penicillin/streptomycin and 0.5 mg/ml geniticin G418. To perform a reporter assay with a wild-type or lipid-binding mutant of A46, HEK293-TLR4 cells were grown in 24-well plates and transfected with 80 ng pNF-κB-luc reporter plasmid (Firefly luciferase), 20 ng of pRL-TK (Renilla luciferase) internal control and 300 ng of the respective A46 containing plasmid. The supernatant from HEK293-MD2 cells was filtered and added in a ratio of 1:4 with DMEM to HEK293-TLR4 in the stimulation assay. 40 hours post transfection, HEK293-TLR4 cells were stimulated by addition of MD-2 supernatant, DMEM and 500 ng/ml of LPS. After 7 h, cells were collected, lysed in Passive Lysis Buffer (Promega) and whole cell lysates were analyzed for luciferase activity using the Dual-Luciferase Reporter Assay (Promega). Firefly luciferase activity was normalized by Renilla luciferase activity.
Expression levels of myc-A46-FLAG and myc-A46-FLAG I72A in HEK293T cells were estimated by western blotting. Myc-tagged A46 variants were detected with monoclonal anti-myc 4A6 antibody at a dilution of 1:1000 (Millipore), ϒ-tubulin was used as a loading control and detected with monoclonal anti-tubulin GTU-88 antibody at a 1:5000 dilution (Sigma).
Lipid extractions from purified recombinant proteins were achieved by two different methods. Method A: 3 successive vigorous extractions with 10 volumes of diethyl ether after treatment with 2 volumes of 6M HCl overnight. The ether extracts were evaporated under nitrogen and analysed by electrospray mass spectrometric and tatty acid methyl ester analysis as described below.
Method B: 3 successive vigorous extractions with ethanol to fully denature proteins (final 90% v/v) [54]. The pooled extracts were dried by nitrogen gas in a glass vial and analysed by electrospray mass spectrometry.
For electrospray mass spectrometry analysis, extracts were analyzed on a Absceix 4000 QTrap, a triple quadrupole mass spectrometer equipped with a nanoelectrospray source as described previously [55].
Quantification of the fatty acids from method A were done by conversion to the corresponding fatty acid methyl esters (FAME) followed by GC-MS analysis as described previously [56] using the following GC temperature program: 70°C for 12 min followed by a gradient to 220°C at 4°C/min and held at 220°C for a further 10 min. Mass spectra were acquired from 50–500 amu. The identity of FAMEs was carried out by comparison of the retention time and fragmentation pattern with mixtures of FAME standards.
|
10.1371/journal.pcbi.1001055 | Benchmarking Ontologies: Bigger or Better? | A scientific ontology is a formal representation of knowledge within a domain, typically including central concepts, their properties, and relations. With the rise of computers and high-throughput data collection, ontologies have become essential to data mining and sharing across communities in the biomedical sciences. Powerful approaches exist for testing the internal consistency of an ontology, but not for assessing the fidelity of its domain representation. We introduce a family of metrics that describe the breadth and depth with which an ontology represents its knowledge domain. We then test these metrics using (1) four of the most common medical ontologies with respect to a corpus of medical documents and (2) seven of the most popular English thesauri with respect to three corpora that sample language from medicine, news, and novels. Here we show that our approach captures the quality of ontological representation and guides efforts to narrow the breach between ontology and collective discourse within a domain. Our results also demonstrate key features of medical ontologies, English thesauri, and discourse from different domains. Medical ontologies have a small intersection, as do English thesauri. Moreover, dialects characteristic of distinct domains vary strikingly as many of the same words are used quite differently in medicine, news, and novels. As ontologies are intended to mirror the state of knowledge, our methods to tighten the fit between ontology and domain will increase their relevance for new areas of biomedical science and improve the accuracy and power of inferences computed across them.
| An ontology represents the concepts and their interrelation within a knowledge domain. Several ontologies have been developed in biomedicine, which provide standardized vocabularies to describe diseases, genes and gene products, physiological phenotypes, anatomical structures, and many other phenomena. Scientists use them to encode the results of complex experiments and observations and to perform integrative analysis to discover new knowledge. A remaining challenge in ontology development is how to evaluate an ontology's representation of knowledge within its scientific domain. Building on classic measures from information retrieval, we introduce a family of metrics including breadth and depth that capture the conceptual coverage and parsimony of an ontology. We test these measures using (1) four commonly used medical ontologies in relation to a corpus of medical documents and (2) seven popular English thesauri (ontologies of synonyms) with respect to text from medicine, news, and novels. Results demonstrate that both medical ontologies and English thesauri have a small overlap in concepts and relations. Our methods suggest efforts to tighten the fit between ontologies and biomedical knowledge.
| Controlled terminologies and ontologies are indispensable for modern biomedicine [1]. Ontology was historically restricted to philosophical inquiry into the nature of existence, but logicians at the turn of the 20th Century translated the term into a precise representation of knowledge using statements that highlight essential qualities, parts and relationships [2]. In the early 1970's, explicit approaches to knowledge representation emerged in artificial intelligence [3], and in the 1990's were christened ontologies in computer science [4]. These representations were promoted as stable schemas for data—a kind of object-oriented content—to facilitate data sharing and reuse. Ontologies have since been used intensively for research in biomedicine, astronomy, information science and many other areas. Biomedical scientists use ontologies to encode the results of complex experiments and observations consistently, and analysts use the resulting data to integrate and model system properties. In this way, ontologies facilitate data storage, sharing between scientists and subfields, integrative analysis, and computational reasoning across many more facts than scientists can consider with traditional means.
In addition to their computational utility, key biomedical ontologies serve as lingua franca: they allow numerous researchers to negotiate and agree on central, domain-specific concepts and their hierarchical interrelations. Concepts commonly modeled with ontologies include organismal phenotypes [5]–[7] and gene functions in genetics and genomics [1], [8]; signs, symptoms and disease classifications in medicine [9]; species, niche names and inter-species relations in ecology and evolution [10]. Building an ontology in any of these areas faces similar challenges: lack of an external standard that defines the most critical concepts and concept linkages for the ontology's proposed function; vast numbers of aliases referring to the same concept; and no yardstick with which to compare competing terminologies. This paper considers scientific ontologies generally and then develops a framework and validates a family of measures that helps to overcome these challenges.
The word ontology historically represented the product of one person's philosophical inquiry into the structure of the real world: What entities exist? What are their properties? How are they grouped and hierarchically related?
While this original definition still holds in philosophy, the computational interpretation of an ontology is a data structure typically produced by a community of researchers through a procedure that resembles the work of a standards-setting committee or a business negotiation (L. Hunter, 2010, personal communication). To agree on the meaning of shared symbols, the process involves careful utility-oriented design. The collective ontologies that result are intended to be used as practical tools, such as to support the systematic annotation of biomedical data by a large number of researchers. A standard domain-specific ontology used in the sciences today includes a set of concepts representing external entities, a set of relations, typically defined as the predicates of statements linking two concepts (such as cat is-an animal, cat has-a tail), and taxonomy or hierarchy defined over concepts, comprised by the union of relations. An ontology may also explicitly represent a set of properties associated with each concept and rules for these properties to be inherited from parent to child concept. Furthermore, formal ontologies sometimes incorporate explicit axioms or logical constraints that must hold in logical reasoning over ontology objects.
In practice, what different research groups mean by the term ontology can range from unstructured terminologies, to sets of concepts and relations without complete connection into a hierarchy, to taxonomies, to consistent, formal ontologies with defined properties and logical constraints.
An ontology developed by group represents a glimpse into the specific worldviews held within that group and its broader domain. By the same logic, we can consider the union of all published articles produced by a scientific community as a much more complete sample of scientific worldviews. While a research team that writes a joint paper agrees on its topic-specific worldview to some extent, its collective domain ontology is neither explicitly defined, nor free from redundancy and contradiction. Insofar as scientists communicate with each other and respond to prior published research, however, these worldviews spread and achieve substantial continuity and homogeneity [11]. A large collection of scientific documents therefore represents a mixture of partially consistent scientific worldviews. This picture is necessarily complicated by the flexibility and imprecision of natural language. Even when scientists agree on specific concepts and relations, their corresponding expressions often differ, as the same meaning can be expressed in many ways.
Nevertheless, if we accept that the published scientific record constitutes the best available trace of collective scientific worldviews, we arrive at the following conclusion: Insofar as an ontology is intended to represent knowledge within a scientific domain, it should correspond with the scientific record. Moreover, an ontology would practically benefit from evaluation and improvement based on its match with a corpus of scientific prose that represents the distribution of its (potential) users' worldviews.
Previously proposed metrics for ontology evaluation can be divided into four broad categories: Measures of an ontology's (1) internal consistency (2) usability (or task-based performance), (3) comparison with other ontologies and (4) match to reality. While this review is necessarily abbreviated, we highlight the most significant approaches to ontology evaluation.
Metrics of an ontology's internal consistency are nicely reviewed by Yu and colleagues [12]. They especially highlight: clarity, coherence, extendibility, minimal ontological commitment, and minimal encoding bias [4]; competency [13]; consistency, completeness, conciseness, expandability, and sensitiveness [14]. The names of these metrics suggest their purposes. For example, conciseness measures how many unique concepts and relations in an ontology have multiple names. Consistency quantifies the frequency with which an ontology includes concepts that share subconcepts and the number of circularity errors.
Measurements of an ontology's usability [15]–[17] build on empirical tools from cognitive science that assess the ease with which ontologies can be understood and deployed in specific tasks [18]. Results from such studies provide concrete suggestions for improving individual ontologies, but they are also sometimes used to compare competing ontologies. For example, Gangemi and colleagues [19] described a number of usability-profiling measures, such as presence, amount, completeness, and reliability, that assess the degree to which parts of an ontology are updated by ontologists [19]. The authors also discuss an ontology's “cognitive ergonomics”: an ideal ontology should be easily understood, manipulated, and exploited by its intended users.
Approaches to ontology comparison typically involve the 1) direct matching of ontology concepts and 2) the hierarchical arrangement of those concepts, often between an ontology computationally extracted and constructed from text and a reference or “gold standard” ontology built by experts. Concept comparison draws on the information retrieval measures of precision and recall [12], [20], [21] (sometimes called term [22] or lexical precision and recall [22]; see Materials and Methods section below for precise definitions of precision and recall). Matching ontology terms, however, raises challenging questions about the ambiguity of natural language and the imperfect relationship between terms and the concepts that underlie them. Some ignore these challenges by simply assessing precision and recall on the perfect match between terms. Others deploy string similarity techniques like stemming or edit distance to establish a fuzzy match between similar ontology terms [23], [24].
The second aspect of ontology matching involves a wide variety of structural comparisons. One approach is to measure the Taxonomic Overlap, or intersection between sets of super- and subconcepts associated with a concept shared in both ontologies, then averaged across all concepts to create a global measure [23]–[25]. Another uses these super and subconcept sets to construct asymmetric taxonomic precision and recall measures [26], closely related to hierarchical precision and recall [27], [28]. A similar approach creates an augmented precision and recall based on the shortest path between concepts [29] or other types of paths and a branching factor [30]. An alternate approach is the OntoRand index that uses a clustering logic to compare concept hierarchies containing shared concepts [31]. The relative closeness of concepts is assessed based on common ancestors or path distance, and then hierarchies are partitioned and concept partitions are compared.
Approaches for matching an ontology to reality are more diverse and currently depend heavily on expert participation [12]. For example, Missikoff and colleagues [32] suggested that an ontology's match to reality be evaluated by measuring each ontology concept's “frequency of use” by experts in the community. Missikoff and colleagues' ultimate goal was to converge to a consensus ontology negotiated among virtual users via a web-interface. Smith [33] recommended an approach to ontology evolution which rests on explicitly aligning ontology terms to unique entities in the world studied by scientists. Ontology developers would then be required to employ a process of manual tracking, whereby new discoveries about tracked entities would guide corresponding changes to the ontology. In a related effort, Ceusters and Smith suggested studying the evolution of ontologies over time [34]: they defined an ontology benchmarking calculus that follows temporal changes in the ontology as concepts are added, dropped and re-defined.
A converse approach to matching ontologies with domain knowledge appears in work that attempts to learn ontologies automatically (or with moderate input from experts) from a collection of documents [35]–[38] using machine learning and natural language processing. The best results (F-measure around 0.3) indicate that the problem is extremely difficult. Brewster and colleagues [36], [39] proposed (but did not implement) matching concepts of a deterministic ontology to a corpus by maximizing the posterior probability of the ontology given the corpus [36], [39]. In this framework, alternative ontologies can be compared in terms of the posterior probability conditioned on the same corpus. Their central idea, which shares our purpose but diverges in detail, is that “the ontology can be penalized for terms present in the corpus and absent in ontology, and for terms present in the ontology but absent in the corpus” (see also [19]). Each of these approaches to mapping ontologies to text face formidable challenges associated with the ambiguity of natural language. These include synonymy or multiple phrases with the same meaning; polysemy or identical expressions with different meanings; and other disjunctions between the structure of linguistic symbols and their conceptual referents.
In summary, among the several approaches developed to evaluate an ontology's consistency, usability, comparison and match to reality, metrics that evaluate consistency are the most mature among the four and have inspired a number of practical applications [40]–[42]. The approach that we propose and implement here belongs to the less developed areas of matching ontologies to each other and to discourse in the world. When considering approaches that compare ontologies to each other and to discourse, metrics comparing ontologies to one another jump from the comparison of individual concepts to the comparison of entire concept hierarchies without considering intermediate concept-to-concept relationships. This is notable because discourse typically only expresses concepts and concept relationships, and so the measures we develop will focus on these two levels in mapping ontologies to text.
Our purpose here is to formally define measures of an ontology's fit with respect to published knowledge. By doing this we attempt to move beyond the tradition of comparing ontologies by size and relying on expert intuitions. Our goal is to make the evaluation of an ontology computable and to capture both the breadth and depth of its domain representation—its conceptual coverage and the parsimony or efficiency of that coverage. This will allow us to compare and improve ontologies as knowledge representations. To test our approach, we initially analyzed four of the most commonly used medical ontologies against a large corpus of medical abstracts. To facilitate testing multiple ontologies in reference to multiple domains we also analyzed seven synonym dictionaries or thesauri—legitimate if unusual ontologies [43]—and compared their fit to three distinctive corpora: medical abstracts, news articles, and 19-century novels in English.
Medical ontologies have become prominent in recent years, not only for medical researchers but also physicians, hospitals and insurance companies. Medical ontologies link disease concepts and properties together in a coherent system and are used to index the biomedical literature, classify patient disease, and facilitate the standardization of hospital records and the analysis of health risks and benefits. Terminologies and taxonomies characterized by hierarchical inclusion of one or a few relationship types (e.g., disease_conceptx is-a disease_concepty) are often considered lightweight ontologies and are the most commonly used in medicine [44], [45]. Heavyweight ontologies capture a broader range of biomedical connections and contain formal axioms and constraints to characterize entities and relationships distinctive to the domain. These are becoming more popular in biomedical research, including the Foundational Model of Anatomy [46] with its diverse physical relations between anatomical components.
The first, widely used medical ontology was Jacques Bertillon's taxonomic Classification of Causes of Death, adopted in 1893 by the International Statistical Institute to track disease for public health purposes [47]. Five years later, at a meeting of the American Public Health Association in Ottawa, the Bertillon Classification was recommended for use by registrars throughout North America. It was simultaneously adopted by several Western European and South American countries and updated every ten years. In the wake of Bertillon's death in 1922, the Statistics Institute and the health section of the League of Nations drafted proposals for new versions and the ontology was renamed the International List of Causes of Death (ICD). In 1938 the ICD widened from mortality to morbidity [48] and was eventually taken up by hospitals and insurance companies for billing purposes. At roughly the same time, other ontologies emerged, including the Quarterly Cumulative Index Medicus Subject Headings, which eventually gave rise to the Medical Subject Headings (MeSH) that the NIH's National Library of Medicine uses to annotate biomedical research literature [49], [50]. By 1986 several medical ontologies were in wide use and the National Library of Medicine began the Unified Medical Language System (UMLS) project in order to link many of them to facilitate information retrieval and integrative analysis [51]. By far the most frequently cited ontology today in biomedicine is the Gene Ontology (GO), a structurally lightweight taxonomy begun in 1998 that now comprises over 22,000 entities biologists use to characterize gene products [52].
We propose to further test and evaluate our ontology metrics using the fit between a synonym dictionary or thesaurus and a corpus. A thesaurus is a set of words (concepts) connected by synonymy and occasionally antonymy. Because synonymy constitutes an is-equivalent-to relationship (i.e., wordx is-equivalent-to wordy), thesauri can be viewed as ontologies, albeit rudimentary ones. Moreover, because a given thesaurus is intended to describe the substitution of words in a domain of language, the relationship between a thesaurus and a corpus provides a powerful model for developing and testing general measures of the fit between ontology and knowledge domain. Most useful for our purposes, the balance between theoretical coverage and parsimony is captured with the thesauri model: A bloated 100,000 word thesaurus is clearly not superior to one with 20,000 entries efficiently tuned to its domain. A writer using the larger thesaurus would not only be inconvenienced by needing to leaf through more irrelevant headwords (the word headings followed by lists of synonyms), but be challenged by needing to avoid inappropriate synonyms.
Synonymy is transitive but not necessarily symmetric – the headword is sometimes more general than its substitute. Occasionally thesauri also include antonyms, i.e., is-the-opposite-of, but fewer words have antonyms and for those that do, antonyms listed are far fewer than synonyms.
A typical thesaurus differs from a typical scientific ontology. While ontologies often include many types of relations, thesauri contain only one or two. Thesauri capture the natural diversity of concepts but are not optimized for non-redundancy and frequently contain cycles. Any two exchangeable words, each the other's synonym, constitute a cycle. As such, thesauri are not consistent, rational structures across which strict, logical inference is possible. They instead represent a wide sample of conflicting linguistic choices that represent a combination of historical association and neural predisposition. Despite these differences, we believe thesauri are insightful models of modern, domain-specific ontologies. Working with thesauri also contributes practically to evaluating the match between ontologies and discourse. Because all of our measures depend on mapping concepts from ontology to text, assessment of the match between thesaurus and text can directly improve our identification of ontology concepts via synonymy.
Our proposed approach to benchmarking an ontology X with respect to a reference corpus T is outlined in Figure 1. The essence of the approach requires mapping concepts and relations of the test ontology to their mentions in the corpus – a task as important as it is difficult [53]. Given this mapping, we show how to compute ontology-specific metrics, Breadth and Depth, defined at three levels of granularity (see Materials and Methods). We also define another important concept – the perfect ontology with respect to corpus T. This ideal ontology represents all concepts and relations mentioned in T and can be directly compared to X. If corpus T is sufficiently large, the perfect ontology is much larger than the test ontology X. This allows us to identify a subset of the perfect ontology that constitutes the fittest ontology of the same size as test ontology X –the one with maximum Breadth and Depth. Finally, given knowledge about the fittest ontology of fixed size and metrics for the test ontology X, we can compute loss metrics, indicating how much ontology X can be improved in terms of its fit to the corpus. All definitions are provided in the Materials and Methods section.
To demonstrate our approach to the comparison of biomedical ontologies, we identified concepts associated with disease phenotypes and relations in four medical ontologies: ICD9-CM [48], [54], CCPSS [55], SNOMED CT [56] and MeSH (see Table 1 and Figure 2). Comparing each medical ontology concept-by-concept (as assessed with UMLS MetaMap—see Materials and Methods), we found that despite a reasonable overlap in biomedical terms and concepts, different ontologies intersect little in their relations (see Figure 2 A and B). This suggests that each ontology covers only a small subset of the full range of possible human disease concepts and circumstances. This likely results from the different ways in which each ontology is used in biomedicine.
To evaluate the fit between an ontology and a corpus, we first estimated the frequency of ontology-specific concepts and relations in the corpus. We mapped ontology concepts to the biomedical literature and then estimated their frequency using MetaMap, which draws on a variety of natural language processing techniques, including tokenization, part-of-speech tagging, shallow parsing and word-sense disambiguation [57]. We then estimated the frequency of concept relations in the literature (see Materials and Methods). We parameterized these relation frequencies as the probability that two concepts co-occur within a statement in our medical corpus (see Table 2, Materials and Methods).
Our measures of ontology representation build on established metrics from information retrieval (IR), which have been previously used in ontology comparison. IR tallies the correspondence between a user's query and relevant documents in a collection: When the subset of relevant documents in a collection is known, one can compute IR metrics such as recall, precision and their harmonic mean, the F-measure, that capture the quality of a query in context (see Materials and Methods). We compute these measures as first-order comparisons between ontologies in terms of whether concept-concept pairs “retrieve” contents from the corpus.
The major rift between IR metrics and the nature of ontologies lies in the binary character of IR definitions: IR measures weight all relations in an ontology equally, but concepts and relations from an ontology vary widely in their frequency of usage within the underlying domain. Further, unlike IR documents retrieved from a query, concepts and relations present in an ontology but not a corpora should not be considered “false positives” or nonexistent in scientific discourse. Unless the ontology contains explicit errors, it is reasonable to assume that by expanding the corpus, one could eventually account for every ontology relation. Formulated differently, we cannot justifiably classify any ontology relation as false, but only improbable. This logic recommends we avoid IR measures that rely on false-positives (e.g., precision) and augment the remaining metrics to model theoretical coverage and parsimony as functions of concept and relation importance rather than mere existence in the domain of interest.
To do this, we first define the complete ontology that incorporates every concept and relation encountered in a corpus. In our implementation, we approximate this with all of the concepts and relations that appear in the corpus and are identified by UMLS MetaMap with the semantic type “disease or syndrome.” We then define two measures, breadth and depth, to describe the fit between an ontology and a corpus. Breadth2 (see Materials and Methods for definition of several versions of Breadth and Depth) is a generalization of recall that substitutes true-positives and false-negatives with real-valued weights corresponding to the frequency of concepts and the probability of relations in text. Depth2 normalizes breadth by the number of relations in the ontology (see Materials and Methods) and so captures the average probability mass for each ontology relation in the corpus. Large ontologies tend to have better breadth of coverage relative to a corpus, but not necessarily more depth: They may be padded with rare concepts lowering their corpus fit compared with small, efficient ontologies containing only the most frequent ones.
Breadth and depth allow us to compare ontologies of different size, but do not account for the fact that as ontologies grow, each incremental concept and relation necessarily accounts for less of the usage probability in a corpus. To address this challenge, we define the fittest ontology of fixed size (with a predetermined number of relations) such that depth is maximized over all possible concepts and relations. Furthermore, for an arbitrary ontology we can compute its depth loss relative to the fittest ontology of same size (see Materials and Methods). This approach allows us to more powerfully control for size in comparing ontologies.
Our analysis of the disease-relevant subsets of four medical ontologies indicates that CCPSS, despite having the smallest number of concepts and a moderate number of relations, performs comparably or better with respect to our clinical corpus than its larger competitors. When we consider concepts and relations jointly (see Table 3), CCPSS outperforms the three other terminologies in terms of Breadth2 and Relative Depth2, while being second only to MeSH in Depth2. ICD9-CM and SNOMED rank last in Breadth2 and Depth2, respectively. When only concepts (but not relations) are considered (Table 3), SNOMED CT has the greatest Breadth1 and Relative Depth1 but the worst Depth1, whereas MeSH and CCPSS lead in terms of Depth1. It is striking that the relatively small CCPSS matches clinical text equally or better than the three other ontologies. Table 3 also indicates that Depth2 Loss is smallest for the largest ontology, SNOMED CT and that CCPSS is next. Given its small size, CCPSS is still less likely to miss an important disease relation than MeSH or ICD9-CM. ICD9-CM, with the highest Relative Depth1,2 Loss, would benefit most by substituting its lowest probability concepts with the highest probability ones missed.
In order to demonstrate the power of our metrics to capture different dimensions of the fit between ontology and knowledge domain, we compared 7 of the most common English thesauri (see Table 1 and Figure 2) against three corpora that sampled published text from the domains of medicine, news and novels (see Table 2). Our thesauri included (1) The Synonym Finder, (2) Webster's New World Roget's A–Z Thesaurus, (3) 21st Century Synonym and Antonym Finder, (4) The Oxford Dictionary of Synonyms and Antonyms, (5) A Dictionary of Synonyms and Antonyms, (6) Scholastic Dictionary of Synonym, Antonyms and Homonyms, and (7) WordNet (see Materials and Methods).
While comparing multiple thesauri word-by-word, we found a pattern similar to our medical ontologies. Despite a larger overlap in headwords than medical ontology concepts, different dictionaries intersect little in their relations. (A headword in a thesaurus is a word or phrase appearing as the heading of a list of synonyms and antonyms. Not every word or phrase that is listed as a synonym in a thesaurus also occurs as a separate headword.) On average, only one relation per headword is found in all three of the largest dictionaries (see Figures 2 C and D). This trend persists as we consider a longer list of thesauri (see Table 2 in Text S1) and indicates that any single dictionary covers only a small portion of synonyms used in the body of English. But some dictionaries are better than others.
To evaluate the fit between thesaurus and corpus, we estimated the frequencies of thesauri headwords and synonyms in the corpus. We assessed headword frequency as we did with medical ontology concepts. In the case of synonymy relations, we parameterize the synonym frequencies as the probability that a headword is substituted with each of its synonyms within a specific four-word context (see Materials and Methods).
While thesauri typically aim to capture universal properties of language, corpora can be surprisingly dissimilar and sometimes disjoint in their use of words and synonym substitutions. Figures 3 and 4 visualize ten words whose synonym substitution probabilities are most unlike one another across the medicine, news and novels corpora. Some words carry a different semantic sense in each corpus (e.g., cat as feline versus CT scan versus Caterpillar construction equipment), while other words have very different distributions of common senses.
It is illuminating to consider the dominant substitutions for the three corpora: The noun insult translates most frequently to injury in Medicine, slur in News, and shame in Novels; the verb degrade to impair, demean, and depress in the same respective corpora (see Figures 3 and 4); the adjective futile to small, fruitless and vain. In some contexts words are used literally and consistently, while in others, metaphorically and widely varying. The meaning of the noun headache in our medical corpus is always literal: the closest synonyms here are migraine and neuralgia – with no other synonyms used. In novels and news the predominant meaning of headache is metaphorical. Novels are replete with headache's synonym mess, a disordered and problematic situation (i.e., headache-inducing). The news corpus also predominantly uses headache to mean problem, but the most frequent synonyms are more precise and literal (problem, concern, worry, trouble). The metaphorical mess and hassle are also present, but at far lower frequencies than in novels. The verb stretch is treated as equivalent to develop, increase, prolong, and enlarge in the medical corpus. In novels it means open, spread, and draw. The news corpus hosts dozens of distinct synonyms for stretch, the most frequent three being extend, widen, and sprawl.
Figure 5, a–i and table 2 in the Supplement compare all metrics discussed for all seven thesauri and three corpora. From Figure 5 d and g, we observe that our importance-based breadth corresponds to counts-based recall (a). The correspondence is not perfect, however: Oxford and WordNet have greater breadth than 21st Century, but this is reversed in recall. On the other hand, larger thesauri tend to lead in both recall and breadth, but small thesauri excel in precision and depth, as shown in Figure 5 e and h. The rankings of depth across all seven thesauri on three corpora, however, are very different from those of precision, which suggests that depth captures a different internal characteristic of ontology. For fixed precision and recall, we can define multiple equal-sized corpus-matched ontologies with widely varying depth and breadth by sampling from the complete ontology. The converse, however, is not true: Our breadth and depth metrics uniquely define an ontology's precision and recall. Figure 5 f and i indicate that depth loss is negatively correlated with the size of our seven thesauri (see Discussion). This is likely because a large thesaurus nearly exhausts the common relations in all domains by including synonyms that are rare in one context but common in another. Small dictionaries must focus. Unless explicitly tuned to a domain, they are more likely to miss important words in it.
Finally, we can compare corpora to each other with respect to all thesauri. As clearly shown in Figure 5, our three corpora map onto the seven thesauri non-uniformly. Precision, for example, is significantly lower across all thesauri for the medical corpus than for news or novels. This is likely due to the specialized and precise medical sublanguage, which renders a large portion of common synonyms irrelevant.
We introduced novel measures that assess the match between an ontology and discourse. These differ from former approaches to ontology comparison by focusing on concept and concept-to-concept relations, as these are the ontology elements present in textual statements. Moreover, our measures account for conceptual distinctions between comparing ontologies to one another versus to the discourse associated with a knowledge domain. In the latter comparison, the notion of a false positive, or a concept that appears in ontology but not in text is misleading, as it does not necessarily indicate the concept was not in discourse, but that the discourse was insufficiently sampled. Building on these insights, we introduce novel measures that capture the Breadth and Depth of an ontology's match to its domain with three versions of increasing complexity. Breadth is the total probability mass behind an ontology's concepts and relations with respect to the reference corpus. Depth, in contrast, is its average probability mass per concept and relation. Metaphorically, if breadth is “national income,” then depth is “income-per-capita.” An ontology with greater breadth captures more concepts and relations; an ontology with greater depth better captures its most important ones.
By measuring the match between a medical ontology and a corpus of medical documents, we are also assessing the utility of each ontology's terms and relations for annotating that corpus. In this sense, breadth measures the overall utility of a given ontology in annotation, whereas depth measures the average annotation utility per ontology constituent.
We also defined the fittest ontology of fixed size such that depth is maximized over all concepts and relations in order to more carefully compare ontologies of different sizes. For an arbitrary ontology we also computed its depth loss relative to the fittest ontology of same size (see Materials and Methods). This approach not only allows us to control for size in comparing ontologies, but also has direct application for pruning an ontology of its most improbable parts.
To illustrate the meaning and relation of depth loss to depth and breadth, imagine a casino with an enormous roulette wheel on which numbers may appear more than once, and some much more frequently than others. A gambler has limited time to observe the wheel before picking a set of numbers on which to bet. In this analogy, the numbers correspond to concepts and relations in science, the gambler to an ontologist, and a win to an efficient representation of science. The probability of winning or achieving a good scientific representation given a set of bets maps to breadth and the probability of winning normalized by number of bets to depth. The fittest ontology of given size is an optimal bundle of bets: the gambling ontologist can still lose by missing any particular concept or relation, but her risk is minimized. Depth loss, then, is the unnecessary risk of losing a gamble beyond that required by the constrained number of bets. As an ontology grows in size, the overall probability of missing an important scientific concept or relation shrinks. Therefore, depth loss will usually decrease as ontologies grow, even if the smaller ontology has greater depth.
By capturing the breadth and depth of an ontology's coverage, our measures suggest precisely what the analyst gains by assessing the direct match between ontology and discourse, rather than attempting to extract or “learn” an ontology from discourse and subsequently compare it with a reference ontology. When an ontology is developed from discourse, all information about the relative frequency with which concepts and relations occur in the domain is lost. Consequently, a match with such an ontology can only grossly capture the representativeness of relations in the reference ontology. The larger difference between these approaches, however, is in the position of authority. Our measures suggest that discourse is the authoritative source of a community's scientific knowledge and should be the reference against which most scientific ontologies are judged. Measures that assess “learned ontologies” with a gold standard, by contrast, assume that ontologists and their constructions are the ultimate reference.
Our approach to ontology evaluation has several limitations. It may be viewed as restrictive due to its reliance on the availability of a large corpus related to the domain of interest. This is usually not a problem for biomedical ontologies as the amount of biomedical text is typically overwhelming. For esoteric ontologies, however, it may be difficult to locate and sufficiently sample the textual domain they are intended to map. At the extreme, consider a hypothetical ontology configuring entities corresponding to a novel theory.
Further, one can imagine ontologies for which any degree of match to an external domain is meaningless. For example, a hypothetical mathematical ontology should be, first and foremost, clear and internally consistent. As is common in mathematics, relevance to external research may not be required. This level of abstraction and invariance to reality, however, is atypical for biomedicine and other areas of science where the corpus of published research indicates much of what is known.
Our approach addresses only one dimension of ontology quality: its match to collective discourse. Other quality dimensions such as consistency and usability are also clearly important. We do not advocate retiring other views of ontology quality: our measures of external validity can be used synergistically with assessments of internal validity to expand the overall utility of an ontology.
Another limitation of our method is that we assume that formal relations among ontology concepts are represented explicitly in text, like the concepts themselves. As Brewster and colleagues have pointed out [36], this is often not the case. More advanced methods are needed to improve on our use of concept co-occurrence. Our approach depends heavily on the advancement of parsing and mapping technologies to enable linkages between ontology concepts and their textual instances. It is particularly dependent on quality in the part-of-speech tagging, recognition of verb nominalization [58] and the association of inflectional and morphological variations in vocabulary.
In this way, proper application of our proposed method demands that users surmount significant technical hurdles. It is not trivial to map concepts and relations from an ontology to a real corpus considering the ambiguities and complexities of unstructured discourse. Although we believe that these technical problems can be resolved with a reasonable degree of accuracy, there remains a lingering concern that ontology evaluation is confounded by imperfections in the analysis of text. To address this concern, our analysis of synonym substitution probabilities suggests a practical approach for generating probabilistic domain-specific thesauri that can be immediately used in more closely mapping arbitrary ontologies to text. These substitution probabilities can also be deployed to improve the cross-mapping of ontologies, expanding database queries, and text mining.
Several previous approaches to ontology comparison involve explicit comparison of the entire taxonomy of relations. Our approach instead emphasizes comparison of ontology relationships individually. This is because metrics of taxonomic distance between two ontologies [23]–[28] are not easily transplanted to the comparison of ontology with text. Ontology comparisons often weight the match between concepts by the centrality of those concepts in each ontology's hierarchy [26]. The upper-level – the most central and abstract – relations in an ontology, however, are rarely mentioned explicitly in prose. This is partly because of the indexical power of context: an article published in the journal Metabolism does not need to mention or describe metabolism to its audience. The publication alone signals it. In contrast, specific concepts that are taxonomically close to the bottom of the hierarchy – the “leaves” of the tree – are often mentioned in text with disproportionate frequency. In short, while centrality denotes importance within an ontology, and ontology importance should correlate with frequency in discourse, we expect that this relationship is confounded in scientific domains where the most central “branching” concepts are likely so conditioned by context (e.g., a biology journal) that they remain unspoken.
In summary, our measures provide a reliable assessment of ontologies as representations of knowledge. We demonstrated their utility using biomedical ontologies, English thesauri and corpora, and we showed that different corpora call for different representations. We believe our straightforward approach can be extended to arbitrary ontologies and knowledge embedded in the literature of their communities. For example, our approach can directly assess the degree to which other popular ontologies represent published knowledge in their respective domains. Our approach would also recommend how these ontologies could be made more efficient or parsimonious. Finally, our measures facilitate comparison between competing ontologies. In conjunction with efforts to make ontologies logically consistent, greater external validity will insure that ontological inferences anchor to the most salient concepts and relations used by the community of science.
We used four medical ontologies, seven English thesauri (Table 1), and three corpora (Table 2) from the areas of medicine, news, and novels. The four biomedical ontologies we used were ICD9-CM, CCPSS, SNOMED-CT, and MeSH each described in the following paragraphs.
ICD9-CM [48], [54], the International Statistical Classification of Diseases and Related Health Problems, is a taxonomy of signs, symptoms, abnormal findings, complaints, social circumstances, and external causes of injury or disease. It uses predominantly one type of relation (is-a), whereas CCPSS and SNOMED CT employ richer repertoires of relation types. The International Classification of Diseases is published by the World Health Organization (WHO) and is used worldwide for morbidity and mortality statistics, reimbursement systems, and automated decision support in medicine. The ICD9-CM version was created by the U.S. National Center for Health Statistics as an extension of the ICD9 system to include diagnostic and operative procedures – the CM referring to clinically modified. Here we use the 2009 version of ICD9-CM. A typical relation between two concepts in ICD9-CM looks as follows:
CCPSS, the Canonical Clinical Problem Statement System [55], is a knowledge base that encodes clinical problems encountered by ailing humans. It is specifically designed to encode clinical knowledge regarding relations between medical conditions. Typical relations encoded in CCPSS look as follows:
SNOMED CT, Systematized Nomenclature of Medicine – Clinical Terms [56], is a synthesis of terminologies produced by the College of American Pathologists and by the National Health Service of the United Kingdom. The American component is called SNOMED Reference Terminology, and the British one is referred to both as Clinical Terms and Read Codes. SNOMED CT is the most comprehensive clinical terminology in existence and includes ∼350,000 concepts. A typical relation in SNOMED CT looks as follows:
Medical Subject Headings (MeSH) [49] is a comprehensive controlled vocabulary designed by the United States National Library of Medicine (NLM). Its intended use is information retrieval; MeSH was not designed as a formal ontology. The 2009 version contains a total of 25,186 subject headings spanning anatomy; organism classification; diseases; chemicals and drugs; food and beverages; analytical, diagnostic and therapeutic techniques and equipment; health care, psychiatry and psychology; biological and physical sciences; anthropology, education, sociology and social phenomena; persons; technology and information science; humanities; publication characteristics and geographic locations. It is mainly used by the MEDLINE/PubMed article database for indexing journal articles and books. A typical relation present in the MeSH is-a hierarchy looks like
We tested the medical ontologies against a corpora of modern medicine comprised of clinical journal article abstracts from the PubMed database. We limited ourselves only to English abstracts in the core clinical journals for the entire period covered by PubMed, 1945 through February of 2009. The resulting corpus included 786,180 clinical medicine-related abstracts (see Table 2).
Our broader analysis of synonym dictionaries included seven of the most common, sampling from very different kinds of thesauri. These include the large thesauri (1) The Synonym Finder and (2) Webster's New World Roget's A–Z Thesaurus; moderately-sized thesauri (3) 21st Century Synonym and Antonym Finder and (4) The Oxford Dictionary of Synonyms and Antonyms; and portable, compact thesauri (5) A Dictionary of Synonyms and Antonyms and (6) Scholastic Dictionary of Synonym, Antonyms and Homonyms. Each thesaurus shared a common layout involving alphabetically arranged headwords followed by synonyms (and antonyms). Finally, we included the electronic dictionary (7) WordNet, which arranges its words asymmetrically into sets of synonyms or “synsets.”
To evaluate the match between these thesauri and a variety of text corpora, we added English news and novels to our sample of clinical medicine (see Table 2). The news corpus covered all Reuters news stories between 08/20/1996 and 08/19/1997. The novels corpus contained 50 of the most influential novels of the 19th Century, written or translated into English. Complete information regarding each of these data sources can be found in the supplement.
To map biomedical concepts to our clinical corpus we used MetaMap. MetaMap [59] is a knowledge-intensive natural language processing program developed at the National Library of Medicine for mapping snippets of biomedical text to the UMLS Metathesaurus [60], [61].
MetaMap uses the SPECIALIST minimal commitment parser [62] to conduct shallow syntactic parsing of text – using the Xerox part-of-speech tagger. For each identified phrase its variants are generated using the SPECIALIST lexicon and a supplementary database of synonyms. A phrase variant comprises the original phrase tokens, all its acronyms, abbreviations, synonyms, derivational variants, meaningful combinations of these, and inflectional and spelling variants. Given a collection of phrase variants, the system retrieves from the Metathesaurus a set of candidate strings each matching one of the variant constituents. Each Metathesaurus-derived string is evaluated against the input text by a linear combination of four metrics, called centrality, variation, coverage and cohesiveness. The first two metrics quantify matches of dictionary entries to the head of the phrase, and the mean inverse distance between dictionary and text phrases. The latter two metrics measure the extent and sparsity of matches between the textual and dictionary strings. The candidate matches are then ordered according to mapping strength, and the highest-rank candidate is assigned as the final match. We used MetaMap's Strict Model to filter matches in order to achieve the highest level of accuracy [57].
The UMLS (Unified Medical Language System) Metathesaurus is a rich terminological resource for the biomedical domain [63], [64]. All concepts in the UMLS Metathesaurus are categorized into 135 semantic types (or categories). In this work we focused on the semantic type of “Disease or Syndrome”. This is why the counts of concepts and relations in Table 3 are much less than the total number of concepts and relations from each of the four ontologies in Table 1.
We used the Stanford POS tagger [65], [66] to parse the news and novels corpora comparable to MetaMap's parsing of medical texts. After parsing, we processed the inflectional and morphological variations of each word. For the medical corpus, we retrieved the base form of a word by querying the UMLS Specialist Lexicon based on its appearance in the text (e.g., singular or plural for a noun, different tenses for a verb). For the news and novels corpora, we converted all words to their base word form (e.g., translating nouns from plural to singular and verbs from past and future to present tense) with a rich set of morphological rules. Then we used these base word forms, in addition to their part of speech, to indicate word context for the calculations below. We also used these base forms to match against thesaurus entries.
In this section, we define several metrics for mapping an ontology to a corpus, arranging the metrics by increasing complexity. The simpler metrics do not distinguish between multiple predicate types in an ontology, summarizing all relations between the same pair of concepts, i and j, with a single association probability, pij. More general versions of our metrics account for multiple relation types that occur in more complex ontologies, but these involve numerous additional parameters that require estimation from real data and therefore are more challenging to implement. For this reason, we count relations represented in a test ontology X in two separate ways. |ℜX | is the number of ordered pairs of concepts with at least one relation defined between them in ontology X, while |RX| is the total number of all relations in the ontology. For predicate-poor ontologies such as thesauri, these two ways of counting relations are equivalent. In predicate-rich ontologies with more than one relation between the same pair of concepts, |RX|>|ℜX |.
Suppose an ontology has N concepts and each concept i has relations with other Mi concepts (each denoted as concept j where j = 1, 2, …, Mi). We practically infer the probability pij that concept i is associated with concept j through simple concept co-occurrence in text. Namely, we estimate:(1)
where nij is the number of times concept i co-occurs with concept j in the same unit of text, such as a sentence or a paragraph (the medical abstract in our implementation). Note that when concept i is unobserved in the corpus, we encounter a singularity (zero divided by zero) when applying equation 1 directly and pij violates the basic property of probability by not summing to 1. For this study we pragmatically postulate that if concept i is not observed in the corpus, then the value of pij is set to 0. Datasets S1, S2, and S3 contain complete sets of non-zero estimates of synonym substitution probabilities for our three reference corpora.
The advantage of setting pij to 0 when i is unobserved is that the ontology will be punished for concepts and relations unobserved in the corpus. One could alternately make pij behave as a probability under all conditions (for all values of nij) and still punish the ontology by making pij very small for all unobserved i in the following manner: (2)
where parameter α and are small positive constants (0≤α β 1). This would require us to further add a pseudo-concept , that relates to every concept i with the following probability:(3)
such that is close to 1 when i is not observed and every pij is approximately 0.
One can imagine the use of more advanced natural language processing techniques than co-occurrence to assess the precise semantic relation in text, but we use the probability estimate from equation 1 in our preliminary evaluation of four medical ontologies against our corpus of clinical abstracts.
Consider further an arbitrary ontology that has multiple distinct relations defined for the same pair of concepts. In such a case, we could supplement pij with an additional set of parameters, πk|ij. These new parameters reflect the relative frequency (importance) of textual mentions of the kth relation between concepts i and j, where
In the case of thesauri, in which the primary relation is synonymy, we are able to assess pij more precisely than with medical ontologies. An English thesaurus has N headwords and each headword (denoted as wi where i = 1, 2, …, N) has a list of M synonyms (denoted as wi,j where j = 1, 2, …, Mi). We compute the probability of substituting word wi with its synonym wi,j through probabilistic conditioning on all contexts observed in a corpus in the following way.(4)
where is a shorthand for “sum over all possible contexts of headword ”.
Equation (4) is closely related to distributional similarity metrics explored by computational linguists, e.g. [67]. This notion, that words occurring in the same contexts tend to have similar meanings is called the Distributional Hypothesis and was introduced by Zellig Harris [68], then popularized by Firth—“a word is characterized by the company it keeps” [69]. Some researchers prefer to induce word relationships like synonymy and antonymy from co-occurrence rather than substitution in order to capture lexical as well as semantic similarity [70], [71]. In our analysis, however, we do not induce synonymy, but rather begin with established synonyms from a published Thesaurus. We then simply calculate their substitution frequencies based on shared context.
In our practical implementation, we defined the context of word wi within a sentence as a list of k words immediately preceding and following it, enriched with positional and part-of-speech (POS) information:
To increase the number of comparable four-token contexts for synonyms in our relatively small corpus, we only considered nouns, verbs, adjectives and adverbs in our analysis of context, disregarding tokens with other part-of-speech tags. That is, given a word wi in the text, we select the nouns, verbs, adjectives and adverbs around it within window size 2k = 4 (two before and two after wi), providing a four-word context for all words except those at a sentence boundary. Because many contexts constructed in this way are unique or very rare, we generalize them by ignoring word order and binning words that appear uniquely in the corpus into part-of-speech pseudo-words (e.g., rare-noun, rare-verb, rare-adjective, and rare-adverb). Equation 4 suffers the same limitation as equation 1 for headwords i that do not occur in corpus. One could extend it in the same manner as equation 1 by adding the pseudo-concept such thatcollects the vast majority of the probability mass for unobserved headwords.
In information retrieval (IR), the goal is to identify documents from a large collection most relevant to a user's query. If the subset of relevant documents is known, we can calculate the quality of an information retrieval method with the metrics precision, recall, and the F-measure (harmonic mean of precision and recall). (5)(6)(7)
True positives (tp), false positives (fp), false negatives (fn) and true negatives (tn) are defined by the cross-tabulation between relevance and retrieval: True positives comprise documents that are both relevant to the query and retrieved by the method; false positives are documents retrieved but irrelevant; false negatives are relevant but not retrieved; and true negatives are irrelevant documents not retrieved.
More measures, such as accuracy and fallout, are introduced and computed in Text S1.
For a given reference corpus T, we define the complete ontology , which incorporates all concepts and all relations encountered in corpus T. We also use the corpus to derive a frequency for each concept i in CT, the set of all concepts in T, and concept association probabilityfor each relation in RT, the set of all relations in T. In the special case of a thesaurus, we understand this probability to be the probability of appropriate substitutability, or “substitution probability” for short. It should be noted that our ability to estimate fi depends on mapping concepts from ontology to text. This is why we spent so much time and energy working with thesauri to facilitate the detection of concept synonyms in text. should be normalized in such a way that (N is the total number of concepts in corpus T) and, by definition,is normalized so that for any concept, ci, involved in at least one relationship. In our implementation, we approximate the complete ontology for our medical corpus with all “Disease or Syndrome” concepts in MetaMap, which includes the union of our four medical ontologies in addition to more than a hundred additional terminologies, such as the UK Clinical Terms, Logical Observation Identifiers Names and Codes (LOINC) that identifies medical laboratory observations, RxNorm that provides normalized names for clinical drugs, and the Online Mendelian Inheritance in Man (OMIM) database that catalogues diseases with a known genetic component. The complete ontology only retains those concepts and relations that appear in the corpus. For our thesauri, we approximated the complete ontology with the union of compared thesauri, excluding concepts and relations not found in the corpus.
Consider that we are trying to evaluate arbitrary ontology X with respect to reference corpus T. We define CX and RX as sets of concepts and relations within X, and | CX | and | RX | the cardinalities of those sets. To evaluate X with respect to T, we identify sets CX(true-positives—tp), CX(false negatives—fn), RX(tp), and RX(fn) such that CX(tp) = CX ∩ CT, RX(tp) = RX ∩ RT, CX(fn) = CT — CX(tp), and RX(fn) = RT — RX(tp), where “—” represents set difference.
Then we derive the first ontology evaluation measure—Breadth—to capture the theoretical coverage of an ontology's concepts:(8)
We derive a corollary version of breadth to capture the theoretical coverage of an ontology's concept and relations:(9)
where pij equals 0 if there is no relation between them in X. Both Breadth metrics are defined on the interval [0,1].
By modifying these measures to account for the number of concepts and relations, we develop measures of Depth to capture theoretical parsimony:(10)(11)
where |ℜX| is the number of ordered pairs of concepts with at least one relation defined between them in ontology X. This normalization thus ignores the number of different relations that X may catalog between concepts i and j.
We can also compare an arbitrary ontology X with the fittest ontology of the same size O(X) by including the most representative CX concepts and RX relations from corpus T that maximize depth. In practice, to compute the fittest ontology of fixed size, we have to perform a numerical optimization over a set of concepts and relations where the size of the ontology being optimized is kept fixed, but the concepts and relations taken from the fittest ontology are added or removed to improve the breadth and depth of the optimized ontology. An estimate of the depth of the fittest ontology of fixed size, DepthO(X)(T), allows us to define and compute a Loss measure.(12)
The above measure can be called the Loss of Depth or Depth Loss. In a similar way we can compute an ontology's Loss of Breadth. (In practice, our estimates of the fittest ontology of fixed size were constrained only by the total number of relations in the corresponding test ontology, so that the Depth Loss in Table 2 was computed using equation (19) in Text S1.)
Note that unlike Breadth, Depth is not naturally defined on the interval [0,1], but will rather tend to result in small positive numbers. Therefore, we define normalized versions of Depth and Depth Loss in the following way.(13)(14)
If we consider an arbitrary ontology with multiple types of relations between concepts i and j, we can further extend Breadth2 and Depth2 measures in the following way:(15)(16)
Note that this definition of Depth3 and Breadth3 involves three levels of ontology evaluation: parameter fi captures usage of the ith concept in the corpus; parameter pij reflects the relative importance of all relations between concepts i and j with respect to all relations involving concept i in the corpus; and parameter πk|ij measures the relative prevalence of the kth relation between concepts i and j in the corpus.
Precise implementation of this task would require capturing mentions of every concept i – relation k – concept j triplet in the text using natural language processing tools. The parameter estimates would then be computed by normalizing counts of captured relations and concepts in an appropriate way.
If, on average, there is only one type of relation per pair of concepts, use of metric Depth3 and Breadth3 would be overkill. For computational simplicity, we use only the first- and the second-level Breadth and Depth in our practical implementation.
|
10.1371/journal.pbio.1002381 | Gβ Regulates Coupling between Actin Oscillators for Cell Polarity and Directional Migration | For directional movement, eukaryotic cells depend on the proper organization of their actin cytoskeleton. This engine of motility is made up of highly dynamic nonequilibrium actin structures such as flashes, oscillations, and traveling waves. In Dictyostelium, oscillatory actin foci interact with signals such as Ras and phosphatidylinositol 3,4,5-trisphosphate (PIP3) to form protrusions. However, how signaling cues tame actin dynamics to produce a pseudopod and guide cellular motility is a critical open question in eukaryotic chemotaxis. Here, we demonstrate that the strength of coupling between individual actin oscillators controls cell polarization and directional movement. We implement an inducible sequestration system to inactivate the heterotrimeric G protein subunit Gβ and find that this acute perturbation triggers persistent, high-amplitude cortical oscillations of F-actin. Actin oscillators that are normally weakly coupled to one another in wild-type cells become strongly synchronized following acute inactivation of Gβ. This global coupling impairs sensing of internal cues during spontaneous polarization and sensing of external cues during directional motility. A simple mathematical model of coupled actin oscillators reveals the importance of appropriate coupling strength for chemotaxis: moderate coupling can increase sensitivity to noisy inputs. Taken together, our data suggest that Gβ regulates the strength of coupling between actin oscillators for efficient polarity and directional migration. As these observations are only possible following acute inhibition of Gβ and are masked by slow compensation in genetic knockouts, our work also shows that acute loss-of-function approaches can complement and extend the reach of classical genetics in Dictyostelium and likely other systems as well.
| The actin cytoskeleton of motile cells is comprised of highly dynamic structures. Recently, small oscillating actin foci have been discovered around the periphery of Dictyostelium cells. These oscillators are thought to enable pseudopod formation, but how their dynamics are regulated for this is unknown. Here, we demonstrate that the strength of coupling between individual actin oscillators controls cell polarization and directional movement. Actin oscillators are weakly coupled to one another in wild-type cells, but they become strongly synchronized after acute inactivation of the signaling protein Gβ. This global coupling impairs sensing of internal cues during spontaneous polarization and sensing of external cues during directional motility. Supported by a mathematical model, our data suggest that wild-type cells are tuned to an optimal coupling strength for patterning by upstream cues. These observations are only possible following acute inhibition of Gβ, which highlights the value of revisiting classical mutants with acute loss-of-function perturbations.
| For cells to move, their cytoskeletal structures become spatially organized by internal polarity signals [1–3] as well as external chemoattractant [4–6]. How such signaling cues tame actin dynamics to produce a pseudopod and guide cellular motility remains a key question in eukaryotic chemotaxis.
By now, several key regulators of the actin cytoskeleton have been identified: in most cells, nucleation promoting factors (NPFs) of the Wiskott-Aldrich Syndrome Protein (WASP) and SCAR/WAVE family stimulate actin nucleation through the Arp2/3 complex and are essential for regulating polarity and motility for cells ranging from Dictyostelium [6,7] to metazoans [8–10]. NPFs themselves are regulated by self-association on the plasma membrane [1,11] and actin polymerization-based autoinhibition [1,12,13]; the actin polymer that they generate facilitates the removal of these NPFs from the plasma membrane. These positive and negative feedback interactions of the NPFs [1,14] and other actin regulators give rise to a range of highly dynamic, free-roaming, nonequilibrium actin structures such as flashes and traveling waves [1,2,5,6,15–21], but how the actin machinery is coaxed to form these very different activity patterns is not well understood.
Particularly striking displays of NPF and actin dynamics are actin oscillations, which can be observed in many cell types and contexts [1,2,5,22,23]. Biological oscillations are typically generated through a combination of (1) fast positive feedback, which amplifies small signals into an all-or-none output; and (2) delayed inhibition, which turns the output off and resets the system for the next pulse. By spatially coupling oscillators, spreading or synchronization over long distances can be achieved [24–26].
Recently, small oscillating SCAR/WAVE foci have been discovered at the periphery of Dictyostelium cells [2]. These foci may constitute the basic cytoskeletal units from which pseudopods are formed. In the absence of signaling cues, these oscillators are present but lead to only small undulations of the cell boundary. In response to upstream signals, however, full-blown protrusions emerge [2,27–31], likely from the coordination of these foci. Some intracellular signals (such as Ras and phosphatidylinositol 3,4,5-trisphosphate [PIP3]) have been identified that affect this transition, but whether other signals link receptor activation with the SCAR/WAVE foci, and, more generally, which properties of the foci are modulated to enable large-scale coordination, are not known.
Here, we find that the heterotrimeric G-protein subunit Gβ sets the coupling range of an actin-based activator—inhibitor system. Specifically, acute sequestration of Gβ leads to strong global synchronization of normally weakly coupled cytoskeletal oscillators, and these effects are independent of known upstream regulators of these oscillators, such as Ras and PIP3. We show that this extended range of spatial coupling is detrimental for cell polarity, cell motility, and directional migration. To guide our intuition for how coupling between oscillators could affect the cell’s ability to sense directional cues, we developed a simple mathematical model that represents its minimal features. Simulations show that the ability to sense a noisy input signal is facilitated by an intermediate strength of oscillator coupling, allowing different membrane regions to share information about the stimulus. We propose that in wild-type cells, Gβ sets the coupling strength of actin oscillators to an appropriate level to sense directional upstream cues.
Strong loss-of-function phenotypes in cell motility are rare [6,32–38]. One reason may be that genetic perturbations are slow to act and may give cells time to compensate for gene loss [39–42]. Redundantly controlled processes like actin rearrangements during motility may be particularly susceptible to such compensation. To overcome this limitation, we developed a system that enables fast loss-of-function perturbations to cell signaling events involved in Dictyostelium cell motility. Here, we focus on its application to Gβ.
Heterotrimeric G-proteins consist of one α, β, and γ subunit and link receptor-mediated signals to directed migration and polarization in eukaryotic cells ranging from yeast to neutrophils to Dictyostelium [43–46]. Both intra- and extracellular signals can regulate the cytoskeleton, yet while knockout of the sole Gβ protein in Dictyostelium completely blocks chemotaxis, basal cytoskeletal dynamics and other directional responses such as shear-flow-induced motility and electrotaxis are still present, although somewhat reduced [2,3,44,47,48].
Gβ requires plasma membrane localization in order to signal; thus, removal from the plasma membrane should prevent it from activating downstream effectors. As Gβ is continually exchanged between membrane and cytoplasm with a half-life of 5 s [49], it should be possible to trap it by association with an internal anchor. We built a Gβ sequestration system using a chemical dimerization approach whereby the association of two protein domains (FKBP and FRB) is induced by the small molecule rapamycin [50–54]. Starting with Gβ-null cells [44], we expressed an FRB—Gβ fusion protein and an endoplasmic reticulum (ER)-localized FKBP (FKBP-calnexinA [55]). Thus, addition of rapamycin should drive Gβ relocalization to the ER and suppress its signaling function, effectively rendering cells Gβ-null in an acute fashion (Fig 1A).
To test for rapamycin-induced sequestration, we measured the extent of ER-localized Gβ in single cells over time following rapamycin addition. We computed the correlation between each cell’s fluorescence intensity in the ER anchor and Gβ channels to assess co-localization. FRB-RFP-Gβ was rapidly sequestered from the plasma membrane and increasingly co-localized in large clusters with FKBP-YFP-calnexinA (S1 Movie). Sequestration is fast: half-maximal correlation occurred 5.6 min after addition of the highest dose (5 μM) of rapamycin that was tolerated by cells (Fig 1B and 1C, S1 Data). Sequestration kinetics were similar for both 5 μM rapamycin and 1 μM rapamycin. Therefore, unless indicated otherwise, we used the lower concentration for subsequent experiments.
Gβ-null cells fail to transmit many signals triggered by G-protein-coupled receptors (GPCRs) [44,56–59], and we should be able to recapitulate these defects with our sequestration approach. We thus assayed whether relocalization of Gβ to the ER inhibits transmission of signals from GPCRs to downstream effectors. Stimulating wild-type cells with chemoattractant (cAMP) triggers transient responses, including phosphorylation of PKBR1, and this response is abolished in Gβ-null cells [33,56]. We found that introducing our FRB-Gβ construct in Gβ-null cells rescued the PKBR1 response. Acute sequestration of FRB-Gβ to the ER anchor blocked PKBR1 phosphorylation, but only when all three components of our system—the ER anchor, FRB-Gβ, and rapamycin—are present (Fig 1D). Unfortunately experiments using the inducible sequestration system in developed cells were often problematic: Tagged Gβ and anchor components were frequently degraded during starvation and, likely as a consequence, cells failed to complete their developmental cycle. However, this problem was not observed in vegetative cells, in which the sequestration components remained intact. Gβ-dependent, chemoattractant-stimulated responses in vegetative cells, such as Ras activity and PIP3 production [57,59,60], could also be blocked by Gβ-sequestration (S1 Fig). Taken together, these results demonstrate that in the absence of rapamycin, our inducible sequestration system sustains key Gβ-dependent signaling events. In the presence of rapamycin, Gβ is sequestered from its site of action, thereby blocking receptor-based signaling. In this respect, sequestration of Gβ recapitulates Gβ-null cells.
To probe for phenotypes that may only be apparent after rapid loss of Gβ, we turned to directional motility assays. We measured the behavior of Gβ-sequestered cells presented with two different directional cues—an attractive chemical (folate) or electric fields—and compared their responses with wild-type and Gβ-null cells. While chemotaxis is strictly dependent on Gβ, electrotaxis, the directed migration of Dictyostelium cells in response to electric fields, is not. While Gβ-null cells cannot move up a chemical gradient, they can move down electrical potential [44,47].
We took advantage of the heterogeneity in expression of components in Gβ-sequestered cells to internally control experiments. We can distinguish behavior of cells that, in the presence of rapamycin, are functionally wild-type (expressing RFP-FRB- Gβ, but no CFP-FKBP-anchor), Gβ-null (with no detectable RFP-FRB- Gβ expressed), or Gβ-sequestered (expressing both RFP-FRB- Gβ and CFP-FKBP-anchor). For chemotaxis, we further compared these populations to true wild-type and true Gβ-null cells.
We find that just as unsequestered cells resemble wild-type cells, Gβ-sequestered cells behave similarly to Gβ-null cells in chemical gradients. In the presence of Gβ, cells move directionally, while in the absence of functional Gβ (through sequestration or knockout), directionality is lost (Fig 2A). Furthermore, Gβ-sequestered cells (0.4 +/- 0.1 μm/min; n = 31; +/- SEM) as well as true Gβ-null cells (0.4 +/- 0.05 μm/min; n = 98; +/- SEM) move at a reduced speed compared to unsequestered (1.1 +/- 0.2 μm/min; n = 30; +/- SEM) or true wild-type (2.5 +/- 0.15 μm/min; n = 97; +/- SEM) cells.
In contrast, in electrical fields, the behavior of Gβ-sequestered cells differs from the Gβ knockout. Compared to wild-type and Gβ-null cells, Gβ-sequestered cells show a significant decrease in their directionality during electrotaxis (Fig 2B and S2 Movie). Furthermore, the speed of translocation in Gβ-sequestered cells (2.1 +/- 0.22 μm/min, mean +/- SEM; n = 34) was reduced compared to wild-type (3.8 +/- 0.23 μm/min, mean +/- SEM; n = 34; Student’s two tailed t test: p < 10-6) and Gβ-null cells (2.9 +/- 0.23 μm/min, mean +/- SEM; n = 33; Student’s two tailed t test: p < 0.006).
Closer examination of Gβ-sequestered cells by confocal microscopy revealed a striking change in the organization of the actin cytoskeleton. While wild-type cells have fairly stable levels of cortical and cytoplasmic actin, sequestration of Gβ induces striking oscillations of LimE-GFP, a reporter for dynamic F-actin (Fig 3A and 3B) [61]. Periodic loss of cytoplasmic LimE-GFP intensity is accompanied by a corresponding accumulation of F-actin around the entire periphery of the cell (S2 and S3 Figs). The cytoskeletal oscillations induced by Gβ sequestration are present in the majority of cells and have well-defined characteristics. By automatically tracking cells over time and measuring their cytoplasmic LimE-GFP intensity, we identified oscillating cells from the characteristic peak induced in their Fourier spectrum (S4 Fig). After rapamycin addition, the fraction of oscillating cells rises from 6% to 52%, but only when the ER anchor is co-expressed (Fig 3C and S1 Data). The period of oscillation (measured as the peak frequency of the Fourier-transformed signal) is tightly controlled across all oscillating Gβ-sequestered cells (12.9 +/- 3.2 s, n = 83) (S4 Fig). We also observed a second F-actin phenotype upon acute loss of Gβ. In ~10% of cells, waves of F-actin polymerization travel around the cell perimeter with a similar period as the whole field oscillations, taking 10–20 s for a full cycle (S5 Fig and S3 Movie).
Two lines of evidence confirm that acute Gβ loss of function through sequestration is required to initiate this actin oscillation phenotype. First, oscillations are not observed when the ER is forced into proximity of the plasma membrane, arguing against an ER-specific recruitment phenotype (S3 Fig). Most importantly, when Gβ is overexpressed and sequestered in wild-type cells (which harbor endogenous Gβ that cannot be recruited), no actin oscillations are induced (Fig 3D and S1 Data).
Individual cells transition abruptly into the oscillatory mode. Oscillations become apparent as soon as rapamycin-induced sequestration of Gβ can be observed (S6 Fig and S3 Movie) and can continue for days (see later; Fig 4C and S1 Data). By treating cells with both rapamycin (the FKBP-FRB heterodimerizer) and a competitive inhibitor of heterodimerization (the small molecule FK506, an FKBP-FKBP homodimerizer), we titrated Gβ levels over the full dynamic range of the sequestration system (S7 Fig). As the amount of sequestered Gβ is increased, the properties of the oscillating state such as its period and amplitude did not change (S8 Fig). The oscillations have characteristics of an all-or-none behavior: only the percentage of oscillating cells increased (Fig 4A and 4B, S1 Data).
These phenotypes—whole-cell oscillations and traveling waves of actin polymerization—are reminiscent of previously observed actin-based activator—inhibitor systems [1,2,5,6,16–20]. However, the oscillations we observe here are triggered, persistent, and have an unusually large spatial range and high amplitude. This suggests that acute loss of Gβ pushes the cytoskeleton into an unusual state.
Our observation that cortical F-actin oscillations follow acute sequestration of Gβ raises a key question: why did previous Gβ-null analyses fail to uncover this striking cytoskeletal phenotype? Consistent with published work [2,3], we find that very few Gβ-null cells display LimE-GFP oscillations cells (Fig 4C and S1 Data). We reasoned that if cells compensate for the loss of Gβ function over time, the phenotype induced by acute sequestration of Gβ should approach the Gβ-null phenotype after sufficient time has passed. Consistent with this hypothesis, the fraction of oscillating cells decreases over days of continuous Gβ sequestration and eventually approaches the small fraction seen in Gβ nulls (Fig 4C and S1 Data). Similar compensatory phenomena have been previously observed in other Dictyostelium signaling contexts. For example, the effect of LY294002, a PI3K inhibitor, on Dictyostelium cell migration fades during prolonged treatment [2], likely due to compensation by redundant signaling pathways [35]. In another case, the actin nucleator WASP relocalizes to the leading edge and compensates for SCAR/WAVE function when SCAR/WAVE is deleted [6]. Our findings suggest that a compensatory mechanism is also at work here: the globally oscillating state is suppressed in Gβ-null cells.
Our results highlight the value of using acute inhibition to uncover protein function. We have used rapamycin-induced Gβ sequestration to interrogate loss-of-function phenotypes along two “axes” (Fig 4D). By titrating the amount of sequestration while retaining its fast timescale (axis 1), it is possible to interrogate how a phenotype emerges, distinguishing between an all-or-none or gradual transition. Conversely, varying the timescale of perturbation (axis 2) reveals whether phenomena such as cellular compensation can mask an acutely induced phenotype. Applied to Gβ sequestration, we find that a new phenotype—a globally oscillating F-actin cytoskeleton—can be uncovered at points in this “phenotypic space” that are not accessible to standard genetic perturbations.
Multiple oscillating actin foci localize around the cell periphery on the basal surface of chemotactic cells. These foci often originate from previously aborted pseudopods that remain attached to the substrate. Internal and external signaling inputs are thought to entrain these foci, but how their dynamics are controlled for this to happen remains unknown (e.g., oscillation dynamics are unchanged in Ras, PI3K, and Gβ nulls) [2]. The large-scale cortical actin oscillations we observe here are similar in period to the previously described oscillating foci (13 +/- 3 s versus 9 +/- 2 s, respectively), suggesting that these two forms of cytoskeletal dynamics may be closely related. Thus, we tested whether our acute sequestration of Gβ would reveal signaling control over these oscillatory actin foci.
To analyze individual actin foci, we collected confocal movies imaged in the plane where cells contact the coverslip. We developed a computational approach to comprehensively track and quantify the dynamics of actin foci by automatically identifying each cell’s periphery, subdividing it into ten degree sectors (thereby generating 36 tracked regions per cell), and measuring the mean intensity in each sector over time (Fig 5A). Consistent with previous results [2], we found large-amplitude oscillations in LimE-GFP intensity in some sectors (Fig 5A, right panel) but not others, with a mean period of approximately 10 s (S9 Fig). Regulators of F-actin formation localize to the same structures and oscillate as well: the peak of actin-nucleating SCAR/WAVE complex member HSPC-300 precedes that of LimE by about 2 s; Arp2 and the F-actin binding domain (ABD) of ABP120 peak at about the same time as LimE; and the peak of Coronin, a regulator of actin disassembly [2,62], lags behind LimE by more than 2 s (Fig 5B, S10 Fig, and S1 Data). These data suggest that focal LimE oscillations report cycles of polymerization and disassembly of F-actin.
We next addressed how the dynamics of actin foci compare between wild-type (Gβ-unsequestered) cells and Gβ-sequestered cells that exhibit whole-field oscillation. In both cases, individual sectors oscillate. However, the mean LimE intensity across all sectors in Gβ unsequestered cells does not show a marked oscillatory behavior (Fig 5C), whereas the mean intensity of sectors in Gβ-sequestered cells clearly oscillates (Fig 5D). Thus, the whole-field oscillations we observe upon Gβ-sequestration in the middle plane of cells (Fig 3A) are also reflected in the behavior of membrane-plane actin foci.
What properties of these individual oscillators change as cells transition to whole-field oscillation? We reasoned that changes in the amplitude, period, or the synchronization in phase between individual oscillating sectors could be responsible. We developed an automated approach using the Hilbert transform [63,64], which has been used extensively to analyze neuronal activity [65,66], to quantify the amplitude, period, and phase of individual oscillators over time (S11 Fig). Using this algorithm, we extracted the oscillation phase (i.e., whether currently at a peak or trough) as well as the instantaneous period (i.e., how fast the phase is changing) at each timepoint. Strikingly, only the phase synchrony differs in Gβ-sequestered cells (Fig 5E, S12 Fig and S1 Data). Yet although synchrony increases, it is not perfect: individual sectors can fall in and out of phase with the group over time (S13 Fig). Taken together, our data suggest that global oscillations in Gβ-sequestered cells are caused by increasing synchronization among preexisting membrane oscillators.
Downstream of Gβ, three signaling pathways, defined by PI3K, TORC2, and PLA2, are known to instruct actin-based motility in Dictyostelium (Fig 6D). Ras activity can feed into both PI3K and TORC2, and downstream, Rac activation is thought to connect these signaling modules to the actin cytoskeleton [31,33,38,56]. Enhanced activity of these pathways leads to wider, more stable zones of actin polymerization compared to the isolated oscillating foci.
We investigated whether Gβ uses any of these signaling pathways to regulate spatial coupling of actin foci. First, we analyzed the dynamics of Ras activity, PIP3 levels, and Rac activity in single cells. Gβ sequestration neither induced oscillations nor caused any other apparent changes to these signaling currencies on a timescale of minutes (Fig 6A). Second, we perturbed the activities of members of these pathways in wild-type cells to determine whether global LimE oscillations would emerge. Neither inducing Rac activity (Tet-On: GFP-Rac1A[V12]), blocking all three pathways (using a pharmacological cocktail: BEL|LY294002|pp242), nor raising the levels of intracellular Ca2+ (a messenger commonly oscillating in other systems [22,67]) led to global oscillations of F-actin (Fig 6B and S1 Data). Third, we interfered with these pathways in Gβ-sequestered oscillatory cells to determine whether their activity was required for synchrony. Acute inhibition of all three pathways caused only a very small decrease in the number of oscillating cells, while unbalancing Ca2+ levels did not inhibit global oscillations at all (Fig 6C and S1 Data). We conclude that Gβ’s control over the coupling range of actin oscillators likely involves a different, currently unidentified mediator.
How can hypercoupling between cytoskeletal oscillators lead to a defect in directed cell migration? The coupling state among the oscillators might be an important parameter for upstream cues to polarize the cytoskeleton—a prerequisite for cell motility. To investigate this question, we tracked individual Gβ-sequestration cells over time, simultaneously monitoring cytosolic actin dynamics and cell migration in both the presence and absence of rapamycin.
For this analysis, we returned to confocal imaging in the midplane of the cell. Here, polarization events are distinguished by a relatively stable actin patch that coincides with a substantial drop in cytoplasmic LimE-GFP reporter levels (Fig7A and 7B and S14 Fig). In both control and Gβ-sequestered cells, polarized patches are of similar intensity (S15 Fig), and phases of polarity alternate with apolar phases, which can easily be visualized in t-stack kymographs (Fig 7A and 7B; left panels). In this representation, the y-axis represents time, and the lateral surface of the cell is shown for each timepoint along the x-axis.
We found that Gβ-sequestered as well as Gβ unsequestered cells were capable of cycling between polarized and apolar states (S4 and S5 Movies). Consistent with our prior results, acute sequestration of Gβ induced large-amplitude oscillations of F-actin. However, long-term imaging revealed that these oscillations are largely restricted to apolar phases—times when the cell is not undergoing protrusion or migration (Fig 7B and S5 Movie). Thus, phases of polarization appear to be incompatible with whole cell oscillations. While increased coupling in Gβ-sequestered cells did not affect the lifetime of poles once they successfully formed (Fig 7C and S1 Data), Gβ sequestration significantly (p < 10-4, Student’s two-tailed t test) impaired the establishment of new poles (Fig 7D and S1 Data). Consistent with a reduced number of cell polarization events, sequestered cells translocate at a significantly reduced speed (p < 0.003, Student’s two-tailed t test, Fig 7E and S1 Data).
Taken together, our data show that appropriate control of coupling between localized cytoskeletal oscillators is essential for efficient polarization and motility as well as directional sensing. Increasing the strength of coupling—through acute loss of Gβ—synchronizes actin dynamics, which hampers the entrainment of the actin cytoskeleton by both internal polarity cues as well as entrainment by the external cues that are necessary to direct motility (Fig 8).
One of the most remarkable features of chemotaxis is the ability of migrating cells to accurately sense extraordinarily shallow chemical gradients [68]. Previous work has suggested that the signaling network downstream of Gβ plays a crucial role in this input sensing [5,29,69,70]. Here, we have uncovered a separate link between Gβ and the cytoskeleton in tuning coupling between actin oscillators. Might oscillator coupling also play a role in input sensitivity?
We reasoned that oscillator-to-oscillator coupling might represent a means of sharing information between nearby regions of the cell periphery. By comparing noisy receptor—ligand interactions at multiple locations, cells might improve their ability to discriminate signal from noise when choosing a migration direction. To test this hypothesis in a simple context, we built a mathematical model representing input sensing at the cell’s periphery (Fig 9). It should be emphasized that this model is not meant to capture the full complexity of the cell’s gradient sensing and chemotaxis pathways, but rather represents a minimal model to quantitatively interrogate the essential elements of oscillator-to-oscillator coupling and entrainment to an input. Our model incorporates a circular lattice of actin oscillators representing the cell’s cortex. Oscillators are coupled to one another by a term that increases sinusoidally with their difference in phase [71] and can also be coupled to an oscillating input signal using the same mechanism. Although the chemoattractant signals presented to a real cell are unlikely to oscillate in this fashion, the exact mechanism for input coupling is unknown, and our simplifying assumption allowed us to model oscillator-to-input and oscillator-to-oscillator coupling in a single unified framework. Our model includes three parameters that define the coupling between an external input and the nearby membrane (kIN) and the coupling between membrane oscillators (parameters k1 and k2 for input-coupled and non-input-coupled membrane oscillators). We also include a term (σ) to represent noise in input-to-oscillator coupling.
Our model reproduced well-known features of coupled oscillator systems. Increasing oscillator-to-oscillator coupling showed an abrupt transition to global synchrony, consistent with prior work modeling the synchronization of weakly coupled oscillators as a phase transition (S16 Fig) [68,69]. This is analogous to the effect observed after Gβ sequestration, in which the transition to global oscillation appears to be all-or-none in individual cells (Fig 4 and S1 Data).
To test how coupled oscillators are affected by features of the input signal, we set out to determine how oscillator-to-oscillator coupling affected sensing of weak inputs (low values of kIN) or noisy inputs (high values of σ). We found that increasing coupling could not improve sensing of weak noise-free inputs but rather led to spontaneous synchronization as coupling strength is increased (S21 Fig). In contrast, oscillator-to-oscillator coupling markedly improved sensing of noisy inputs (Fig 9). For simulations with little or no coupling, the effect of noise was dominant, and membrane oscillators were unable to accurately couple to inputs (Fig 9; k1 = 0.1). Conversely, for very strong coupling, oscillators became synchronized to one another so strongly that they were completely input-insensitive (Fig 9; k1 = 3.5) [25,71]. Between these two extremes, our model revealed an optimum of input sensitivity at an intermediate coupling strength (Fig 9; k1 = 2.5).
If weak oscillator-to-oscillator coupling was indeed beneficial for input sensing, one would expect wild-type cells to exhibit some coupling between oscillating foci. Indeed, we find experimentally that in wild-type cells the relative phases of oscillators are not random but loosely correlated (Fig 5E, asynchrony; phase distribution width Θ50 < 90 [deg.]; S12 Fig and S1 Data). Thus, we propose that upstream signaling cues optimally entrain the cytoskeleton when the coupling strength between its dynamic units is of intermediate strength.
A dynamic actin cytoskeleton drives eukaryotic cell migration. Waves, flashes, patches, and oscillatory actin foci have been observed in Dictyostelium [2,5,6,15–19], neutrophils [1,4], and other mammalian cells [17,20,72,73]. Underlying these phenomena are nonlinear reaction processes that exhibit a range of behaviors including excitability and oscillations [1,2,26,29,70,74]. These cytoskeletal dynamics are shaped further by upstream cues such as internal polarity signals [1–3] and external chemoattractant [4,5]. Here, we show that signaling also directly regulates the strength of coupling between local cytoskeletal processes. Acute loss of Gβ leads to strong synchronization of actin oscillators, which has detrimental consequences for cell polarity, motility, and directionality.
Electrotaxis revealed this new role for Gβ in directed cell migration. However, we expect the link between Gβ and cytoskeletal dynamics to be essential for interpreting other cues as well. During chemotaxis, when the activity of the Gαβγ heterotrimer is proportional to the amount of the chemical signal the cell experiences [75], fine control of the magnitude and intracellular distribution [76] of oscillator coupling may be possible.
In future work, it will be important to learn more about how Gβ exerts this control. Common signaling pathways involving PI3K, TORC2, and PLA2 appear to be not essential. Similarly, perturbing levels of Ca2+, a messenger known to oscillate in many systems [22,67], including chemotaxing Physarum polycephalum cells [77], shows no obvious effect on coupling between actin oscillators. P. polycephalum, however, is a beautiful, conceptual precedent for the idea that cell movement may be governed by the coupling between independent oscillators: in this organism, periodic streams of small pieces of cytoplasm can become entrained to each other, which, through further modulation by attractants or repellants, supports directional movement [78].
Recent evidence in Dictyostelium shows that Gβ interacts with Elmo, which suggests a possible direct link to the cytoskeleton bypassing the other signaling pathways [79]. We observed oscillation of HSPC-300, a member of the actin-nucleating SCAR/WAVE complex. This may be the most upstream oscillator, with F-actin reporters and disassembly factors (e.g., Coronin) following its dynamics. In this case, SCAR/WAVE’s relevant regulators will need to be identified [80]. Mechanistically, how could loss of Gβ increase the strength of coupling? Based on the mechanisms through which oscillators are coupled in other systems, possible explanations include (1) increasing the density of oscillators at the periphery while keeping the coupling range of each constant [81], and/or (2) directly increasing the range of a diffusible or mechanical signal that is generated by the oscillators [26]. Our experimental data support the first hypothesis. In strongly coupled Gβ-sequestered cells, a larger fraction of sectors contain actin oscillators compared to Gβ-unsequestered cells or Gβ-null cells (S17A Fig). Moreover, upon Gβ sequestration, the number of membrane sectors that contain an actin oscillator increases, while the amplitude of the oscillators remains constant (S17B Fig). Additionally, we find that during cell polarization, oscillators largely disappear from the sides and back of the cell (S18 Fig). Taken together, our data suggest that the number or density of oscillators is regulated, and this may be used as a mechanism to control coupling strength. Additional mechanisms could affect the firing threshold or the refractory period of the oscillators.
Our simple mathematical model helps to guide intuition on why coupling between oscillators could be advantageous for polarity and directional movement. For both cases, the signals that need to be interpreted can be noisy, and in these scenarios moderate coupling between oscillators can provide an advantage—input-coupled oscillators can “share” information to filter noise and better entrain to an input signal. Our results are consistent with recent predictions in bacterial chemotaxis, in which an optimal membrane distribution of receptors balances sensitivity to spatially correlated external noise and spatially uncorrelated intrinsic noise (which can be filtered out by a similar mechanism of local information sharing) [82].
Our work highlights limitations in classical genetic approaches. Genetic nulls are the most common means of assaying gene function in Dictyostelium. However, many genetic mutants give no or mild phenotypes, and they often require combined hits in multiple signaling pathways to significantly inhibit chemotaxis [6,32–36]. In theory, two mechanisms can account for this: a selection on the population levels can favor a subset of cells (potentially carrying suppressor mutations) that best cope with the genetic change. Alternatively, intrinsic redundancy with parallel pathways or slow compensation via negative feedback can obscure the true role of a gene in cell behavior [42]. Such compensation enables robust function and is a widely employed characteristic of adaptive/homeostatic systems. For example, pharmacological inhibition of synapses transiently inhibits signal transmission, but homeostatic mechanisms restore function within minutes [39–41]. The motor of bacteria is another example. It compensates for persistent changes in the level of internal signaling components to maintain the robustness of chemotaxis [41].
Which mechanism is at play in our case? Both Gβ-null knockout and wild-type cells lack excessive coupling of actin oscillators, albeit likely for different reasons. In wild-type cells, Gβ suppresses the coupling, while Gβ-null knockout cells have, over time, arrived at a Gβ independent steady state that does not support oscillations. Transformation of Gβ-null knockout cells with the sequesterable Gβ construct restores wild-type physiology, which can become transiently unbalanced upon acute Gβ sequestration. This imbalance remains for days—long enough for us to observe its effect on oscillator coupling—but eventually the steady state of Gβ-null knockout cells is assumed again, potentially due to compensation from parallel pathways. We favor this possibility over genetic suppression based on the speed with which oscillations disappear again after induction.
As a consequence of compensation, different modes of gene inactivation can result in strikingly different phenotypes. In zebrafish, gene knockdowns can produce strong phenotypes that are masked by compensation in genetic knockouts [83]. Our data suggest that additional phenotypes appear when proteins become inactivated even more rapidly. Gβ-knockout and knockdown cells have been extensively studied in Dictyostelium and other systems [43,44,84]. Although defects have been reported for a wide range of chemoattractant-stimulated responses, including directed migration [44], these cells display normal basal polarity and actin dynamics [2,3]. Acute sequestration was essential to uncover the role of Gβ in tuning cytoskeletal dynamics and initiating cell polarity. In this light, our work suggests that much can be learned by revisiting classical mutants with acute perturbation approaches, and not only in instances in which a loss-of-function mutation is lethal.
Dictyostelium strains were grown at 22°C in HL5 medium (ForMedium) in Nunclon tissue culture dishes or in suspension in flasks shaken at 180 rpm. Cells were routinely used from nonaxenic cultures. In this case, cells were grown in association with Klebsiella aerogenes (K.a.) on SM agar plates and used for assays when bacteria began to get cleared [85]. Growth under these conditions gave the strongest responses to stimulation with folate, so this condition was used for most subsequent rapamycin-mediated sequestration experiments. However, sequestration of Gβ also induced oscillations in F-actin when cells were grown in HL-5 instead. For imaging experiments, a scrap of cells was seeded in 200 μl HL5 in a Lab-Tek II 8 well chamber (Nunc), allowed to settle, and washed one to two times in KK2 (16.5 mM KH2PO4, 3.9 mM K2HPO4, 2 mM MgSO4) immediately before the assay. Rapamycin (SIGMA) was freshly prepared at 2 μM in KK2 and added 1:1 in sequestration experiments. To render cells responsive to cAMP (Fig 1C), aggregation-competent amoebae were prepared by resuspending washed cells at 2 x 107 cells/ml in KK2, starving them for 1 h while shaking at 180 rpm, followed by pulsing the cells with 70–90 nM cAMP (final conc.) for another 4 h. Before stimulation with 1 μM cAMP, cells were basalated (shaking at 180 rpm in the presence of 5 mM caffeine for 20 min) with or without 5 μM rapamycin, washed in ice-cold KK2, and kept on ice until stimulation. For analysis by western blotting, samples were resolved on 4%–15% SDS-PAGE gels. After electrophoresis, proteins were transferred to PVDF membrane and probed with anti—phospho-PKC (pan) antibody from Cell Signaling Technology (190D10), which was used to detect the activation loop (T309) phosphorylation of PKBR1, and anti—pan-Ras antibody from EMD (Ab-3). Sequestration of Gβ was confirmed by fluorescence microscopy.
PhdA-GFP, LimEΔcoil-GFP, LimEΔcoil-RFP, GFP-Arp2, Hspc300-GFP, ABD-GFP, GBD(PAK)-YFP, Coronin-GFP, and YFP-RBD(PI3K1) have been described previously [2,5,35,86]. Standard methods of molecular biology, including reagents from Quiagen and Zyppy Plasmid Miniprep Kits from Zymo Research, were used to generate the following constructs: SRC-YFP-FRB (pHO34) was assembled in pDXA-YFP by subcloning FRB (XhoI/XbaI) from pOW578 with a synthetic sequence (HindIII/Nsi1) encoding the myristoylation tag from SRC. cAR1-RFP-FRB (pHO39) was assembled in pDXA-YFP by replacing YFP with a fragment containing cAR1-RFP (HindIII/XhoI) and adding amplified FRB (XhoI/XbaI). CalnexinA-CFP-FKBP (pHO232) was assembled in multiple steps. CalnexinA was amplified from a published plasmid [55] and inserted into a variant of pDXA-YFP encoding FKBP (pHO167) or CFP and FKBP (pHO232). A Gateway-compatible vector derived from pDM448 [87] encoding FRB-RFP was generated (pHO436), into which Gβ was inserted with an LR reaction to build FRB-RFP-Gβ (pHO536). A tetracycline-inducible variant of GFP-Rac1AV12 (pHO578) was built by enzymatic assembly (Gibson) in pDM369 [87].
To generate stable cell lines, cells were transformed by electroporation (Genepulser Xcell, Bio-Rad) using 10–20 μg DNA per 4x106 cells (100 μl) in 1 mm cuvettes (Bio-Rad). Two consecutive pulses with a 5-s recovery period between were delivered at 750 V, 25 μF, and 50 Ohm. For overexpression, cells were plated in bulk and selected with G418 (10 μg/ml) and/or hygromycin (50 μg/ml) the next day.
The time course of inducible sequestration (Fig 1B) was benchmarked in strain HO543: A Gβ-null strain (LW6) derived from DH1 [44] was used as the base strain into which the sequestration system was engineered. First, pHO536 (FRB-RFP-Gβ) was introduced, and transformants were selected with hygromycin (50 μg/ml) to give HO535. This strain was then transformed with pHO167 (calnexinA-YFP-FKBP) to give HO543 or simultaneously with pHO232 (calnexinA-CFP-FKBP) and LimEΔcoil-GFP to give HO547, with pHO232 and phdA-GFP to give HO548, with pHO232 and pOH250 to give HO549, and with pHO232 and PAK(GBD)-YFP to give HO630. Additional anchors, such as a NLS or the transmembrane domain of Miro, were tested, but yielded poor depletion of Gβ. Transformants were selected with G418 (10 μg/ml). When appropriate for comparison, parent strains DH1 expressing LimEΔcoil-GFP (HO618) or LW6 expressing LimEΔcoil-RFP (HO595) were analyzed. The following strains were used to control for the effect of rapamycin mediated recruitment: DH1 expressing LimEΔcoil-GFP, pHO232 and pHO39 (HO620; G418 resistant), Ax2 (Kay lab) expressing LimEΔcoil-RFP, pHO232 and pHO34 (HO621; G418 resistant) and Ax2 (Kay lab) expressing LimEΔcoil-GFP, pHO232 and pHO536 (HO626; G418 and hygromycin resistant). For dual color oscillation experiments (Fig 5B), Ax2 (Kay lab) cells expressing LimEΔcoil-RFP together with GFP-Arp2 (HO632), Hspc300-GFP (HO634), ABD-GFP (HO638), or Coronin-GFP were analyzed.
A spinning disc Nikon Eclipse Ti fitted with a spinning disc head, 405 nm, 488 nm, and a 561 nm laser line and appropriate emission filters were used to record CFP, RFP, and GFP (or YFP) double- or triple-labeled cells at room temperature. Images were routinely recorded using a 60x (1.45 NA) objective, a Clara Interline CCD camera (Andor Technologies), and NIS Elements software. After analysis, when necessary for presentation, contrast was adjusted uniformly using ImageJ or Photoshop, and to image sets of some experiments a uniform Gaussian Blur was applied. To quantify oscillations, a single two- or three-channel image was taken to assess Gβ sequestration, followed by a 2-min movie (1 frame/second) to record behavior in the reporter channel at the lowest laser intensity necessary for reasonable signal-to-noise. Longer imaging periods (10 min) and/or adjustment of the focal plane close to the coverslip were used when necessary (e.g., to record individual oscillating foci or alternating polar and apolar states).
For Fig 5, Ax2 cells expressing LimE-RFP were analyzed for 2 min (1 frame/second) immediately before and for 2 min (within 5 min) after applying perturbations. For Gβ sequestration, only oscillating cells (strain HO547) were considered. Ca2+ and ionomycin were used at 10 mM and 10 μM, respectively. For triple drug inhibition, Bromoenol lactone (BEL 5 μM) was washed out after 5 min of treatment, after which acute application of LY294002 (50 μM) together with pp242 (40 μM) followed. BEL and LY294002 have been demonstrated as effective inhibitors of PLA2 and PI3K in Dictyostelium before [38]; pp242 is an inhibitor of TOR kinase and inhibits TORC2-mediated phosphorylation events in Dictyostelium (S19 Fig). Expression of tet-on GFP-Rac1A(V12), was induced overnight with 100 μg/ml doxycycline. The effect on oscillating, Gβ-sequestered cells was additionally tested by treatment with U73122 (5 μM), EGTA (10 mM), and Ca2+ (10 mM).
The electric fields were applied as previously described for vegetative Dictyostelium cells [88] by using μ-Slides (Ibidi). These tissue-culture-treated slides with small cross-sectional area provide high resistance to current flow and minimized Joule heating during experiments. To eliminate toxic products from the electrodes that might be harmful to cells, agar salt bridges made with 1% agar gel in Steinberg’s salt solution were used to connect silver/silver chloride electrodes in beakers of Steinberg’s salt solution to pools of excess developing buffer (5 mM Na2HPO4, 5 mM KH2PO4, 1 mM CaCl2, and 2 mM MgCl2, pH 6.5) [89] at either side of the chamber slide. EF strength is empirically chosen (~10V/cm) based on our previous experience [90] and measured by a voltmeter before and after each experiment. Fields of HO547 cells were chosen based on the presence of Gβ and anchor expressing cells, which were distinguished by fluorescence imaging (see Microscopy section for details). High-definition DIC movies (1 frame/30 s) were recorded at room temperature for at least 30 min after the electric field was switched on. To quantify directionality and speed, time-lapse images were imported into ImageJ (http://rsbweb.nih.gov/ij). Tracks were marked by using the MtrackJ tool and plotted by using the Chemotaxis tool described [91]. All experiments were repeated and produced similar results. Data are combined and presented as means +/- SEM (standard error). To compare group differences, unpaired, two-tailed Student’s t test was used. A p-value of less than 0.05 is considered significant.
HO543, DH1, or LW6 (Gβ null) cells were grown in HL5 medium containing 20 μg/ml G418 and 50 μg/ml hygromycin. Two days before the experiment, 2x105 cells were mixed with an overnight culture of K.a. in 250 μl streptomycin-free HL-5 medium and plated on an SM agar plate. On the day of the experiment, cells were washed off the SM plate with DB buffer, washed once, and resuspended in DB at 2x107 cells/ml. Suitable amount of cells were transferred to LabTek II chambered coverglass (Nalge Nunc) containing DB with 5 μM rapamycin and 0.05% DMSO. For folic acid chemotaxis, Femtotips microcapillary pipettes (Eppendorf) filled with 1 mM folic acid were used. Microscopy for this set of experiments was carried out with a Nikon Eclipse TiE microscope illuminated by an Ar laser (YFP) and a diode laser (RFP). Time-lapse images in bright field, YFP, and RFP channels were acquired by a Photometrics Evolve EMCCD camera controlled by Nikon NIS-Elements. Tracks of cell migration were analyzed in ImageJ to obtain directedness and speed of cells.
For all other analyses, cells were identified, tracked, and processed to extract various properties (e.g., cytoplasmic fluorescence, membrane fluorescence, extent of polarization, angle of polarization) using custom code written in Matlab. First, initial locations for each cell were provided by hand-drawn masks such that each mask contains a single cell at the first timepoint. At each subsequent timepoint, each cell was tracked by extracting a 100x100 pixel box centered at that cell’s prior location in the LimE-GFP fluorescent channel. To identify the cell within this box region, interior pixels were separated from background intensity using a fixed intensity threshold, followed by binary erosion with a single-pixel structuring element (to remove isolated noncell pixels) and a hole-filling operation (to fill all pixels within the cell). The largest connected component within this image was assumed to be the cell.
For each cell and at each timepoint, we extracted the following features:
To identify which cells in a population were oscillating and characterize the timescale of oscillation, we turned to a Fourier approach (for the analyses of Figs 3 and 4). We found that the cytoplasmic LimE-GFP levels undergo strong, regular periodic fluctuations. From each cytoplasmic intensity timecourse, we subtracted a 30 s moving average to center cytoplasmic fluctuations on a mean value of zero (eliminating intensity fluctuations during cell movement or photobleaching) and computed the discrete Fourier transform of this mean-centered signal. Cells were then marked as “oscillating” if any sampling frequency between 0.05 and 0.2 Hz contained at least 10% of the cytoplasmic signal’s total power (see S4 Fig for oscillating and nonoscillating representative cells). These frequencies correspond to periods ranging from 5 to 20 s, which covered the range of frequencies we observed in a preliminary analysis across more than 50 oscillating cells. Each cell’s oscillation frequency was then taken to be the sampling frequency at which the power was maximal.
To understand how cortical LimE dynamics relate to those of other cytoskeletal factors, we sought to correlate LimE-RFP with other reporters (GFP fusions to HSPC300, Coronin, the ABD actin binding domain of ABP120, and Arp2). To identify cells expressing both LimE and a second reporter, we thresholded cells using both GFP and RFP fluorescence. The cell’s cortex was identified as a 5-pixel-wide shell of this thresholded image for each cell. To compute the intensity of cytoskeletal foci around the cell’s cortex, we then subdivided the cortex into 36 equal-angle segments (sweeping out 10 degrees each) and measured the fluorescence intensity in both the GFP and RFP channels.
We then sought to compare the temporal dynamics of GFP and RFP in each spatial region from each cell. To do so, we calculated the cross-correlation between these two channels. For uncorrelated cytoskeletal factors (e.g., myosin, paxillin), we found that dynamics in GFP and RFP were uncorrelated, leading to a low-magnitude, flat cross-correlation. For correlated cytoskeletal factors (e.g., HSPC300, Coronin, Arp2, and the actin binding domain ABD), the cross-correlation peaked at the characteristic delay time between LimE and that particular cytoskeletal factor. We estimated this delay time by fitting a Gaussian distribution to the cross-correlation to identify the location of this peak—the resulting delay times are shown in Fig 5B.
From the centroid and center of mass measurements described above, the direction and extent of polarization was determined by computing the vector between the center of mass (c→) and centroid (n→).
The magnitude of p→ describes the extent of polarization, while its direction reflects the pole’s orientation.
We were also interested in identifying periods of time in which cells exhibit long-term, stable polarization (for the analyses of Fig 7). By inspecting many cell trajectories, we found that stable polarization was associated with a consistent direction of polarity—cells would retain a pole with a similar directional orientation, and changes in direction were associated with the formation of a new pole. Conversely, during unpolarized phases, fluctuations of actin around the membrane would lead to frequent changes in the direction of p→ (S14 Fig; lower panels). Thus, we implemented a greedy search algorithm to find continuous periods of time when the angle of polarization was contained in a 1-radian window and lasted at least 25 s, and measured the number and duration of these polar regions for each cell (S14 Fig shows two representative cells).
To assess the synchrony of oscillation between different membrane regions of a cell, we set out to measure each region’s oscillation phase at each timepoint. The phase of oscillation describes the current position of an oscillating signal on a sinusoidal curve (i.e., the rising or falling edge), and periodically rises from 0 to 2π. Thus, by comparing the phases between different regions of the membrane, we could assess whether they were oscillating in synchrony, with the phase rising and falling together, or whether at a single timepoint different membrane regions were at different points in their oscillating trajectories.
The analytic representation of a signal provided by the Hilbert transform is an ideal way to measure instantaneous properties of a signal containing periodic fluctuations such as the oscillation phase. For the time-varying LimE-GFP intensity in the nth membrane sector xn (t), the analytic signal
x˜n(t)=xn(t)+i xn(t)*1πt
is a complex-valued function from which instantaneous properties of the signal’s oscillation can b e calculated, such as its instantaneous oscillation phase
φn(t)=∠ x˜n(t)
and frequency
ωn(t)=φ˙n(t).
Phase measurement can be improved by first applying a low-pass filter to avoid noisy fluctuations from being interpreted as oscillation. Thus, we first applied a low-pass filter (an 8th order Butterworth filter with a cutoff of 0.2 Hz) to each membrane trajectory before calculating its Hilbert transform, using custom Matlab code. We found this procedure to yield highly robust measurements of oscillation phase (S11 Fig) in both Gβ-sequestered and Gβ-functional cells. The instantaneous frequencies we measure from this approach are closely centered at ~10 s (S9 Fig) and are strikingly similar to those measured by Fourier analysis of cytoplasmic oscillation (Fig 3).
To assess synchrony between different membrane regions, we measured the breadth of spread in oscillation phase between them, at all timepoints during oscillation. We first computed the “group phase”—the vector sum of all regions’ individual phases, weighted identically.
We assessed synchrony by computing the phase difference between each membrane region and the group phase at each timepoint, and measured how broad this distribution is in oscillating Gβ-sequestered and nonoscillatory Gβ-functional cells (S12 Fig shows histograms of two representative cells).
To characterize the migration of Gβ-sequestered and Gβ-unsequestered cells, we tracked individual cells during 10 min movies, where fluorescent images were acquired once per second. Cells were automatically segmented by thresholding the fluorescent channel, and the centroid of each cell was automatically determined at each timepoint. At least 28 cells were tracked in each condition. From each cell’s centroid data, we calculated the root-mean-squared displacement xrms over time for each cell, choosing 300 distinct 5-min intervals for each cell during the 10 min movie. We fit the data to the simple diffusion model
xrms2=2dDt,
where d = 2 is the dimensionality, D is the diffusion constant, and t is the current time. From this model, we estimated the diffusion coefficient for each cell, and computed the p-value for a difference in diffusion coefficients between Gβ-sequestered and Gβ-unsequestered cells (Fig 7E).
|
10.1371/journal.ppat.1007826 | Vaccinia viral A26 protein is a fusion suppressor of mature virus and triggers membrane fusion through conformational change at low pH | Vaccinia mature virus requires A26 envelope protein to mediate acid-dependent endocytosis into HeLa cells in which we hypothesized that A26 protein functions as an acid-sensitive membrane fusion suppressor. Here, we provide evidence showing that N-terminal domain (aa1-75) of A26 protein is an acid-sensitive region that regulates membrane fusion. Crystal structure of A26 protein revealed that His48 and His53 are in close contact with Lys47, Arg57, His314 and Arg312, suggesting that at low pH these His-cation pairs could initiate conformational changes through protonation of His48 and His53 and subsequent electrostatic repulsion. All the A26 mutant mature viruses that interrupted His-cation pair interactions of His48 and His 53 indeed have lost virion infectivity. Isolation of revertant viruses revealed that second site mutations caused frame shifts and premature termination of A26 protein such that reverent viruses regained cell entry through plasma membrane fusion. Together, we conclude that viral A26 protein functions as an acid-sensitive fusion suppressor during vaccinia mature virus endocytosis.
| Vaccinia virus is a complex large DNA virus with a large number of viral membrane proteins to facilitate cell entry. Although it is well established that vaccinia mature virus uses endocytosis to enter cells, it remains unclear how it triggers membrane fusion in the acidic environment of endosomes. Recently, we hypothesized that A26 protein in vaccinia mature virus functions as an acid-sensitive membrane fusion suppressor, which suggests a novel viral regulation not present in other enveloped viruses. We postulated that conformational changes of A26 protein at low pH result in de-repression of viral fusion complex activity to trigger viral and endosomal membrane fusion. Here, we provide structural, biochemical and biological evidence demonstrating that vaccinia A26 protein does indeed function as an acid-sensitive fusion suppressor protein to regulate vaccinia mature virus membrane fusion during endocytosis. Our data reveal an important and unique “checkpoint” for vaccinia mature virus endocytosis that has not been described for other viruses. Furthermore, by isolating adaptive vaccinia mutants that escaped endocytic blockage, we discovered that mutations within the A26L gene serve as an effective strategy for switching the viral infection route from endocytosis to plasma membrane fusion, expanding viral host range.
| Virus entry represents the initial stage of infection and is a target for developing new antiviral therapeutics. Poxvirus is a family of enveloped DNA viruses with genomes of ~200 kilobases [1]. Vaccinia virus, an orthopoxvirus, is a model system for investigating poxvirus entry into host cells, producing mature (MV) and extracellular virus (EV) [2–4].
Vaccinia MV attaches to cell surface glycosaminoglycans and extracellular matrix laminin [5–10]. It then clusters at lipid rafts, triggering the integrin β1-CD98-PI3K signaling cascade [11, 12] to induce actin-dependent endocytosis that may [13] or may not involve apoptotic mimicry [14–16]. After internalization, vaccinia MV is trafficked in vesicles inside the cells, with subsequent endosomal acidification triggering viral membrane fusion with the vesicular membrane to release viral cores into the cytoplasm [17–20].
How vaccinia virus triggers fusion with host cells remains unclear. Many enveloped viruses contain a viral fusion protein that induces conformational changes at low pH [21–23]. Conformational change of viral fusion proteins exposes a hydrophobic terminal fusion peptide [24, 25] or internal fusion loop [26, 27] that can be inserted into host membranes. Subsequent conformational changes and oligomerization of viral fusion proteins then triggers fusion of viral and host membranes. Vaccinia MV employs a highly conserved eleven-component fusion protein complex to mediate virus fusion with cells [3, 28], but how it functions remains unknown [4].
Vaccinia MV exhibits broad infectivity, acting via endocytosis or plasma membrane fusion [9, 29, 30], depending on vaccinia virus strains and cell types [17, 31]. We previously demonstrated that the WR strain of vaccinia MV uses viral A26 protein for the endocytosis pathway, whereas deletion of A26 protein induced the plasma membrane fusion pathway in HeLa cells [32, 33]. However, loss of A26 protein renders viral MV particles resistant to bafilomycin (BFLA) without loss of fusion activity, suggesting that A26 protein is the target of acid regulation, not the viral fusion complex [32]. The A26 protein binds to A16 and G9 proteins of the viral entry fusion complex at neutral pH and, when purified MV was treated with acidic buffer, the A26-A27 protein complex dissociated from MVs at low pH [33], inspiring our model wherein A26 protein is an acid-sensitive fusion suppressor of MV (Fig 1A). In this model, A26 protein binds to viral fusion complex to suppress MV fusion at neutral pH. However, the acidic pH of endosomes triggers conformational changes in A26 protein, which is subsequently released from the viral fusion protein complex, resulting in viral and vesicular membrane fusion. In the absence of A26 protein, viral fusion protein complex becomes fusion-competent at neutral pH, triggering efficient fusion with the plasma membrane, consistent with electron microscopy (EM) data [32, 33]. Here, we provide genetic, biochemical and structural evidence that A26 protein is a fusion suppressor that regulates vaccinia MV membrane fusion through acid-dependent conformational changes.
We generated two N-terminal deletion constructs of the A26 open-reading frame (ORF), in which we removed amino acids (aa) 1–75 or aa 1–320 of A26 protein (Fig 1B). Each deletion construct was fused in-frame with N-terminal flag sequences and inserted into a non-essential thymidine kinase (tk) locus of the WR-ΔA26 virus, which deleted the A26 ORF (Fig 1B). We did not generate any C-terminal deletions of the A26 ORF because this region is required for A26 protein binding to viral A27 protein and subsequent packaging into MV particles [34, 35]. Recombinant viruses expressing WR-A26(76–500) or WR-A26(321–500) were isolated and plaque-purified. A recombinant vaccinia virus expressing full-length flag-tagged A26 protein, named WR-A26, was also included [36]. HeLa cells were infected with individual virus at a multiplicity of infection (MOI) of 5 plaque-forming units (PFU) per cell and harvested at 24 hours post infection (hpi) to determine MV growth (Fig 1C). Control WR-A26 grew approximately 100-fold at 24 hpi. WR-A26(76–500) exhibited significantly reduced MV yield in the same timeframe, whereas WR-A26(321–500) yield was similar to that of control WR-A26 virus. None of these three viruses exhibited defective virus assembly, with each presenting large amounts of MV in cytoplasm of infected cells at 24 hpi (Fig 1D), consistent with previous results demonstrating that A26 protein is not required for MV assembly [10]. We purified MV particles via CsCl gradient purification and found that WR-A26, WR-A26(76–500) and WR-A26(321–500) presented similar morphologies under EM (enlarged MV images in the insets of Fig 1D). Immunoblot analyses of purified MV also revealed comparable A26 protein levels in MV particles of WR-A26, WR-A26(76–500) and WR-A26(321–500) (Fig 1E). We assessed whether the reduced growth of WR-A26(76–500) virus reflects low MV particle infectivity by counting how many virus particles are required to initiate a single infection event in HeLa cells, i.e., to determine the particle-to-PFU ratio (Table 1). For the control WR-A26 MV particles, this ratio is ~43, whereas for WR-A26(76–500) it is ~88, demonstrating that removal of aa 1–75 of A26 protein significantly reduced MV infectivity. Surprisingly, further deletion of A26 protein, aa 1–320, recovered MV infectivity of WR-A26(321–500) to a ratio of ~33, i.e., not statistically different from control WR-A26 MV infectivity (an outcome we explain in the next section).
Based on our model (Fig 1A), A26 protein contains an acid-sensing or acid-sensitive region responsible for inducing acid-dependent conformational changes, as well as a fusion suppressor region to suppress fusion activity at neutral pH. To establish if these region functions are absent in our WR-A26(76–500) and WR-A26(321–500) protein constructs, we adopted a cell-cell fusion assay to investigate virus-cell membrane fusion (Fig 2A, 2B and 2C). Mock-infected cells did not fuse at either neutral or acidic pH, so GFP- and RFP-expressing cells were well separated. Control WR-A26 virus needs a low pH endocytic environment to initiate fusion, so surface-bound MV did not induce cell-cell fusion at neutral pH. However, brief treatment of these surface-bound MV with low pH buffer created an acidic environment that mimicked endosomes, leading to conformational changes of the A26 protein that activated virus-mediated cell-cell fusion to produce double-fluorescent fused cells. In contrast, WR-ΔA26 MV lacked a fusion suppressor, so cell-cell fusion was triggered by virus infection at both neutral and acidic pH. Interestingly, WR-A26(76–500) did not fuse at either neutral or acidic pH, thus acting like a pH-independent fusion suppressor and indicating that aa 1–75 represents the acid-sensitive region of A26 protein. Finally, WR-A26(321–500) triggered robust membrane fusion at both neutral and acidic pH, i.e., similar to WR-ΔA26, suggesting that the fusion suppressor region is absent in WR-A26(321–500) and that aa 76–320 represents the fusion suppressor region of A26 protein.
Based on the fusion quantification data in Fig 2C, we divided the % fusion at low pH (4.7) by that at neutral pH (7.4) in order to obtain an “acid fusion index” that reflects the acid dependence of each MV construct (Fig 2D). Only endocytic WR-A26 had a high acid fusion index, whereas all other viruses lost their acid dependence with WR-ΔA26 and WR-A26(321–500) fusing well at both pH and WR-A26(76–500) fusing poorly at both pH. Therefore, we conclude that the N-terminal region of aa 1–75 of A26 protein is important for acid-sensing or acid sensitivity and the middle region of aa 76–320 is required for fusion suppression (Fig 2E). This conclusion fits well with the low infectivity of WR-A26(76–500) MV and the normal infectivity of WR-A26(321–500) MV (described in the previous section), since this latter virus exhibited the ability to enter cells via plasma membrane fusion, just like WR-ΔA26 virus.
The above-described analyses prompted us to investigate the N-terminal aa sequences of A26 protein. Since two-dimensional (2D) 1H-15N heteronuclear single quantum coherence (HSQC) serves as a reliable measure of secondary structure and conformational change in solution, we then performed 2D HSQC experiment to examine the N-terminus of A26. To achieve better protein solubility and stability critical for NMR study, we fused A26(aa 1–91) coding region with thioredoxin (TRX), and purified the recombinant fusion protein TRX-A26(1–91). In addition, we also purified wild type TRX control protein and a mutant TRX-fused protein TRX-A26(1–91)H48,53R, in which His48 and His53 were mutated to Arg. We then applied 2D HSQC experiment to TRX-A26 (1–91), TRX-A26 (1–91)H48,53R and TRX respectively (Fig 3B–3D). Notably, the 2D HSQC spectra of all three proteins—TRX, TRX-A26(1–91) and TRX-A26(1–91)H48, H53R—at pH 8 are somewhat similar (blue in Fig 3B–3D), suggesting that the 2D spectral signals are dominated by those of TRX protein. Furthermore, the spectral patterns of control TRX at pH 6 and pH 8 are nearly identical (Fig 3C), demonstrating that TRX is pH-insensitive. However, recombinant TRX-A26(1–91) at pH 6 exhibits a different 2D spectral pattern (red in Fig 3B) from that at pH 8 (blue in Fig 3B). At pH 6, the amide-1H signals gave rise to a narrow dispersion within 8–8.5 ppm, indicating a partially unfolded conformation [37]; in contrast, at pH 8, the amide-1H signals displayed a much wider distribution 7–10 ppm, indicative of a structured conformation. The data thus suggested that, when fused with TRX, the A26(1–91) fragment induced significant conformational changes in the TRX-A26(1–91) fusion protein at low pH. However, recombinant TRX-A26(1–91)H48, H53R fusion protein exhibited similar 2D spectral patterning at pH 6 and pH 8 (Fig 3D), confirming that H48 and H53 are responsible for pH sensitivity of TRX-A26(1–91) in vitro.
To further confirm this conclusion, we then removed the TRX tag from the above-mentioned fusion proteins and performed circular dichroism (CD) spectroscopy to analyze conformational changes of recombinant A26(1–91) protein and A26(1–91)H48, H53R mutant protein at different pH, ranging from 5.1 to 8.5 (Fig 3E and 3F). The CD spectrum of recombinant A26(1–91) protein exhibited α-helical ellipticity at 208 and 222 nm (Fig 3E). The ellipticity decreased as the pH value decreased, suggesting a transition from an α-helix to a random coil. In contrast, the A26(1–91)H48, H53R mutant protein was insensitive to pH alteration from 5.2 to 8.5 (Fig 3F) although the secondary structure of the mutant protein may appear similar to A26(1–91) in α-helix content. Based on these data, we concluded that the N-terminal region (1–91) of A26 protein is sufficient for low pH-dependent conformational changes, and that His48 and His53 are essential for acid sensitivity of A26(1–91) protein in vitro.
To demonstrate that His48 and His53 of A26 protein are indeed involved in acid-dependent membrane fusion of MV in cells, we created a recombinant vaccinia virus (WR-A26-H2R) that expresses flag-tagged A26-H48, H53R double mutant protein (A26-H2R protein) in the infected cells (Fig 4A and 4B). Unlike wild-type A26 protein, A26-H2R mutant protein should fail to undergo conformational changes in response to low environmental pH. We also generated another recombinant virus A26-H3R that contains an extra H92R mutation in flag-tagged A26 protein, in addition to H48R and H53R (Fig 4A and 4B).
Immunoblot analyses revealed comparable levels of WR-A26, A26-H2R and A26-H3R proteins in the infected cells (Fig 4C) and in purified MV particles (Fig 4D). Although WR-A26-H2R and WR-A26-H3R mutant viruses exhibited normal MV assembly in the infected cells (Fig 4E), MV growth was reduced (Fig 4F) and MV particles presented very low infectivity (Table 1), reminiscent of our observations of the WR-A26(76–500) deletion virus.
In cell-cell fusion assays, WR-A26-H2R and WR-A26-H3R mutant viruses did not trigger cell-cell fusion at either neutral or low pH (Fig 4G, quantified in Fig 4H), similar to WR-A26(76–500) virus, suggesting that the A26-H2R and A26-H3R proteins are constitutive fusion suppressors, as evidenced by their low acid fusion indexes (Fig 4I). Taken together, these mutational studies show that H48 and H53 are required for the functioning of the acid-sensitive region of A26 protein during vaccinia MV-mediated membrane fusion. Additional mutation of His92 did not enhance the mutant virus phenotype.
To further understand how His48 and His53 mediate A26 protein conformational change, we endeavored to obtain a crystal structure of A26 protein. We generated various A26 gene constructs for protein expression in E. coli and only A26(1–420) and A26(1–420)-C43C342A were successfully purified. However, we failed to obtain crystals from either proteins, suggesting that A26(1–420) and A26(1–420)-C43C342A still contain some disordered regions. Our limited trypsin digestion identified aa 395–420 as an unstable region so we generated recombinant A26(1–397) protein and purified it through affinity and monoQ ion exchange chromatography before solving its crystal structure (Table 2 and Fig 5). The overall A26(1–397) structure consists of 18 α-helices and 6 β-strands (Fig 5A and 5B), with an N-terminal α-helical domain (NTD; aa 17–228) and a C-terminal β-sheet domain (CTD; aa 229–364). The total solvent accessible surface area (SA) of A261-397 is 14606 Å2, and the buried area between these two domains are 4252.3 Å2 the interface (the contact areas on NTD and CTD are 2387 and 1865.3 Å2, respectively). Additionally, an inter-domain disulfide bond is present between Cys43 and Cys342, consistent with our previous mutational analyses [35]. To address the novel fold of this structure, we used the DALI server (http://ekhidna2.biocenter.helsinki.fi/dali/) to perform a structural homolog analysis. The results showed that the overall structure of A261-397 does not have any significant hit. The most similar protein to A261-397 is the C-terminal MIF4G domain in NOT1 (PDB ID 6H3Z with RMSD above 11.4 Å). Moreover, the A261-397 NTD exhibits only minor similarity to importin protein (PDB ID:3zkv; with RMSD above 6 Å). In comparison, the folding of A261-397 CTD is similar to gamma crystallin S (PDB ID: 1m8u; RMSD = 2.2 Å). However, since sequence identity between the A261-397 CTD and gamma crystallin S is below 15%, it does not suggest a close relationship between these two proteins. Consequently, the A261-397 structure appears to present a novel fold with two distinct domains.
Many viral fusion proteins exhibit pH-dependent conformational changes that are mainly controlled by electrostatic repulsion [39]. Although A26 protein is a fusion suppressor and not a viral fusion protein, its ability to respond to acidic environments suggests that electrostatic replusion may also contribute to its conformational changes at low pH. In general, two classes of paired amino acids are involved in pH-dependent electrostatic repulsions within a protein, i.e., His-cation repulsion at acidic pH and anion-anion (Ani-Ani) repulsion at neutral pH [39]. For His-cation pairs, the histidine residues are usually close (< 7 Å) to other His or basic residues (Arg or Lys). Therefore, we investigated whether any His-cation or Ani-Ani pairs are present in the A26(1–397) structure. As shown in Fig 5C, most His-cation or Ani-Ani pairs are located around helix α2 that hosts His53 (green arrow in Fig 5C), so upon encountering the acidic endosomal pH, charge repulsion produced by the His-cation pair between His53 and Arg57 destablizes the conformation of helix α2. It is worth noting that His53 is also cation-paired with Arg312 and His314, both of which are located in the CTD of A26(1–397), so electrostatic repulsions of His53 at low pH may also destabilize the interactions between helix α2 and the CTD. Furthermore, His48 (green arrow in Fig 5C) is His-cation-paired with Lys47 in helix α2. We observed that an Asp308-Asp310 pair is also adjacent to this region. In another notable obervation, the predicted pKa of the residues that involve in His-cation or Ani-Ani pairs (S1 Table) are usually low (below 5) and most residues with low pKa are found in the helix α2 region, indicating that these residues prefer proton release. Thus, our crystal structure of A26(1–397) protein strongly supports that His-cation pairs involving both His48 and His53 within the N-terminal region most likely contribute to structural alterations by increasing electrostatic repulsions under acidic conditions.
We performed in vitro mutagenesis to express an A26 mutant protein (A26-H2-CAT) that contains K47D, R57D, R312D and H314R mutations to reduce cation-mediated repulsion at low pH via His48 and His53 (Fig 6A). As expected, yield of the recombinant WR-A26-H2-CAT virus (Fig 6B) at 24 hpi was significantly reduced (Fig 6C). Purified WR-A26-H2-CAT MV particles contained mutant A26 protein of the correct size (Fig 6D), but exhibited low infectivity with an increased particle-to-PFU ratio of ~147 (Table 3). Importantly, the WR-A26-H2-CAT mutant virus triggered less cell-cell fusion at both neutral and acidic pH (Fig 6E and 6F) and it presented a low acid fusion index (Fig 6G). Thus, we conclude that electrostatic repulsion induced by the N-terminal-protonated His48 and His53 residues and their surrounding basic aa (K47, R57, R312 and H314) is essential for conformational changes of A26 protein at low pH. Interference with these conformational changes will inhibit subsequent membrane fusion of an endocytic vaccinia virus.
During our experiments on HeLa cells, we noticed that WR-A26(76–500) recombinant virus formed plaques that were slightly smaller than those of control WR-A26 and recombinant WR-A26(321–500) viruses (Fig 7A). Interestingly, WR-A26-H2R, WR-A26-H3R and WR-A26-H2-CAT recombinant virus all formed tiny plaques on HeLa cells. These A26 mutant viruses appeared unstable, generating spontaneous “large plaque” revertants during early virus passaging and propagation (red arrows in Fig 7A). Therefore, we isolated several of the large plaque revertants (Rev) from the WR-A26-H2R, WR-A26-H3R and WR-A26-H2-CAT mutant viruses and analyzed their protein expression in the infected HeLa cells. As shown in Fig 7B, the size of the A26 protein in all these large-plaque revertant viruses was either smaller relative to control or the protein was completely absent in the infected cells. We suspected that this outcome was due to second-site mutations within the A26 ORF so we purified viral genomic DNA from cells infected with the WR-A26-H2R-Rev1, WR-A26-H3R-Rev1 and WR-A26-H2-CAT-Rev1 revertant viruses and subjected it to whole genome sequencing (results summarized in Table 4).
All revertant viruses retained the designed A26 mutations present in the parental WR-A26-H2R, WR-A26-H3R and WR-A26-H2-CAT mutant strains, but they also all contained an extra intragenic deletion in the A26 ORF that resulted in a frame-shift and premature termination of A26 protein translation (Fig 7C). Our sequencing results are consistent with the immunoblots (Fig 7B), although some small A26 fragments were not detected in the lysates, probably due to rapid degradation. WR-A26-H2R-Rev1 and WR-A26-H2-CAT-Rev1 genomes contained no other gene mutations, whereas WR-A26-H3R-Rev1 contains a C-to-A mutation in a pseudogene B3R ORF (a truncated ortholog of camelpox viral gene 176R). The camelpox 176R encodes a schlafen-like protein in virus-infected cells, but a screening of 16 vaccinia viruses revealed no evidence of B3R expression [40]. We conclude that all three revertant viruses host second-site mutations that only affect A26 protein function. Since the A26 fragments in all of these revertant viruses are much shorter and lack the C-terminal A27-interacting region (Fig 7C), they are unlikely to be packaged into revertant MV particles. Accordingly, we anticipated that these three revertant viruses would exhibit a phenotype similar to that of WR-ΔA26 virus. Indeed, the WR-A26-H2R-Rev1, WR-A26-H3R-Rev1 and WR-A26-H2-CAT-Rev1 viruses mediated clear cell-cell fusion under neutral pH, just like WR-ΔA26 (Fig 7D), and presented acid-independent fusion activity (Fig 7E). Therefore, by mutating A26 protein to eliminate His-cation-mediated repulsion at low pH, we created a constitutive suppressor for viral membrane fusion so mutant MV infectivity diminished significantly. However, second-site mutations in the A26 gene resulted in revertant viruses regaining MV infectivity (Tables 1 and 3) and exhibiting normal virus yields (Fig 7F) through plasma membrane fusion. Successful selection of these revertant viruses provides strong evidence that vaccinia MV can switch between the endocytosis and plasma membrane fusion entry pathways, mediated by A26 protein on MV. Most importantly, we have uncovered the structure of the N-terminal region of A26 protein and provide mechanistic insights demonstrating that electrostatic repulsion of His48 and His53 is critical for controlling acid-dependent conformational change of A26 protein prior to virus-mediated endocytic membrane fusion.
Poxviruses are very large and are known to contain multiple proteins of overlapping or redundant functions. Vaccinia virus contains four envelope proteins for cell attachment [5–8, 10], whereas viral membrane fusion requires a separate fusion protein complex of 11 components that specifically performs membrane fusion (reviewed in [3, 4]). Therefore, vaccinia virus has evolved two separate sets of envelope proteins specialized for cell attachment and membrane fusion, respectively, during cell entry.
Many studies have reported that vaccinia virus entry pathways vary depending on virus strains and cell types [17, 31, 41]. How can virus entry pathways be strain-dependent? Using proteomics and genetic complementation analyses, we previously showed that A26 protein determines MV entry pathways in several cell lines [32]. A26+ strains, such as the WR and IHD-J strains, employ endocytosis to enter HeLa cells, whereas A26- strains, such as MVA and Copenhagen, employ a plasma membrane fusion pathway [32, 33]. Viral endocytosis would appear to be an optimal mode of virus entry into cells since no envelope proteins or viral membranes remain on the host cell surface for host B and T cell detection. However, under certain conditions when endocytosis becomes a less optimal route for A26+ vaccinia virus to enter cells, deletion of the A26 ORF results in A26- MV progeny that can infect cells through plasma membrane fusion, thereby broadening the host range. Another advantage of having multiple entry pathways is to avoid innate immune sensing and antiviral signaling activation. We recently infected murine bone marrow-derived macrophages (BMDM) with WR or WR-ΔA26 virus and found that the IFNβ-Stat1 signaling pathway was preferentially induced by endocytic WR virus but not by WR-ΔA26 virus [36]. Consequently, WR-ΔA26 exhibited enhanced virulence in mice compared to WR vaccinia virus [36].
Here in this study, we have used genetic, biochemical and structure analyses to provide strong evidence supporting that vaccinia A26 viral protein functions as an acid-sensitive fusion suppressor of MV particles during virus endocytosis. To demonstrate the critical role of the pH-dependent conformational changes, we purposely generated His48R and His53R mutations (A26-H2R) in the N-terminal domain of A26 protein so that the acid-sensing region is rendered pH-independent. We assume that the N-terminal domain in the A26-H2R mutant may structurally mimic the low pH conformation because of constitutive repulsion forces even in neutral pH environments. However, this scenario does not necessarily mean that the conformation of the fusion suppressor domain also changes during assembly into MV particles. Based on the A26 crystal structure, we generated H2-CAT mutations such that the resulting N-terminal domain structure of the A26-H2-CAT protein also becomes pH-independent but, in this case, His-Cation repulsion was replaced by His-Anion attraction. Therefore, we anticipated that the N-terminal domain of A26-H2-CAT mimics the neutral pH conformation, even in acidic environments. Despite different structural mimics being generated in the N-terminal regions of A26-H2 and A26-H2-CAT, both proteins were acid-insensitive and constitutively suppressed fusion. These outcomes demonstrate that pH-dependent conformational changes via His-Cation repulsion, as opposed to a particular acid-stable N-terminal domain structure per se, are essential for regulating membrane fusion activation. The A26-H2R and A26-H2R-CAT mutations eliminated the acid-dependent response and these mutant proteins retain fusion-suppressing functions. Taken together, our deletion and mutagenesis data support that His48 and His53 in the N-terminal domain are protonated at low pH, creating electrostatic repulsion with surrounding residues (K47, R57, R312 and H314) that results in conformational changes. Our model is also consistent with the data from the A26(76–500) deletion protein, which behaves as a constitutive fusion suppressor upon deletion of the acid-sensing domain. Finally, although we do not have the crystal structure of the C-terminal region of A26 protein, we have generated other C-terminal His-to-Arg mutant viruses, such as A26H357R, A26H425, 432 439R, A26H439, 452, 453R and A26H425, 432, 439, 452, 453R. All these A26 C-terminal mutant viruses expressed A26 mutant proteins of correct size, formed plaques of normal size, and grew to high titers, suggesting that, in contrast to His48 and His53, these C-terminal His residues have a limited role in A26-mediated MV entry and fusion regulation.
In our A26 crystal structure obtained at neutral pH, His48 and His53 are located in the helix α2, which is strategically sandwiched between N-terminal helix clusters and C-terminal beta strands (Fig 5C). This implies that the helix α2 is important for maintaining protein structure stability at neutral pH. The A26 protein structure revealed that these C-terminal beta sheets are distinct from the N-terminal helix cluster, with a sole intra-molecular cysteine disulfide bond formed between C43 and C342, suggesting that the C-terminal domain may stabilize the helix-rich N-terminal domain or vice versa. Currently, we do not have an A26 crystal structure under conditions of low pH nor for A26-H2R mutant protein at neutral pH so we do not know how the N-terminal domain alters A26 protein structure under the acidic condition.
To investigate pH-mediated changes of A26 by using the current results, we employed Discovery Studio [42] to produce an A261-397 model at pH 4.7 [A261-397 model (pH 4.7)]. We first compared the surface electrostatic potential between A261-397 and A261-397 (pH 4.7) model (S1 Fig). Our results show that the NTD of A261-397 is highly positively charged, and the CTD of A261-397 is relatively negatively charged. Furthermore, we identified a positively-charged cavity (dotted green box in S1 Fig, panel A) between two domains. The cavity is formed by the α2 helix region, which we propose plays an important role in the pH-dependent regulation of A26.
Next, we endeavored to address the issue of pH-mediated changes in A26. Since the crystal structure of A261-397 was obtained from a neutral pH, we first used Discovery Studio to produce a model of A261-397 at pH 4.7, as described in materials and methods. We used the final conformation for subsequent analysis (e.g. to calculate surface electrostatic potential and solvent accessibility, etc.). We then compared the structures and surface charges of the A261-397 and the A261-397 (pH 4.7) model. In the A261-397 (pH 4.7) model, the low pH enriches the positive charges and reduces the negative charges on the protein surface (S1 Fig, panel B). We also observed a significant difference in the low pH model in terms of the region comprising the α1 and α2 helices, both of which presented a loosely coiled structure. This outcome may be related to electrostatic repulsions caused by the enrichment of positive charges in this region. Although our low pH model seems to support our hypothesis, it will be necessary to resolve the actual structure of A261-397 at low pH for verification.
The solvent accessibility (SA) of A261-397 and A261-397 (pH 4.7) model were calculated using Discovery Studio [42]. However, the SA difference between overall protein of A261-397 and A261-397 model (pH 4.7) is not much (14606 and 14357 Å2, respectively, representing a difference of 1.7%). We also established the SA of each residue for both structures (S2 Table). Using a threshold for individual residues of a 15% difference in SA between structures, we predicted the SA of 27 residues is reduced in the A261-397 (pH 4.7) model, whereas it was increased for another 17 residues. Notably, although the SA of most residues that involve His–Cation and Ani-Ani pairs were not greatly altered, we observed pronounced SA changes adjacent to the α2 helix (S2 Fig), suggesting that this region may undergo a conformational change at low pH. Again, this analysis was based on a model of A261-397 at low pH, it is necessary to resolve the actual structure of A26 at low pH to precisely elucidate the pH-dependent changes in A26.
It is worth noting that though A26 protein only exists on MV and not on EV, it has been hypothesized that A26 protein in cells can negatively regulate MV egress to Golgi to form EV [43, 44]. Since all of our designed acid-insensitive A26 mutant viruses exhibit small plaque sizes, we rationalized that conformational changes of A26 protein may also control MV to EV egress at late phase in infected cells. Subsequent isolation of revertant viruses from the WR-A26-H2R, WR-A26-H3R and WR-A26-H2-CAT mutant viruses was unexpected. However, these revertant viruses clearly demonstrate how vaccinia mature virus can cope with detrimental mutations of A26 protein and how, through second-site mutations in A26 protein, the revertant mature virus regains host cell entry ability by switching from endocytosis to plasma membrane fusion (Fig 8). Apart from the three revertants reported in Fig 7, we analyzed additional revertants with the large plaque phenotype—Rev2 and 3 from WR-A26-H2R; Rev2, 3 and 4 from WR-A26-H3R; and Rev2, 3, and 4 from WR-A26-H2-CAT (S3 Fig.)—and immunoblots showed that all of these revertants contain a smaller form of A26 protein or no A26 protein at all (S4 Fig). PCR amplification and sequencing of A26L genes from these viral genomes revealed additional second-site mutations within the A26L ORF (S5 Fig.) that resulted in a frame-shift and premature termination of A26 protein (S6 Fig.), consistent with our immunoblot data and supporting our model shown in Fig 8. As expected, all of these revertant viruses infected cells via plasma membrane fusion, and displayed robust cell-cell fusion at both neutral and acidic pH (S7 Fig).
Previous crystallization experiments on viral fusion proteins under different pH have provided strong evidence to support acid-dependent conformational changes of viral fusion proteins [23, 45–51]. At low pH, Type I, II and III fusion proteins exposed an N-terminal fusion peptide or internal fusion loop for target membrane insertion, followed by fusion protein oligomerization, hemifusion and subsequent complete fusion between viral and host membrane [23, 26, 52, 53]. Although A26 protein is not a viral fusion protein, its acid-dependent conformational changes provide a new paradigm of membrane fusion activation, i.e. activation of viral membrane fusion by “de-repression” of a viral fusion suppressor. In the report by Gershon et al., the N-terminal region of A26 was proposed to interact with the N-terminal of G9 and the C-terminal of ATI based on crosslinking analysis [54]. However, the protein structures of G9, A16 and ATI remain unresolved. Without this structural information, it is difficult to evaluate the interfaces between these proteins and A261-397. We speculate that a conformational change at low pH may affect the G9 and ATI interfaces on A26, resulting in disassociation of A26 from these two proteins, but the actual structures of A26 at low pH, as well as G9, A16 and ATI will be necessary to verify this possibility. A recent crystal structure study of the Gn glycoprotein of Rift Valley Fever Virus (RVFV) also revealed that Gn protein shields the hydrophobic fusion loops of the Gc fusion protein to prevent premature fusion of RVFV [55]. Despite a lack of similarity between the RVFV Gn protein and vaccinia A26 protein, conformational changes of viral regulatory proteins may represent a new mechanism to activate viral fusion proteins for membrane fusion.
An African green monkey kidney cell line BSC40 (provided by Dr. Sridhar Pennathur), a human cervical adenocarcinoma cell line HeLa (Obtained from American Type Culture Collection (ATCC CCL-2) and murine L (ATCC CRL-2648) cells expressing GFP or RFP were cultured in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% fetal bovine serum (Invitrogen). The wild-type Western Reserve (WR) strain of vaccinia virus, the A26L deletion virus (WR-ΔA26), and a revertant WR-Flag-A26 virus (WR-A26) in which an N-terminal flag-tagged A26L ORF was reinserted back into the genome of WR-ΔA26 were all described previously [33]. Viruses were propagated in BSC40 cells and purified through a 36% sucrose cushion and a 25–40% sucrose gradient, followed by CsCl gradient centrifugation as previously described [56, 57]. Anti-A27 [5] and anti-D8 [7] antibodies were described previously. Anti-flag monoclonal antibody was purchased from Sigma Inc.
Two A26L N-terminal deletion ORFs, A26(aa 76–500) and A26(aa 321–500), were generated by PCR using the WR-A26L ORF as template. PCR fragments containing N-terminal flag sequences were individually cloned into pMJ601 plasmid so that each flag-A26L deletion ORF was expressed from a synthetic late promoter and flanked by the left and right arm viral tk sequences. In addition, we performed in vitro mutagenesis (QuickChange Lightning site-directed mutagenesis kit; Agilent Tech. Inc.) on the pMJ601-flag-A26L ORF plasmid to generate His-to-Arg mutations at His48 and His53 (A26-H2R), as well as three histidine residues at His48, His53 and His92 (A26-H3R). We also performed in vitro mutagenesis on the pMJ601-flag-A26L ORF plasmid to generate an A26 mutant protein (A26-H2-CAT) that contains K47D, R57D, R312D and H314R mutations to reduce cation-mediated repulsion at low pH via His48 and His53. All mutant A26L plasmids were sequenced to confirm accuracy. To generate N-terminal A26 deletion viruses and the other A26 mutant viruses, CV-1 cells (A kidney cell line of Cercopithecus aethiops purchased from ATCC CCL-70) were infected with WR-ΔA26 at an MOI of 1 PFU/cell, subsequently transfected with each plasmid and cultured for another 3 days. Lysates were subsequently harvested for recombinant virus isolation on BSC40 cells via three rounds of plaque purification in agar containing 150 μg/ml 5-bromo-4-chloro-3-indolyl-β-D-galactopyranoside (X-Gal), as described previously [58].
CsCl purified vaccinia mature virions (1μg) were loaded on SDS-PAGE gels for immunoblot analyses as previously described [32]. Alternatively, BSC40 or HeLa cells were infected with various vaccinia viruses at an MOI of 5 PFU per cell for 1 h at 37°C and harvested at 24 hpi. Cell lysates were separated on SDS-PAGE gels and transferred onto nitrocellulose membranes for immunoblot analysis with anti-flag, anti-A27, and anti-D8 antibodies as previously described [32].
EM analyses of virus-infected cells were performed as previously described but with minor modifications [59]. In brief, BSC40 and HeLa cells were infected with each virus at an MOI of 5 PFU per cell and harvested at 24 hpi. After embedding and sectioning, samples in thin sections were stained with 1% uranyl acetate in H2O and Reynold’s lead citrate solution. The samples were subsequently examined under a Tecnai G2 Spirit TWIN electron microscope (FEI Company, The Netherlands) and photographed using a CCD camera (4*3 model 832, Orius SC1000B).
EM analyses of purified vaccinia MV were performed as previously described [8]. In brief, CsCl-purified vaccinia MV samples were serially diluted, and then half of the samples were stained with 1% Sodium phosphotungstate (PTA) and loaded onto 400-mesh, 10 nm Formvar and 1 nm carbon-coated grids to count MV particle numbers under a Tecnai G2 Spirit TWIN electron microscope (FEI Company, The Netherlands). The other half of the serially-diluted MV samples was used to infect BSC40 cells for plaque assays to determine virus titers. Vaccinia MV infectivity is calculated as the Particle-to-PFU ratio = (Particle number per ml) / (PFU per ml) as previously described [8]. The infectivity assays for each virus were repeated three times and statistical analyses were performed using Student's t test in Prism (version 5) software (GraphPad). Statistical significance is represented as *, P value <0.05; ** <0.01; and *** <0.001.
Cell-cell fusion assays induced by vaccinia MV infections were performed as previously described [32]. In brief, L cells expressing EGFP or RFP were mixed at a 1:1 ratio and seeded in 96-well plates. The next day, cells were pretreated with 40 μg/ml cordycepin (Sigma) for 60 min and subsequently infected with each vaccinia virus at an MOI of 50 PFU per cell in triplicate. Cordycepin was present in the medium throughout the experiments. After infection at 37°C for 30 min, cells were treated with PBS at either pH 7.4 or pH 4.7 at 37°C for 3 min, washed with growth medium, further incubated at 37°C, and then photographed at 2 hpi using a Zeiss Axiovert fluorescence microscope. Five images for each virus were recorded and the % fusion was calculated using the image area of GFP+RFP+ double-fluorescent cells divided by that of single-fluorescent cells.
“Acid Fusion Index” was calculated to represent the acid-dependence of each A26 deletion protein, i.e., the occurrence of A26 protein conformational change. The index for each virus was obtained by dividing the percentage of cell fusion at low pH with that recorded for neutral pH. To quantify fusion activity of each virus as described above the fusion assays were repeated three times and statistical analyses were performed using Student's t test in Prism (version 5) software (GraphPad). Statistical significance is represented as *, P value <0.05, ** <0.01, and *** <0.001.
An NdeI-EcoRI DNA fragment containing a thioredoxin (TRX)-hexahistidine-Xa cutting site was synthesized and cloned into the bacterial expression vector pET21a, resulting in pET21a-TRX. Subsequently, an EcoRI-XhoI fragment containing the A26L ORF encoding aa 1–91 was synthesized and cloned into pET21a-TRX, resulting in pET21a-TRX-A26(1–91) that expresses a bacterial fusion protein containing N-terminal TRX fused with A26 (aa 1–91) (Yao-Hong Biotechnology Inc., Taiwan). Next, in vitro mutagenesis was performed using pET21a-TRX-A26(1–91) plasmid as template to generate the pET-21a-TRX-A26(1–91)H48, 53R mutant plasmid. To generate a control TRX expression plasmid, we inserted a stop codon immediately before the A26(1–91) sequence in pET-21a-TRX-A26(1–91) to generate a control plasmid that only expresses TRX protein (XL Site-Directed Mutagenesis Kit, Agilent Technologies, Santa Clara, CA).
Each plasmid was transformed into BL21(DE3) and recombinant proteins were expressed via 0.2 mM isopropyl 1-thio-β-d-galactopyranoside (IPTG) induction for 4 h, before harvesting for protein purification by nickel column chromatography as suggested by the manufacturer. All of the recombinant A26 proteins used in our NMR analyses contained the TRX fusion tag. To render recombinant TRX fusion proteins suitable for heteronuclear NMR studies, bacterial cultures were incubated at 37°C in M9 medium supplemented with [15N]ammonium chloride (1 g/liter) (Sigma-Aldrich Co., St. Louis, MO) to an absorbance at 600 nm of 0.8, induced for 4 h at 37°C with 0.2 mM IPTG, and harvested for protein purification through nickel-nitrilotriacetic acid affinity column chromatography. The bound recombinant TRX-fusion proteins were eluted with 0.3 M imidazole and dialyzed overnight against 0.1 M MES buffer pH 6.0 at 4°C before use.
For NMR measurement, all 1H-15N HSQC spectra were recorded at pH 6.0 or 8.0 and at 25°C on a Bruker Avance 600 MHz spectrometer equipped with a 5 mm QXI (1H/13C/15N) z-axis gradient probe. For 15N-labeled proteins (0.8–1.0 mM), Shigemi NMR tubes (5 mm outer diameter) were used. The pH values were measured at 25°C with a Suntex TS-100 pH meter. Proteins started to aggregate under acidic conditions (pH<6.0). All spectra were processed using Topspin version 3.2 (Bruker, Karlsruhe, Germany).
For CD measurements, the TRX tag was removed from the fusion A26 proteins using ~2 μg Factor Xa (Sigma Aldrich) in 20 mM Tris·HCl, pH 6.5 with 50 mM NaCl and 1 mM CaCl2 at 4°C. After cleaving, purification was carried out using Ni-NTA resin to separate the cleaved His-TRX tag from the A26 protein samples. The CD spectra were recorded on a Jasco J-815 spectrometer equipped with a water bath for temperature control. All CD spectra were collected at 25°C using a quartz cuvette with a 1 mm path length and a protein concentration of 15.4 μM. The step size was 0.2 nm with a 1.0 nm bandwidth at a scan speed of 50 nm/min. Each spectrum represents the average of three measurements. All spectra were collected in 20 mM potassium phosphate buffer with background buffer correction.
Four A26L DNA fragments coding for (1) full-length A26 protein, (2) residues 1–420 (A26(1–420)), (3) residues 1–420 with Cys43/Cys342 mutations (A26(1–420)C43C342A), and (4) residues 1–110 (A261-110) were each ligated into pET-16 vector (Novagen). Each construct was transformed into Escherichia coli BL21(DE3). After induction with 1 mM IPTG, each recombinant protein was expressed at 16°C for 16 hours. Each soluble A26 protein was purified by immobilized metal-ion chromatography with a Ni-NTA column (GE Healthcare).
Since we failed to crystallize A26(1–420) and A26(1–420)C43C342A, we used a limited trypsin digestion assay to determine the core structure of A26 protein. Briefly, 500 μg purified A26(1–420)C43C342A was incubated in 30 μl reaction buffer (30 mM Tris pH 8.0, 100 mM NaCl, 1 mM dithiothreitol and 5% glycerol) either alone or in the presence of trypsin at a 1:500 ratio (w/w; trypsin:protein). Digestion was carried out at 4°C for 16 hours. The reaction products were analyzed by 12% Bis-tris SDS-PAGE and stained with coomassie blue. The bands containing A26 fragments were excised from the SDS PAGE gel and then subjected to in-gel digestion with trypsin. The digested peptide mixtures were then subjected to a NanoLC−nanoESI-MS/MS analysis. MS data were analyzed using the MASCOT server (http://www.matrixscience.com/search_form_select.html).
Based on the above-described digestion analysis, we decided to use A26(1–397) for further crystallization experiments using the (smt3/Ulp) system provided by Dr. C. D. Lima for recombinant protein expression and purification as previously described [60]. The pET-His10-SUMO-A261-397 DNA construct (codon-optimized in bacteria) was transformed into the E. coli BL21(DE3) strain (Novagen), and the cells were cultured in LB broth containing 50·μg/ml Kanamycin until the optical density at 600 nm (OD600) reached 0.6–0.8 at 37°C. A final concentration of 0.1 mM isopropyl-®-thiogalactopyranoside (IPTG) was added to induce expression and cultured overnight at 17°C for 20 h until the OD600 reached 1.20. Bacterial pellets were harvested by centrifugation at 6,000×g for 30 min at 4°C and then disrupted by sonication in lysis buffer [20 mM Tris pH 8.0, 20 mM imidazole, 0.5 M NaCl, 10% (w/v) glycerol, 1 mM PMSF, 1 mg/ml lysozyme, 0.1 mg/ml DNase I, 1 mM benzamidine (Novagen), and EDTA-free protease inhibitor cocktail (Roche)]. SUMO-A26(1–397) protein was loaded onto a Ni-NTA affinity chromatography column (GE Healthcare), washed first with 40 volumes of binding buffer 1 [20 mM Tris pH 8.0, 20 mM imidazole, 0.5 M NaCl, 10% (w/v) glycerol], then with 40 volumes of binding buffer 2 [20 mM Tris pH 8.0, 100 mM imidazole, 0.5 M NaCl, 10% (w/v) glycerol], before elution with a linear gradient of up to 100% (v/v) elution buffer [20 mM Tris pH 8.0, 0.5 M imidazole, 0.5 M NaCl, 10% (w/v) glycerol]. The eluted SUMO-A26(1–397) protein was dialyzed three times against 7.5 liters of buffer [20 mM Tris pH 8.0, 0.3 M NaCl, 10% (w/v) glycerol] and then subjected to Ubiquitin-like-specific protease 1 (Ulp1) treatment to remove the histidine-tagged SUMO fusion protein. The histidine-tagged SUMO fusion protein was cleaved using Ulp1 at a ratio of 1:500 (w/w; Ulp1:protein) that was later removed with a Ni-NTA affinity chromatography column (GE Healthcare). The untagged A26(1–397) proteins were purified using another HiPrep Q FF 16/10 column (GE Healthcare). These untagged A26(1–397) proteins were purified through a HiPrep Q FF 16/10 column (GE Healthcare), the column was washed with 10 volumes of Q binding buffer [20 mM Tris pH 8.0, 1 mM DTT, 50 mM NaCl], and eluted with a linear gradient of up to 100% (v/v) elution buffer [20 mM Tris pH 8.0, 1 mM DTT, 0.5 M NaCl]. The A26(1–397) proteins were stored in a buffer containing 20 mM Tris pH 8.0, 1 mM DTT, and 100 mM NaCl at 4°C. We used a 12% SDS PAGE gel to confirm the purity of A26(1–397) proteins as being above 99%. Selenomethionine-labeled A26(1–397) protein (SeMet A26(1–397)) was labeled using a SelenoMethionine medium complete kit (Molecular Dimensions) and purified according to the same procedures.
Since A26 shows no sequence homology to any reported protein structure, we produced SeMet A26(1–397) for X-ray analysis. SeMet A26(1–397) was crystallized by the sitting drop method, in which 2 μl of the purified protein (15 mg/ml) was mixed with 2 μl of a reservoir containing 0.2 M sodium acetate trihydrate, 0.1 M Tris pH 8.5, 30% w/v PEG 4000, and equilibrated with 200 μl of the reservoir at 25°C. For X-ray data collection, 15% ethylene glycol was used as a cryoprotectant. A single-wavelength anomalous dispersion (SAD) X-ray diffraction dataset was collected from Taiwan Photon Source (TPS) beamline 05A at the National Synchrotron Radiation Research Center (NSRRC) in Hsinchu, Taiwan. The X-ray data were processed by using HKL2000 [61].
The space group of the SeMet A26(1–397) crystal is P21, with unit cell dimensions of a = 45.38 Å, b = 80.72 Å, c = 53.81 Å and β = 113.8° (Table 2). The initial electron density map of SeMet A26(1–397) was calculated by using the peak dataset collected at wavelength 0.97907 Å and the program Shelix CDE [62]. The program BUCCANEER [63] was then used to produce the initial model. Only one SeMet A26(1–397) monomer was found in each asymmetric unit. We used the programs COOT [64] and Refmac [65] for model refinement. Finally, residues 17–364 of SeMet A26(1–397) were successfully built. In addition, resdiues 382–385 of SeMet A26(1–397) (Thr-Pro-Ile-Pro) were also built as a separate fragment. Data collection and refinement statistics are shown in Table 2. The completeness of 95.2% is actually for the outermost shell and this is now clarified in Table 2. We also analyzed our SeMET A26 structure by MolProbity (http://molprobity.biochem.duke.edu/) [38]. The resulting MolProbity score and Clashscore are 1.10 and 2.68, confirming the good quality of this structure. We used the CCP4 package [66], Chimera program [67], Areaimol [68], PROPKA3 [69] and Discovery studio [42] for structural analyses and to generate figures. The corresponding positions of the regions of constructs used in this study are highlighted in S8 Fig. Discovery studio was also used to build a A261-397 model at pH4.7 [A261-397 model (pH4.7)].
Then, the A26 1-397model (pH4.7) was subjected to molecular dynamic simulations following the Standard Dynamics Cascade protocol. The dynamic simulations were performed for a production time of 1,000 ps using default parameters. The final conformation of A26 1–397 model (pH4.7) was then used in subsequent analysis (e.g. calculate Surface electrostatic potential and solvent accessibility, etc).
WR-A26-H2R, WR-A26-H3R and WR-A26-H2-CAT mutant viruses formed tiny plaques of HeLa cells. Revertant (Rev) viruses displaying a large plaque phenotype were spontaneously derived at an initial rate of ~0.1% during early expansion of WR-A26-H2R, WR-A26-H3R and WR-A26-H2-CAT viruses. We independently isolated three revertant viruses (Rev1-3) from WR-A26-H2R, four revertant viruses (Rev1-4) from WR-A26-H3R, and four revertant viruses (Rev1-4) from WR-A26-H2-CAT, respectively. All of the revertant viruses were subsequently purified to 100% purity. Viral genomic DNA was purified from all the revertant viruses but only WR-A26-H2R-Rev1, WR-A26-H3R-Rev1 and WR-A26-H2-CAT-Rev1 viruses were sent for whole genome sequencing. Viral genomic DNA of all other revertant viruses were used in PCR amplification and sequencing to determine the location of the second-site mutations in A26L ORF.
Vaccinia viral genomic DNA was extracted as previously described [70]. All genomic DNA was quantified by Qubit ds DNA BR assays using a Qubit 3.0 fluorometer (Life Technologies, Carlsbad, US-CA). Genomic DNA (2 μg) was sheared to an average length of 550 bp, end-repaired and A-tailed, and then ligated with indexed adapters for PCR-free library preparation [71] based on the manufacturer’s protocols (Illumina TruSeq DNA PCR-Free Library Preparation kit protocol 15036187 Rev.B) (Illumina Inc., San Diego, CA, USA). The quality and size distribution of the genomic libraries was verified using an Agilent DNA High Sensitivity kit (5067–4626) and Agilent 2100 Bioanalyzer. Viral genomic sequencing was performed using an Illumina Miseq 2x300 cycle or NextSeq500 2x150 cycle paired-end run at the Genomics Core facility of the Institute of Molecular Biology, Academia Sinica.
The sequence data were performed using CLC Genomics Workbench 11.0.1 (Qiagen, Aarhus, Denmark) for raw sequencing trimming, sequence mapping, and variant detection. Raw sequencing reads were trimmed by removing adapter sequences, low-quality sequences (Phred quality score of less than Q20) and sequencing fragments of shorter than 30 nucleotides. Sequencing reads were mapped to the human genome (GRCh38, from ftp.ensembl.org/pub/release-82/fasta/homo_sapiens/dna/) with the following parameters: mismatches cost = 2, insertion cost = 3, deletion cost = 3, minimum fraction length = 0.8, minimum fraction similarity = 0.8. All host genome sequences that met the above parameters were removed, as were duplicate reads, before mapping the remaining paired-end reads to the vaccinia virus WR genome (GenBank NC_006998) [72] with much more stringent parameters (mismatches cost = 2, insertion cost = 3, deletion cost = 3, minimum fraction length = 0.9, minimum fraction similarity = 0.9). We used the Basic Variant Detection tool in CLC Genomics Workbench 11.0.1 to call single nucleotide polymorphisms (SNPs) and insertions/deletions (InDels) with customized parameters to identify mutation positions: (1) minimum frequency of 10% and minimum coverage 10 reads; and (2) minimum quality of SNPs/InDels should be larger than Q25 and the neighborhood quality (upstream/downstream five bases) should be larger than Q20.
We also used the paired-end reads after removing host genome sequences to generate mutant and revertant viral genomes by de novo genome assembly but excluding the terminal repeat sequences of vaccinia virus. Using the alignment program MAFFT version 7[73, 74], we aligned all viral genome sequences with the reference WR strain (GenBank NC_006998) to identify the second-site mutations in revertant viruses. The multiple alignments of viral genomic sequences of WR-A26, WR-A26-H2R-Rev1, WR-A26-H3R-Rev1 and WR-H2-CAT-Rev1 are included in Supplemental S1 Appendix.
Coordinates and structure factors of A26 have been deposited in the Protein Data Bank with 6A9S accession number (PDB ID: 6A9S).
|
10.1371/journal.pntd.0004483 | Burkholderia pseudomallei Colony Morphotypes Show a Synchronized Metabolic Pattern after Acute Infection | Burkholderia pseudomallei is a water and soil bacterium and the causative agent of melioidosis. A characteristic feature of this bacterium is the formation of different colony morphologies which can be isolated from environmental samples as well as from clinical samples, but can also be induced in vitro. Previous studies indicate that morphotypes can differ in a number of characteristics such as resistance to oxidative stress, cellular adhesion and intracellular replication. Yet the metabolic features of B. pseudomallei and its different morphotypes have not been examined in detail so far. Therefore, this study aimed to characterize the exometabolome of B. pseudomallei morphotypes and the impact of acute infection on their metabolic characteristics.
We applied nuclear magnetic resonance spectroscopy (1H-NMR) in a metabolic footprint approach to compare nutrition uptake and metabolite secretion of starvation induced morphotypes of the B. pseudomallei strains K96243 and E8. We observed gluconate production and uptake in all morphotype cultures. Our study also revealed that among all morphotypes amino acids could be classified with regard to their fast and slow consumption. In addition to these shared metabolic features, the morphotypes varied highly in amino acid uptake profiles, secretion of branched chain amino acid metabolites and carbon utilization. After intracellular passage in vitro or murine acute infection in vivo, we observed a switch of the various morphotypes towards a single morphotype and a synchronization of nutrient uptake and metabolite secretion.
To our knowledge, this study provides first insights into the basic metabolism of B. pseudomallei and its colony morphotypes. Furthermore, our data suggest, that acute infection leads to the synchronization of B. pseudomallei colony morphology and metabolism through yet unknown host signals and bacterial mechanisms.
| Melioidosis is a common disease in Northern Australia and East Asia, with regional mortality rates of up to 40%. Clinical manifestations range from soft tissue infections to severe sepsis. It is caused by the Gram negative saprophytic water and soil bacterium Burkholderia pseudomallei, which forms a variety of colony morphologies on solid agar. Various morphotypes appear after the bacterium is exposed to physiological stress conditions or underwent the process of infection, yet the physiological function is unclear. Metabolism is closely linked to virulence in many pathogens, and since metabolic data are not available so far for this bacterium, we monitored the nutrition uptake and metabolite secretion of B. pseudomallei morphotypes. Interestingly, despite typical genes responsible for gluconate production are missing in the B. pseudomallei genome, we observed high amounts of gluconate in the extracellular space. Furthermore, we were able to investigate metabolic differences among the morphotypes and identified synchronization in morphology and metabolism after infection as an adaptation to the host environment.
| Colony morphology variants are described for many bacterial pathogens that are able to induce pneumonia including Pseudomonas aeroginosa [1] Staphylococcus aureus [2] and Burkholderia cepacia complex [3]. It is also a long known phenomenon of the Gram negative water and soil proteobacterium Burkholderia pseudomallei, which is the causative agent of melioidosis [4,5]. This disease occurs predominantly in Northern Australia, Southeast Asia, China and Taiwan but additional cases and environmental isolates of B. pseudomallei have been reported from several regions between latitude 20°N and 20°S [6–8]. Clinical manifestations are highly diverse including soft tissue lesions, abscess formation, sepsis and pneumonia, with the latter being the most frequent clinical presentation of this disease [4,8]. The isolation of isogenic B. pseudomallei morphotypes out of patients indicates a significant role for these morphotypes in adaptation during human melioidosis [9]. In vitro studies have shown that the appearance of various B. pseudomallei morphotypes can be linked to starvation stress, iron limitation, growth temperature and presence of antibiotics [9]. However a clear connection between morphotype formation and metabolism as described for S. aureus or Pseudomonas fluorescence has not been established so far [10–12]. Functional studies indicate that infection relevant parameters like adhesion and intracellular replication differ between the morphotypes [9,13]. Additionally, a higher susceptibility to reactive oxygen species, overexpression of the arginine deaminase system and flagellin was observed [13,14]. However, the question, if morphotype formation affects the bacterial metabolism should be addressed, since in many intracellular pathogens metabolism is closely connected to virulence and depends on host derived nutrients (reviewed in [15,16]). Metabolic adaptation of B. pseudomallei to the host environment includes the expression of metabolic genes for alternative carbon sources and the downregulation of TCA-cycle, glycolysis and oxidative phosphorylation [17,18]. Yet gene expression data only can indicate metabolic activities and to our knowledge actual metabolic investigations have not been carried out on B. pseudomallei so far. We therefore aimed in this study to gain a deeper insight into the diversity of colony morphology variants of B. pseudomallei on a metabolic level. Thus we investigated the uptake of amino acids, glucose and other carbon sources as well as the secretion of metabolites into the extracellular space. Thereby, we found so far unknown shared metabolic characteristics of B. pseudomallei morphotypes and morphotype specific secretion patterns. Our intention was also to address the question whether these metabolic features are affected by the process of infection in a murine macrophage cell culture model and in an acute pneumonia mouse model. Altogether our data indicate that, despite some variation, metabolic principles are shared among B. pseudomallei colony morphology variants and that the acute infection event synchronizes colony morphology as well as metabolism.
All the animal experiments described in the present study were conducted in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All animal studies were conducted under a protocol approved by the Landesamt für Landwirtschaft, Lebensmittelsicherheit und Fischerei Mecklenburg-Vorpommern (LALLF M-V; 7221.3–1.1-020/11). All efforts were made to minimize suffering and ensure the highest ethical and human standards.
All experiments with B. pseudomallei were carried out in biosafety level 3 (BSL3) laboratories. Two different B. pseudomallei strains were used in this study. The sequenced B. pseudomallei strain K96243 was originally isolated in 1996 from a 34-year-old female diabetic patient in Thailand [19] and has been used as a reference strain in many studies [20–22]. The second strain is the yet unsequenced B. pseudomallei strain E8, an environmental isolate from Ubon Ratchathani, Thailand [23,24].
Both strains were cultivated at 37°C and 140 rpm in Tryptone Soja Broth (TSB) to an optical density of 0.8 (OD 650nm) and cells were stored in Luria broth (LB) media with additional 20% glycerol at– 80°C until usage. Additionally, cells were plated on LB agar plates and incubated at 37°C for 4 days, but no different morphotypes could be detected. To generate strain specific morphotypes, different nutritional limitation conditions were used as described previously [7] with the following modifications. In brief, 10 μl of the strain K96243 and strain E8 glycerol stocks were used for the inoculation of i) 3 ml LB media, ii) 3 ml LB media with 5% glycerol, iii) 3 ml Dulbecco’s Modified Eagle’s Medium (DMEM), iv) 3 ml DMEM media with 5% glycerol, v) 3 ml RPMI-1640 (supplemented with 15 mM citrate) media and vi) 3 ml RPMI-1640 (supplemented with 15 mM citrate) media with 5% glycerol. The cells were aerobically cultured at 37°C under static conditions for 24 days. Afterwards, serial dilutions were plated onto Ashdown agar, incubated at 37°C in air for 4 days and the different colony morphotypes were photographically documented and used for the generation of further morphotype stocks and stored at -80°C. To confirm their stability after storage at -80°C, all morphotypes were plated again on Ashdown agar and incubated for 4 days at 37°C in air, but no further morphotype switching was observed. Three colonies of every morphotype were picked and used as biological replicates in further growth experiments.
For in vivo experiments, female 8 to 12-week-old BALB/c mice were purchased from Charles River (Germany). BALB/c mice were housed under specific-pathogen-free conditions. All mice received 100 colony forming units (CFU) of B. pseudomallei strain K96243 colony morphotypes intranasally and were monitored daily after infection. After 72 hours, mice were sacrificed and the lungs were homogenized and plated onto Ashdown agar plates in appropriate dilutions. Occurring morphotypes were photographically documented and their colonies were stocked in LB with 20% glycerol at -80°C until usage. Also their stability after storage was confirmed as described above.
For in vitro experiments, we used the murine macrophage cell line RAW 264.7 purchased from the American Type Culture Collection (ATCC) (Rockville, MD). Briefly, RAW 264.7 cells were seeded in six-well plates (5 x 105 cells/well) and grown for 24 hours in DMEM + 10% fetal calf serum (FCS) before starting the infection experiments. Cells were infected at a multiplicity of infection of ~5 of all six isolated B. pseudomallei morphotypes from strain K96243 for 30 min, washed twice with phosphate buffered saline (PBS), then DMEM medium + 10% FCS containing 250 μg/ml kanamycin (Km) was added for 6 hours and finally the cells were further incubated for 18 h with DMEM medium + 10% FCS and 125 μg/ml Km. After a total of 24 hours, cells were washed three times with PBS before lysis with 1 ml Aqua destillata at 37°C for 15 minutes. Serial dilutions were spread plated onto Ashdown agar, incubated at 37°C in air for 4 days and occurring colony morphotypes were photographically documented and used for the generation of further in vitro morphotype stocks and stored at -80°C. Their stability after storage was confirmed as described above.
All morphotypes (obtained from nutrients starvation, in vivo and in vitro experiments) were spread on Ashdown agar and cultivated at 37°C for 4 days. Then, relevant colonies were scraped from the plates and diluted in 700 μl PBS to an OD 650nm of about 10. These dilutions were used for the inoculation of 50 ml RPMI-1640 media containing 15 mM citrate to an OD650 of 0.05. All samples were cultured at 37°C and with vigorous agitation for 60 hours. During and after the cultivation an aliquot of the culture was plated on Ashdown agar as described above to confirm the colony morphotype. Cultivations of every morphotype and every condition were carried out as triplicates.
2 ml of cell free culture medium were taken at 5.5, 8.25, 11.5, 16.5, 30.5 and 60 hours by sterile filtration and directly frozen until measurement. 1H-NMR analysis was carried out as described previously [25]. In brief, 400 μl of the sample was mixed with 200 μl of a sodium hydrogen phosphate buffer (0.2 M, pH 7.0) to avoid chemical shifts due to pH, which was made up with 50% D2O. The buffer also contained 1 mM trimethylsilyl propanoic acid (TSP) which was used for quantification and also as a reference signal at 0.0 ppm. To obtain NMR spectra a 1D-NOESY pulse sequence was used with 64 FID scans with 600.27 MHz at a temperature of 310 K using a Bruker AVANCE-II 600 NMR spectrometer operated by TOPSPIN 3.1 software (both from Bruker Biospin).
For qualitative and quantitative data analysis we used AMIX (Bruker Biospin, version 3.9.14). We used the AMIX Underground Removal Tool on obtained NMR-spectra to correct the baseline. Thereby we used the following parameters: left border region 20 ppm and right border region -20 ppm and a filter width of 10 Hz. The region of noise, used for final baseline correction was between 5.5 ppm and 5.6 ppm. In some cases noise or unknown signals appeared in regions of integrated metabolites after these were consumed completely during cultivation. If these signals were different from the signals in our database they were regarded as false positive. These false positive signals were replaced with an adequate integral of the noise region mentioned. Absolute quantification was performed as previously described [25]. In brief, a signal of the metabolite, either a complete signal or a proportion, was chosen manually and integrated. The area was further normalized on the area of the internal standard TSP and on the corresponding amount of protons and the sample volume.
Microsoft Excel 2007 was used for all final calculations and for generation of tables. The software PAST was used for the generation of principle component analysis (PCA) [26]. For that matter the calculated concentrations of amino acids were mean centered and autoscaled before being applied to the PCA [27]. Single values from all K96243 morphotypes and ex vivo isolates were grouped by the time point of sampling. For statistical comparison of extracellular metabolite data and growth we performed the two-way ANOVA provided by Prism (version 6.01; GraphPad Software). Multiple comparisons were corrected by applying the Holm-Šidák approach and the level on confidence (alpha) was set to 0.01. Bar-charts and XY-plots were also done using PRISM software. Heatmaps of extracellular amino acids were created using MeV v4.8.1 [28] with the following settings for hierarchical clustering: optimized gene leaf order, euclidean distance metric and average linkage method.
We were able to identify 6 morphotypes of the K96243 strain and 8 morphotypes of the E8 strain by plating on Ashdown agar after 24 days of starvation in 6 different media (Fig 1A(iii)). Both strains were able to form rough and smooth colonies with variations in color and inner and outer shape (Fig 1B). We found very similar morphotypes in both strains (E8 MT01 and K96243 MT03; and E8 MT08 and K96243 MT10) but also unique morphologies (E8 MT07, E8 MT12 and K96243 MT20). We used these 14 morphotypes to elucidate differences in uptake and secretion of diverse nutrients and metabolites (Fig 1A(vi)).
When we cultivated the morphotypes in modified RPMI for 60 h, no lag phase was detected (Fig 2). After 5 hours the midexponential growth phase was reached and the optical densities of morphotypes differed significantly. After 8.25 hours cells entered the transition phase and subsequently after 30 hours the stationary phase. K9 morphotypes reached maximum optical densities between 1.62 (K96243 MT01) and 3.64 (K96243 MT03) whereas E8 derived morphotypes showed a higher variation between 1.08 (E8 MT12) and 3.73 (E8 MT01). After 60 hours of cultivation we observed cell aggregation in some morphotypes cultures and in accordance to that the optical density, especially in the culture of E8 MT07 and K96243 MT20, dropped slightly.
Beside glucose, citrate and myo-inositol are available in the modified RPMI medium and serve as carbon sources. Notably, citrate was present after supplementation in a concentration of 15.65 mM whereas myo-inositol was rather low concentrated at 0.29 mM. All morphotypes showed citrate uptake with some variation over time (Fig 3A/3B). Unlike glucose or amino acids, citrate was not completely depleted after 60 h of cultivation but the uptake was strongly reduced after 30 h for most morphotypes. The decrease in concentration of citrate within the first 30 h of cultivation was measured between 4.91 mM and 9.55 mM and only between 0.65 mM and 3.49 mM during the second 30 hours. A significant uptake of myo-inositol was first measured 16 h or 30 h after inoculation, depending on the morphotype (Fig 3A/3B). K96243 MT01 and K96243 MT10 and E8 MT08 and E8 MT12 showed the fastest myo-inositol uptake. E8 MT11 however showed only very little uptake of myo-inositol.
Based on the decline in concentration, amino acids were classified in two sets (Fig 3C). Set A consists of the amino acids glutamine (gln), glutamate (glu), aspartate (asp), asparagine (asn), serine (ser), glycine (gly) and proline (pro) and showed a decline in concentration below detection limit within the exponential growth phase. In the culture of the morphotypes K9 MT01 and E8 MT08 the uptake of most set A amino acids was completed even before reaching midexponential phase. Set B includes aromatic amino acids (tyrosine (tyr), phenylalanine (phe) and histidine (his)), branched chain amino acids (isoleucine (ile), leucine (leu) and valine (val)), positive charged amino acids (lysine (lys) and arginine (arg)), non-proteinogenic (4-hydroxyproline, 5-oxoproline) and threonine (thr). Amino acids of set B were consumed slower during growth and for some morphotypes almost no uptake was observed until midexponential phase (5 hours). This applies especially to thr, of which more than 96% of the initial amount was still being available in all morphotypes cultures after 5 h. Also only little amounts of arg, tyr, 4-hydroxyproline and 5-oxoproline are taken up in that time. In the late transition phase the diversity between the uptake profiles was relatively high, since some morphotypes took up amino acids of set B more efficiently than others. Interestingly, E8 MT11 showed the fastest uptake of the non-proteinogenic amino acid 5-oxoproline, whereas being rather slow in the uptake of other amino acids of set B. After 30 hours only two morphotype cultures still contained amino acids in measurable amounts (ile; val; leu in the culture of E8 MT11 and arg in the culture of K96243 MT20). However, at the end of cultivation a complete uptake of amino acids was observed for all morphotypes.
The first step in branched chain amino acid (BCAA) degradation is the deamination via IlvE (branched chain amino acid aminotransferase). We found the corresponding α-keto-acids (leu→ 4-methyl-2-oxovalerate; ile→ 3-methyl-2-oxovalerate; val→ 2-oxoisovalerate) to be secreted into the medium by several morphotypes in minor but distinct amounts (Fig 4). Overall, K96243 MT20, E8 MT11 and E8 MT15 secreted the highest amounts. Notably for E8 MT11 no reuptake of these compounds was observed, whereas the other morphotypes consumed the secreted metabolites again. The reuptake for K96243 MT20 and E8 MT15 started after 30 hours of growth, when branched chain amino acids were mostly depleted, yet for E8 MT11 the uptake of branched chain amino acids was still in progress. Therefore the reuptake of α-keto-acids may be connected to the depletion of corresponding amino acids.
Isovalerate, which might be a degradation product of leucine, was solely found in the culture medium of two E8 morphotypes (MT03 and MT15) (S2 Fig). Secretion of isovalerate started in the transition phase and no reuptake for this substance was observed in both cultures.
Several other signals were appearing in the spectroscopic data and point to the secretion of other organic compounds into the medium especially, when cells enter the stationary phase. Unfortunately we were not able to identify most of these metabolites (detailed signal pattern are summarized in S1 File, Table A). However, even if the signals are unknown, they can be used for a discrimination of morphotypes by their secretion pattern (S2 Fig).
To elucidate the effect of infection on morphotype stability and metabolic activity we performed infection experiments in mice (Fig 1iv/2) and murine macrophage cell cultures (Fig 1iv/3) and isolated bacteria out of the lung and the intracellular compartment of macrophages. Post in vivo and in vitro infection (p.i.), the isolated K96243 morphotypes showed a homogenous rough colony, which was similar to the K9 MT03/MT02 morphotype prior to infection (a.i.) (Fig 1B). These isolates were cultivated again in liquid modified RPMI medium for 60 hours and samples to investigate the metabolic footprint were taken. We found that, in contrast to a stationary phase, a second lower growth rate was established after the exponential phase and remained until the end of the experiment (Fig 2). Most isolates of in vitro infections were growing within 60 hours to similar optical densities between 1.97 and 2.24, whereas MT20 grew up to 3.21 and MT02 only reached 0.79. A slightly higher growth was determined for isolates after in vivo infection which grew to optical densities between 2.57 and 3.22 with MT02 being an exception with a maximum of only 0.96.
After in vitro and in vivo infection the isolates were grown again in modified RPMI medium for 60 h. The glucose concentration decreased significantly within the first 8.5 h (exponential phase) in all isolate cultures but slower compared to the a.i. cultivation (Fig 5). Whereas after cultivation for 16 hours pre infection glucose was depleted in the culture media, the lowest glucose concentration was measured in the K96243 MT02 (ex vitro) culture with still 4.97 mM. Other morphotypes were found with even higher extracellular glucose concentrations of up to 8.58 mM. Consequently, we detected gluconate in morphotype cultivations p.i. in slower rising concentrations (Fig 5). The maximum extracellular gluconate concentration was measured for most morphotypes p.i. after 30 h of cultivation and reached up to 9.00 mM. At 60 h of cultivation the concentration of gluconate decreased in all morphotype cultures p.i. except for MT18 ex vitro and MT20 ex vivo. A significant reduced combined concentration of glucose and gluconate was first measured 30.5 h after inoculation for all isolates except MT02 in vitro and MT18 in vitro (lower but not significant amount) (Fig 5). Whereas prior to infection most of the glucose and gluconate was taken up during the transition phase, in the cultivation p.i. the majority was consumed in the slow growth phase between 30 h and 60 h. In this time period the variation in glucose and gluconate consumption among the morphotypes increased significantly. Other carbon sources were used less after infection. The total amount of consumed citrate was reduced from 8.58 ± 0.74 mM in average pre infection to 3.52 ± 0.42 mM and 3.90 ± 0.65 mM post in vivo and in vitro infection respectively (S3 Fig). Notably the vast majority of citrate was taken up after 16 h of growth. The highest variation in extracellular citrate concentrations was observed at the very end of the experiment. Myo-inositol remained at the initial level for most morphotypes and no uptake was observed (S3 Fig).
The amino acid uptake profile of cultivated morphotypes p.i. showed similarities to the morphotypes a.i. with regard to the classification of amino acids. We found a set of amino acids with a fast uptake profile (set A: gly, ser, pro, gln, asp, asn) and a slow uptake (set B: arg, leu, ile, val, his, lys, thr, tyr, phe, 5-oxoproline and 4-hydroxyproline) (Fig 6A). Additionally, glutamate showed rising concentrations in the extracellular space of some morphotypes indicating a secretion of this amino acid. Subsequently, it was consumed in a fast manner. In all morphotype cultures p.i. (in vitro and in vivo) amino acids of set A are still present 5 h after inoculation, which was not the case before infection. Some amino acids of set B, especially branched chain amino acids, were taken up to a much lesser extend and were not depleted after 60 h of cultivation. For threonine no significant uptake could be measured until 30 h of cultivation for any morphotype. For isoleucine and arginine no uptake between mid-exponential phase and 16 h, 30 h or 60 h depending on the morphotype was observed. In general the data show, that usage of both sets was slower after infection, compared to the cultivation prior to infection (Fig 6B). All K96243 morphotypes showed p.i. (in vivo and in vitro) a more focused use of set A amino acids during the exponential phase. After 16.5 hours approximately 25% of set B amino acids were used in comparison to about 75% prior to infection in the same time (Fig 6B). In the later growth phase, the content of set B amino acids was decreased to 60% in the p.i. cultivation but completely consumed in the a.i. cultivation. Furthermore, the morphotypes p.i. showed very small variation in the usage of amino acids. This can be also shown when we apply the amino acid concentrations to principle component analysis (Fig 6C). The PCA plot of PC1 versus PC3 shows that K96243 morphotypes a.i. form wide spread and overlapping groups which correlate with diverse amino acid uptake profiles compared to morphotypes p.i., which group in smaller clusters and more distinct. The PCA plot of PC1 versus PC2 shows a orthogonal structure caused by the successive uptake of the two amino acid sets as indicated by the loading plot of PC2 (S4 Fig). Even if the morphotypes p.i. are very similar, especially MT02 showed differences compared to the other morphotypes. After in vivo and in vitro infection, the uptake of glu, gln, leu, ile and arg was faster in MT02 than in other isolates. In fact MT02 was the only isolate, which consumed arg completely within 60 h. Additionally a faster uptake was also seen for his, asp, asn, and thr only after in vitro infections. This is indeed surprising since MT02 showed reduced growth compared to the other morphotypes with regard to the optical density.
Prior to infection K96243 MT20 was found to produce the majority of α-keto-acids of all K96243 morphotypes and an additional reuptake of these compounds was observed over time. However, after infection these metabolites were minimally secreted in any morphotype culture within the first 30 h of cultivation. Isolates of K96243 MT02 were found to produce the majority of α-keto-acids starting at 11.5 h or 16 h respectively (Fig 7). Also no reuptake of secreted metabolites was measured. Similar to growth, amino acid uptake and carbon assimilation isolates of in vivo and in vitro infections show rather similar secretion patterns during cultivation (S4 Fig).
In our study, we observed major qualitative similarities in metabolic behaviour of the investigated B. pseudomallei morphotypes like gluconate production and amino acid uptake. However, we also observed various differences in growth, uptake rates and metabolite secretion patterns supporting the idea, that beside the similarity, morphotypes exhibit different regulations on a metabolic level. Interestingly, these variations were strongly reduced, after the morphotypes were isolated from the lung of infected mice or from the intracellular environment of murine macrophages.
Extracellular gluconate production was described before for other proteobacteria like Pseudomonas fluorescence, Pseudomonas cepacia, and B. cepacia and is thought to be connected to the mineral phosphate solubilizing ability of gluconate [29–31]. In a metabolic profiling study, it was shown, that clinical P. aeroginosa isolates from various cystic fibrosis patients do secrete gluconate, and that a small non-coding RNA (CrcZ) regulates gluconate production [32]. Extracellular gluconate is produced by a periplasmatic glucose dehydrogenase (Gdh) and can be further converted by a gluconate dehydrogenase (Gad) into 2-keto-gluconate. Gluconate or 2-keto-gluconate can then be taken up by the cells and further metabolized by either the Entner-Doudoroff-pathway (ED-pathway) or by the pentose-phosphate-pathway (PP-pathway). Consistently, a study investigating intracellular carbon fluxes for various bacteria (among these P. fluorescence and P. putida) showed, that the ED-pathway is the main pathway for carbon catabolism in many bacteria [33]. Surprisingly, neither gdh, nor gad is present in the genome of B.pseudomallei and moreover, we did not detect 2-keto-gluconate in the extracellular space. Therefore, it is still unknown whether gluconate is produced by an unassigned glucose dehydrogenase in the periplasm or by an intracellular glucose dehydrogenase followed by a subsequent secretion into the medium. In the case of P. fluorescence it was described, that gluconate is taken up when glucose is depleted [33]. We also observed gluconate uptake when the extracellular glucose level reached a minimum. Yet with regard to the extracellular gluconate concentration we found two variations. Some morphotypes reached high extracellular gluconate concentrations (71%-88% of the provided glucose) and some cultures contained less gluconate (31%-45%). Either gluconate production is lower in some morphotypes and glucose is taken up directly or the gluconate uptake rate is higher in these morphotypes. If direct glucose uptake is enhanced in these morphotypes they do not show a significant difference in glucose depletion compared to the other morphotypes. Therefore, we would rather suggest that gluconate uptake is elevated in these morphotypes. We found, that low extracellular gluconate concentrations correlated with higher final optical densities, suggesting a faster gluconate uptake favors growth. In general, these data imply that typical gluconate catabolizing pathways as the ED-pathway or the PP-pathway are the preferred pathways for carbohydrate catabolism instead of the Embden-Meyerhof-Parnas-pathway in B. pseudomallei.
After being isolated from the intracellular environment, we found that the uptake of glucose and gluconate was strongly delayed and very similar among all isolates, which consumed mainly certain amino acids during exponential growth. For many bacteria a regulative process called carbon catabolite repression (CCR) is known, which regulates the hierarchy of carbon assimilation and might also play a role in B. pseudomallei. As mentioned above there are similarities between the genus Burkholderia and the close related bacterial family of Pseudomonads. Pseudomonas species exhibit a CCR-system that is different from model organisms like Escherichia coli and Bacillus subtilis. Contrary to these species, Pseudomonads prefer organic acids and amino acids as carbon sources rather than glucose or other sugar derivatives including gluconate [34,35]. Not only the preference for the carbon source is different, but rather the whole CCR-system. Whereas a phosphotransferase-system senses the availability of preferred carbon sources in E. coli and B. subtilis, in Pseudomonas sp. a protein called Crc acts as a global regulator controlled by CrbA, CbrB and the above mentioned small non coding RNA CrcZ [34,36]. Crc does not only distinguish between amino acids and sugars it is also able to establish a hierarchy of amino acid uptake [37]. Unfortunately, very little is known about the process of carbon catabolite repression in Burkholderia species. For B. cepacia the expression of a gene (alkB) involved in alkane degradation as a carbon source seems to be also regulated by catabolite repression [38]. A genome wide search in B. pseudomallei for CrcZ, CbrA and CbrB was without result, but we found a homolog to Crc from P. aeruginosa with 41.9% identity to BPSL0191 in B. pseudomallei. Therefore a similar CCR as described for P. aeruginosa is possible but still speculative in Burkholderia species. Even though the molecular mechanisms of carbon catabolite repression in Burkholderia sp. are unclear, our data indicate that carbon source uptake might be regulated in B. pseudomallei and that certain amino acids (gln, glu, asn, asp, ser, gly, pro) are the preferred carbon sources during exponential growth, especially after infection. Some amino acids of set B (thr, arg, and ile) showed not only a slower uptake over time but rather a delayed uptake, which also points to a hierarchical uptake of amino acids. Another indication towards carbon catabolite repression is the uptake of myo-inositol. It only occurred when glucose and gluconate were depleted, in opposite to citrate which was consumed mostly with other nutrients still being present.
A possible explanation for the distinct regulation of nutrition uptake after infection might be an adaptation to the environment during infection. After in vivo infection, bacteria were isolated from the lung and not differentiated between intra- and extracellular bacteria, whereas in our in vitro approach, bacteria were isolated solely from the intracellular compartment. Interestingly, there was almost no difference in the metabolic behavior between ex vivo and ex vitro isolates suggesting, that the trigger for morphotype switches and changes in metabolic behavior is present in both infection models. We therefore favor the idea, that the trigger that induces morphotype switching is an intracellular signal. Especially intracellular pathogens must have the ability to use nutrients provided by the host cell to satisfy their own needs for replication, since the host-cell cytosol is not a suitable growth medium for bacteria in general [39]. Therefore intracellular pathogens like Francisella tularensis and Legionella pneumophila have developed strategies to manipulate cellular host processes like autophagy or proteasomal degradation to elevate the amount of free amino acids [40,41]. So far, such mechanisms are not known for B. pseudomallei, but like the two former mentioned bacteria, it also shows defects in intracellular replication in a mouse model when the stringent response, an amino acid sensing system, is defect [20]. The slower uptake of certain amino acids after infection could be explained, if, due to the low intracellular abundance in the host, these metabolites have to be synthesized and therefore uptake systems for these metabolites might be downregulated.
Biosynthesis of BCAAs is essential for B. pseudomallei to establish long time persistence and the demand cannot be satisfied by host cell derived BCAAs [42]. Indeed, our data suggest that B. pseudomallei is rather used to low amounts of BCAAs since we found that some morphotypes secreted 3-methyl-2-oxovalerate, 4-methyl-2-oxovalerate and 2-oxoisovalerate during cultivation. Interestingly, the secretion of isovalerate by only two E8 morphotypes indicates morphotype specific metabolic aspects of leucine degradation, whereby the exact pathway remains unclear [43]. A secretion of degradation products of branched chain amino acids has been described previously for two S. aureus strains that were grown in RPMI medium [25]. These metabolites are either substrates of branched-chain amino acid aminotransferase (IlvE) during leucine, isoleucine and valine synthesis, respectively, or products of IlvE, when BCAAs are deaminated during degradation. We also could observe a reuptake of these metabolites in stationary phase, indicating that during exponential growth BCAAs are consumed in excess and cannot be used for biosynthesis or catabolism. Similar to auxothrophic mutants for BCAA-biosynthesis mutants in other pathways like purine synthesis, histidine synthesis and p-aminobenzoic acid synthesis showed attenuated virulence and growth defects in a mouse model [44]. This confirms the lack of specific nutrients inside the host. Such mutants in biosynthesis pathways of essential metabolites are of clinical importance because of their display of potential candidates for vaccination [45].
However, the metabolic adaption and morphotype switch we present in our study is caused by acute infection in a mouse pneumonia model or in a murine macrophage cell line and might only represent a short episode of adaptation. Obligate intracellular bacteria like Buchnera, Wigglesworthia and Blochmannia, which are highly adapted to the host, show extensive gene reductions of housekeeping genes and thereby strong dependences of host nutrient supply [16]. And indeed, a recent study shows that after long-term (12 years) persistence biosynthetic pathways for amino acids are lost due to genome reduction in B. pseudomallei which indicates usage of host provided amino acids or redundant pathways [46].
Further research is needed to investigate i) the biosynthesis pathways, which are required for B. pseudomallei to establish an infection and ii) the intracellular conditions in the host cell, which allow microbial replication. Additionally to metabolic footprints, the intracellular metabolome, the fingerprint, should be examined to identify metabolic differences between B. pseudomallei morphotypes and particularly to uncover the metabolic state of the morphotype after infection. We would favor a combination of biosynthesis pathway mutants and metabolome approaches to uncover the dependence of the microbial metabolism on host derived metabolites.
Overall our metabolic footprint study provides for the first time insights into the so far unknown metabolic characteristics of B. pseudomallei morphology variants. The finding of gluconate production points out, that metabolome approaches are needed to describe the metabolism of an organism, despite the availability of genomic data, since no gluconate production enzymes were assigned so far for B. pseudomallei. We identified a synchronization effect in colony morphology and metabolism due to acute infection that might play an important role in the pathogenicity of B. pseudomallei. Our metabolomic study therefore contributes to the necessary knowledge about a hazardous pathogen and its adaption to the host in the acute phase of melioidosis.
|
10.1371/journal.ppat.1002461 | Slit2/Robo4 Signaling Modulates HIV-1 gp120-Induced Lymphatic Hyperpermeability | Dissemination of HIV in the host involves transit of the virus and virus-infected cells across the lymphatic endothelium. HIV may alter lymphatic endothelial permeability to foster dissemination, but the mechanism is largely unexplored. Using a primary human lymphatic endothelial cell model, we found that HIV-1 envelope protein gp120 induced lymphatic hyperpermeability by disturbing the normal function of Robo4, a novel regulator of endothelial permeability. HIV-1 gp120 induced fibronectin expression and integrin α5β1 phosphorylation, which led to the complexing of these three proteins, and their subsequent interaction with Robo4 through its fibronectin type III repeats. Moreover, pretreatment with an active N-terminus fragment of Slit2, a Robo4 agonist, protected lymphatic endothelial cells from HIV-1 gp120-induced hyperpermeability by inhibiting c-Src kinase activation. Our results indicate that targeting Slit2/Robo4 signaling may protect the integrity of the lymphatic barrier and limit the dissemination of HIV in the host.
| The most common route of HIV transmission is through unprotected sexual contact. By this route, HIV first infects cells in the mucous membranes of the mouth, vagina or rectum. From the mucosa, virus and virus-infected cells move through lymphatic endothelial channels to draining lymph nodes where they infect various cells, including their major target cells, CD4+ T lymphocytes. The virus and infected cells then transmigrate through the lymphatic barrier, enter the blood stream, and spread throughout the body. We found that HIV-1 gp120 compromises the lymphatic endothelial barrier by inducing hyperpermeability. We hypothesize that an impaired barrier may facilitate the dissemination of HIV. Likewise maintaining a “normal” barrier may help slow the dispersal of HIV, thereby protecting the body from HIV spread and progression after initial mucosal exposure. We demonstrated that in lymphatic endothelium the interactions among Robo4, Slit2, fibronectin and α5β1 integrin modulate the effect of HIV-1 gp120 on lymphatic permeability. Moreover, we found that Slit2 inhibits the complexing of Robo4 with fibronectin and protects cells from gp120-induced hyperpermeability. These data suggest that by interacting with Robo4, Slit2 may help maintain the integrity of the lymphatic barrier, thereby interfering with the dissemination of HIV beyond the draining lymph nodes.
| HIV becomes established at mucosal sites by infecting dendritic cells, CD4+ T lymphocytes and macrophages in the lamina propria after its entry. From there, virus and infected cells disseminate via lymphatic endothelial channels to the draining lymph nodes, and subsequently pass into the bloodstream [1]–[4]. An impaired lymphatic barrier may accelerate HIV dissemination.
Generally, endothelial cells do not express CD4, the major receptor of HIV, but express varying levels of CXCR4 [5] and CCR5 [6], the co-receptors of HIV, depending on the tissue of origin [7]. While HIV can infect endothelial cells, its biological importance in the pathogenesis of AIDS is unclear [8]–[10]. The HIV-1 envelope glycoprotein gp120 and the HIV transactivator of transcription (Tat) may contribute to HIV-associated vasculopathy. HIV-1 gp120 induces apoptosis in endothelial cells [11], [12] and Tat stimulates angiogenesis [13], [14], which is often concomitant with hyperpermeability. Current knowledge of the effects of HIV-associated hyperpermeability are limited to disrupting the integrity of vascular structures and/or enhancing inflammatory reactions. However, these phenomena are characteristic of many infectious diseases [15] and do not explain the unique biology of HIV. In addition, while a pivotal role for the lymphatic system in the pathogenesis of HIV/AIDS has been suggested [16], the pathobiology of HIV interaction with lymphatic endothelium has not been extensively characterized.
The Slit2/Robo4 (Roundabout 4) signaling pathway is a recently identified regulator of endothelial permeability [17]. The Slit/Robo family members were originally discovered as axon guidance molecules that mediate repulsive signaling mechanisms in the central nervous system [18]–[20]. Recent studies from animal models strongly implicate a central role for Slit/Robo in vascular biology [17], [21]. For example, Robo4 knockdown zebrafish embryos have vascular sprouting defects [22], and Robo4 knockout mice display abnormal vascular hyperpermeability [17]. Moreover, Slit2/Robo4 interactions can maintain the integrity of the vascular network and its barrier function by inhibiting cytokine-mediated vasculogenesis and enhanced permeability [17], [23], and the Slit2-Robo4-paxillin-GIT1 network inhibits neovascularization and vascular leakage [24].
Slit2 belongs to a family of three glycosylated extracellular proteins containing at least four different motifs and sharing cognate Robo receptors (Robo1-4) [25], [26]. Slits are secreted by midline glial cells and other tissues [19], [27], [28], and can be processed by proteolytic cleavage to yield a shorter C-terminus fragment of unknown function and a longer, active N-terminus fragment that agonizes the Robos [29], [30]. Robo4 is predominantly expressed in endothelial cells, including embryonic endothelium and tumor vascular endothelium, and shows significant structural differences from the other Robos [26], [31]. Robo4 has only two immunoglobulin (Ig) domains and two fibronectin type III domains in the extracellular region, whereas the other Robos have five and three, respectively [32], [33]. The Robo4 cytoplasmic domain also differs from the other family members, e.g. while Robo1 has four conserved motifs in this region, Robo4 retains only two [26]. Structure-effect studies have revealed that the Slits bind via their N-terminal leucine-rich repeat domain to the Robos, and that the first Ig domain of the Robos is highly conserved and important for Slit binding [34]–[36].
Slit2/Robo4 signaling activates Rho GTPases in endothelial cells, but the precise mechanism by which they interact with each other remains controversial [37]–[39]. There are two prevailing hypotheses for their interaction. One posits that Slit2 activates Robo4 and initiates a signaling cascade [17], [21]. Alternatively, Slit2 may interact with Robo1, and then transactivate Robo4 [39], [40].
In this study, we explored if and how HIV-1 gp120 modulates the Slit2/Robo4 signaling pathway in primary human lung lymphatic endothelial cells. We found that HIV-1 gp120 elevated fibronectin levels, activated fibronectin and α5β1 integrin, and induced a physical association between α5β1 and Robo4. This complexing of Robo4 resulted in hyperpermeability in a lymphatic cell monolayer; however, pretreatment with Slit2N, an active N-terminal fragment of Slit2, inhibited significantly these HIV-1 gp120-induced effects. We suggest that the Slit2/Robo4 pathway may play a key role in modulating HIV-1 gp120-induced lymphatic hyperpermeability, and its manipulation may be used to inhibit the dissemination of HIV in the host.
The effects of HIV-1 gp120 on vascular endothelium have been well characterized [41]–[43], however, very little is known about how HIV-1 gp120 specifically affects the lymphatic barrier. To address this issue, we studied the effects of HIV-1 gp120 from two different HIV-1 strains (M-gp120 which utilizes the CCR5 co-receptor on target cells, and T-gp120 which utilizes the CXCR4 co-receptor) on lung lymphatic endothelial cells (L-LECs) in an in vitro, vascular permeability assay. Permeability was quantified by the translocation of FITC-conjugated Dextran particles through an L-LEC cell monolayer seeded in the top chamber of a transwell plate, into the bottom chamber, after incubation with specified concentrations of M-gp120 or T-gp120. We observed a significant increase in permeability of the lymphatic cell monolayer after treatment with both M-gp120 and T-gp120 (Figure 1A). We then assessed in the L-LECs, the expression of CD4 (the major receptor for HIV-1 gp120 on target cells) and the co-receptors, CCR5 and CXCR4, by immunohistochemistry. While we detected no expression of CD4 or CCR5 in these cells (data not shown), we observed a robust expression of CXCR4 on the cell surface and in the nucleus (Figure 1B). However, inhibiting the effects of CXCR4 with a neutralizing antibody had no effect on the HIV-1 gp120-induced permeability of the monolayer (data not shown). These data suggest that HIV-1 gp120 induces hyperpermeability in an L-LEC monolayer by a mechanism independent of CD4, CCR5 and CXCR4 binding.
Fibronectin is important for maintaining vascular integrity [44] and is involved in lymphangiogenesis [45], [46]. Previous studies showed that HIV-1 gp120 can bind to fibronectin through its heparin-binding domains, and facilitate HIV infection [47]–[49]. Therefore, we assessed fibronectin expression by Western blot analysis in L-LECs and their supernatant after incubation with various concentrations of HIV-1 gp120 (M-gp120 was used in all experiments unless specifically stated otherwise). We observed marked increases of fibronectin (predominantly as a dimer) in cell lysates after treatment with HIV-1 gp120, and less pronounced increases of soluble, monomeric fibronectin in the supernatant (Figure 2A). We interpret our data to indicate that HIV-1 gp120 can enhance fibronectin expression in lung lymphatic endothelial cells. Interestingly, we observed that low concentrations of gp120 (10–50 ng/ml) induced a decrease in FN secretion (vs. untreated) as compared with higher gp120 concentrations (100–500 ng/ml). Few experimental studies focus on gp120 at such low levels, however, we hypothesize that the effects of gp120 at these low concentrations may be an in vitro correlate for HIV latent infection in vivo and a low viral load, although this has yet to be confirmed.
With the recent discovery that Slit2/Robo4 signaling regulates endothelial permeability [17], [24], and our data that demonstrate HIV-1 gp120-induced hyperpermeability and fibronectin up-regulation in L-LECs, we postulated that fibronectin, Slit2 and Robo4 may be interacting to regulate lymphatic permeability after HIV exposure. By confocal microscopy, we observed the expression and localization of fibronectin and Robo4 in L-LECs with or without treatment with Slit2N or HIV-1 gp120. After stimulation with Slit2N, no co-localization of fibronectin and Robo4 was observed (Figure 2B, middle panel). When the L-LECs were treated with HIV-1 gp120, however, fibronectin and Robo4 displayed strong co-localization (Figure 2B, right panel). These expression patterns and interactions were corroborated by a Robo4 immunoprecipitation assay in which L-LECs, stimulated with HIV-1 gp120, showed fibronectin activation (by serine/threonine phosphorylation, Figure 2C) and a significantly enhanced physical association between fibronectin and Robo4 (Figure 2C).
Since Slit2/Robo4 signaling is known to inhibit cytokine-induced vascular permeability [17], [23], we compared Slit2 expression in L-LECs in the presence or absence of HIV-1 gp120, a known inducer of endothelial permeability. Using a semi-quantitative RT-PCR assay, we found that at low concentrations, HIV-1 gp120 enhanced Slit2 expression in L-LECs, while higher concentrations of HIV-1 gp120 inhibited the expression of Slit2 (Figure 3). The inhibition of Slit2 by HIV-1 gp120 at 250 ng/ml and 500 ng/ml is consistent with the characterization of Slit2 as an inhibitor of pathological hyperpermeability [24]. Taken together, these data suggest that fibronectin, Robo4 and Slit2 may cooperate in mediating permeability induced by HIV-1 gp120 in lymphatic endothelium.
When fibronectin interacts with vascular endothelium it commonly binds to either α5β1 integrin or αvβ3 integrin on the cell surface. Activation of these integrins dramatically enhances this interaction [50]. Therefore, we examined the expression of these two integrins in L-LECs by Western blot analysis. Since we detected α5β1, but not αvβ3 in L-LECs (data not shown), we investigated only α5β1 in subsequent experiments. We treated L-LECs for various times with either HIV-1 gp120 or Slit2N. By Western blotting we measured the levels of β1 phosphorylation, a reflection of α5β1 activation (Figure 4A). We observed no change in α5β1 activation after Slit2N treatment, however, incubation with HIV-1 gp120 induced significant phosphorylation of β1 (Figure 4A). Furthermore, we observed co-localization of HIV-1 gp120 and activated α5β1 integrin on the L-LEC cell surface by confocal microscopy (Figure 4B, “Merge” panel).
Based on these results, we hypothesized that the physical interaction between HIV-1 gp120 and integrin α5β1 may play a role in HIV-1 gp120-induced effects. Therefore, we examined the effect of blocking this interaction on lymphatic hyperpermeability. Using the previously described in vitro transwell permeability assay, cells were pre-treated with either a neutralizing anti-α5β1 antibody or an isotype control before incubation with M-gp120 or T-gp120. We observed increased permeability through the L-LEC monolayer after treatment with either of the HIV-1 gp120 isotypes (Figure 4C); pretreatment with the anti-α5β1 antibody prevented much of the increase in permeability associated with HIV-1 gp120 (Figure 4C). These data indicate that the increased activation of α5β1 integrin by HIV-1 gp120 and their physical association are required for HIV-1 gp120-induced hyperpermeability of L-LECs.
Based on the results from our expression and co-localization studies of Slit2, Robo4, gp120, fibronectin and α5β1, we sought to investigate further the physical interactions that contribute to HIV-1 gp120-induced hyperpermeability, and to explore the specific effects of Slit2 on these processes. To these ends, we examined the physical interaction of Robo4 and α5β1 in L-LECs after treatment with Slit2N or HIV-1 gp120 in a Robo4 immunoprecipitation assay. The basal association between Robo4 and α5β1 integrin was not affected by the differential expression of Slit2N (Figure 5A). However, we observed a significant increase in this physical association after treatment with HIV-1 gp120 (Figure 5A). We then pretreated L-LECs with Slit2N or a negative control before incubating the cells with HIV-1 gp120. While the association between Robo4 and α5β1 integrin appeared to peak 15 minutes after HIV-1 gp120 incubation (Figure 5B), pretreatment with Slit2N greatly diminished this interaction (Figure 5B). Based on these data, we theorized that Slit2 may antagonize the effects of HIV-1 gp120 on a lymphatic cell monolayer, and therefore, may protect lymphatic endothelium against HIV-1 gp120-induced hyperpermeability. To test this hypothesis we utilized the L-LEC transwell permeability assay previously described. While incubation with M-gp120 and T-gp120 increased L-LEC monolayer permeability (Figure 5C, “Control” bars), the extent of this HIV-1 gp120-induced hyperpermeability was significantly inhibited by pretreatment with Slit2N (Figure 5C, “Slit2N” bars). We interpret these data to indicate that Slit2N significantly inhibits HIV-1 gp120-induced hyperpermeability in lymphatic endothelium by blocking the physical association between Robo4 and α5β1 integrin.
To demonstrate that Slit2N and gp120 can induce similar effects in various types of lymphatic endothelium, we repeated the lymphatic permeability assay using primary human dermal lymphatic endothelial cells (D-LECs). Similar to the results using L-LECs, gp120 increased the permeability of D-LEC monolayers in a dose-dependent manner, and pretreatment with Slit2N significantly decreased the gp120-induced hyperpermeability (Figure 6A). To confirm that the changes in permeability were not due to the origin of the gp120, we repeated this experiment with another M-tropic gp120 protein, gp120CM, and observed similar effects (data not shown).
To demonstrate that the effects of the gp120 protein reflect accurately those of intact HIV-1 virions on lymphatic hyperpermeability, we pretreated L-LEC monolayers with Slit2N or a negative control, followed by incubation with HIV-1 virions or gp120. We found that HIV-1 virions significantly increased lymphatic permeability within 5 hours, whereas gp120 induced only a mild increase during the same time period (overnight incubation was needed for full in vitro effect of gp120 on permeability) (Figure 6B). Pretreatment with Slit2N significantly inhibited the permeability induced by both the HIV-1 virions and the gp120 protein (Figure 6B). Taken together, our results indicate that intact HIV-1 virions increase lymphatic monolayer permeability, and preincubation with Slit2N can effectively inhibit this increase. These data indicate that HIV-1 virions can induce lymphatic endothelial monolayer permeability similar to that induced by gp120, suggesting that our in vitro model of gp120-induced endothelial cell monolayer permeability may reflect the actions of HIV-1 in vivo.
To elucidate the signaling cascade(s) responsible for HIV-1 gp120-induced hyperpermeability in lymphatic endothelium, we analyzed the effects of HIV-1 gp120, Slit2N and Robo4 by Western blot analysis on Src kinase, a key molecule in the regulation of vascular endothelial permeability [51], [52]. We found that preincubation of L-LECs with Slit2N significantly inhibited HIV-1 gp120-induced phosphorylation of c-Src (Figure 7A), indicating inhibition of the Src signaling pathway. We theorized that the modulation of Src kinase signaling by Slit2N and Robo4 may be the result of a physical complexing between the two proteins. To test this hypothesis, we transiently expressed both Robo4 and Myc-tagged Slit2 in 293 cells, and examined their physical association in a Robo4 immunoprecipitation assay. We observed a physical association between Slit2 (c-Myc) and Robo4 in these cells (Figure 7B).
Since pretreatment with Slit2N inhibited c-Src signaling and there appeared to be a physical association between Slit2 and Robo4, we asked if the inhibition of c-Src signaling was a result of Slit2 sequestering Robo4 to deplete its cellular levels and render it unavailable for binding to a competing protein. To approximate this situation, we pretreated L-LECs with a mixture of Robo4 siRNAs or a control siRNA before incubating the cells with HIV-1 gp120. We did not observe the same inhibition of c-Src activation as we had with the Slit2N preincubation (Figure 7A). Instead, the constitutive activation of c-Src increased dramatically in the Robo4 knockdown cells as compared with the control siRNA-transfected cells (Figure 7C). These findings are consistent with the phenotype of Robo4 knockout mice which display heightened vascular permeability [17]. These data suggest that a sufficient endogenous level of Robo4 in lymphatic endothelium is necessary to block c-Src signaling, and that its binding to Slit2 is required to protect against lymphatic hyperpermeability. Additionally, HIV-1 gp120 did not enhance c-Src signaling in the Robo4 knockdown cells as it did in the control siRNA-transfected cells (Figure 7C). We hypothesize that the elevated constitutive level of c-Src kinase signaling in the Robo4 knockdown cells prevented HIV-1 gp120 from enhancing this effect in the L-LECs.
To determine if Src signaling is involved in HIV-1 gp120-induced lymphatic permeability, we pretreated L-LECs with a Src kinase inhibitor or a DMSO control before measuring HIV-1 gp120-induced permeability, as described previously. While treatment with HIV-1 gp120 resulted in increased permeability through the L-LEC monolayer preincubated with DMSO, HIV-1 gp120 had no effect on the L-LECs preincubated with a Src kinase inhibitor (Figure 7D). Taken together, we interpret these data to indicate that Src kinase signaling is required for HIV-1 gp120-induced lymphatic hyperpermeability, and that Slit2/Robo4 interactions can inhibit this signaling cascade.
To characterize more precisely the role of Robo4 in HIV-1 gp120-induced effects on lymphatic permeability, we transfected L-LECs with control siRNAs or Robo4-specific siRNAs (to reduce Robo4 levels), and confirmed a decrease in Robo4 expression by Western blot analysis 24 hours later (Figure 8A). We compared the permeability of L-LEC monolayers expressing endogenous levels of Robo4 (Figure 8B, “Control siRNA” columns) and reduced levels of Robo4 (Figure 8B, “Robo4 siRNA” columns) in the presence or absence of Slit2N or HIV-1 gp120. In the L-LEC monolayers with endogenous levels of Robo4, incubation with Slit2N had no significant effect on permeability, but HIV-1 gp120 significantly increased the permeability of this monolayer. We observed a significantly higher basal level of permeability in the L-LEC monolayers with reduced Robo4 levels. Slit2N had no significant effect on the permeability of these monolayers, and HIV-1 gp120 failed to cause any significant change in the permeability of the L-LEC monolayers with reduced Robo4 levels. We hypothesize that HIV-1 gp120 did not enhance the permeability of these monolayers, because reducing Robo4 levels had already markedly increased their permeability. These data suggest that sufficient endogenous levels of Robo4 are required to maintain an intact lymphatic barrier.
The Robo4 receptor contains two fibronectin (FN) type III domains in its extracellular region [32], [33]. While a study by Kaur et al., demonstrated that these motifs are important for the interaction of Robo4 with fibronectin [38], no other function for the domains has been documented. Fibronectin regulates the permeability of vascular endothelium [44]. We observed that HIV-1 gp120 elevated FN levels and enhanced lymphatic monolayer permeability in L-LECs. Therefore, we examined the effects of fibronectin on c-Src activation, and its effects after pretreatment with Slit2N. We observed that fibronectin enhanced the activation of c-Src, and that pretreatment with Slit2N significantly inhibited the FN-induced activation of c-Src (Figure 9A). We hypothesize that Slit2N may be interacting with Robo4 to block the FN-induced c-Src activation, and that the FN domains of Robo4 may be involved in the inhibition of FN-induced c-Src activation by Slit2 and L-LEC monolayer hyperpermeability.
To explore the potential role of the Robo4 FN domains in HIV-1 gp120-induced effects, we compared the effects of HIV-1 gp120 on the permeability of L-LEC monolayers transfected with wild-type Robo4 (WT), mutant Robo4 (MT), which lacks the FN type III domains, or a vector control (V). We found that HIV-1 gp120 induced significantly less permeability in L-LEC monolayers with elevated levels of wild-type Robo4 as compared to those with endogenous Robo4 levels (Figure 9B). These results indicate that Robo4 inhibits HIV-1 gp120-induced permeability in L-LEC monolayers, and may protect the integrity of the lymphatic barrier after HIV-1 infection by interacting with Slit2. We also observed that HIV-1 gp120-induced permeability was inhibited to a significantly greater extent in the L-LEC monolayers transfected with mutant Robo4 vs. wild-type Robo4. In fact, treatment with HIV-1 gp120 resulted in no change in the permeability of the L-LEC monolayers expressing mutant Robo4 (Figure 9B). We interpret these results to indicate that the complexing of FN and Robo4 (through its FN type III domains) is necessary for HIV-1 gp120-induced hyperpermeability of L-LEC monolayers, and that the FN type III domains of Robo4 may be required for the interaction of HIV-1 gp120, Robo4 and FN.
To explore this hypothesis, we transiently transfected L-LECs with plasmids encoding wild-type Robo4 (WT), mutant Robo4 (MT), or a vector control (V). After 48 hours, we analyzed the effects of HIV-1 gp120 on c-Src pathway activation in each of the transfected cell types by Western blot analysis. We observed that the basal level of c-Src activation was lower in L-LECs with elevated Robo4 expression as compared to those with endogenous Robo4 expression (Figure 9C, WT/− and V/−, respectively). HIV-1 gp120 increased c-Src activation in both cell types (Figure 9C), however, overall HIV-1 gp120-induced c-Src activation levels were significantly lower in L-LECs with elevated Robo4 levels as compared to those with endogenous Robo4 levels (Figure 9C, WT/+ and V/+, respectively). In L-LECs expressing elevated levels of mutant Robo4 (MT), both basal c-Src activation and HIV-1 gp120-induced c-Src activation were equivalent to the L-LECs expressing elevated wild-type Robo4 (Figure 9C). These data indicate that elevated levels of Robo4 inhibit basal c-Src activation and HIV-1 gp120-induced c-Src activation. We hypothesize that since Slit2 inhibits c-Src activation, elevated Robo4 levels after transfection may magnify the effects of Slit2, by providing more receptors to which endogenous Slit2 can bind.
We also examined the levels of HIV-1 gp120-induced phosphorylation of ERK1/2, key signaling molecules for endothelial cell function, by Western blot analysis, using the same three groups of L-LEC transfectants. HIV-1 gp120 induced a significant increase in ERK1/2 phosphorylation in the L-LECs transfectants with elevated wild-type Robo4 expression as compared to those with endogenous Robo4 expression (Figure 9C). Although HIV-1 gp120 increased the phosphorylation of ERK1/2 in the L-LECs transfected with mutant Robo4, the increase was significantly lower than the wild-type Robo4 transfectants (Figure 9C). These data indicate that while the FN domains of Robo4 are not required for the inhibition of gp120-induced c-Src activation, they are required for gp120-induced phosphorylation of ERK1/2. We hypothesize that the FN domains of Robo4 may participate in the activation of other key signaling molecules like ERK1/2, however, further investigation is needed to fully understand their function.
The integrity of the lymphatic barrier requires a dynamic interaction between fibronectin, other extracellular matrix (ECM) proteins, and their receptors, cell-surface integrins [46], [53]. As a result of HIV-1-induced inflammation and increased protease expression, fibronectin fragments are detected in the blood of HIV-infected patients [54]–[56]. These fragments are believed to promote the transendothelial migration of HIV-1-infected and non-infected leukocytes, and to promote viral stability and cell-to-cell transmission [48], [56], [57]. We found that lymphatic endothelial cells produce elevated levels of cell-bound fibronectin after exposure to HIV-1 gp120 (Figure 2A). This elevation appeared to modulate the integrity of the lymphatic barrier. In particular, HIV-1 gp120 induced activation of α5β1 integrin which enhanced the physical complexing of HIV-1 gp120, fibronectin, α5β1 integrin and Robo4, and resulted in lymphatic hyperpermeability (Figures 1A, 2B, 2C, 4A and 4B).
While FN/integrin [45] and Slit2/Robo4 [17] interactions are both important for endothelial permeability, little is known about their relationship. α5β1 and αvβ3 are two major integrins expressed on the surface of endothelial cells [45]. We and others have shown that integrin α5β1, but not αvβ3, clustered in focal contacts of endothelial cells during stressful cellular conditions (Figure 4B) or incubation with fibronectin [58]–[60]. In this study, we found that Robo4 formed a complex with fibronectin and integrin α5β1 at low, basal levels in uninfected lymphatic endothelial cells (Figures 2B, 5A and 5B). Slit2N did not alter this association, which is important for maintaining the integrity of the lymphatic barrier (Figures 5A and 5C). However, exposure to HIV-1 gp120 enhanced the association of α5β1 and Robo4 (Figure 5A) and resulted in increased lymphatic permeability (Figures 1A and 5C). Moreover, pre-incubation with Slit2N blocked the HIV-1 gp120-induced enhanced complexing of Robo4 and α5β1, (Figure 5B) and lymphatic hyperpermeability was reduced (Figure 5C). These data suggest that α5β1 integrin may also participate in the effects of HIV-1 gp120 on lymphatic permeability, and Slit2 may help sustain the integrity of the lymphatic barrier after HIV-1 exposure.
Activation of the Src kinases modulates cytoskeletal remodeling and affects cell-to-cell and cell-to-ECM adhesion [61], [62]. Our data indicate that HIV-1 gp120 and Slit2 exert opposing effects on c-Src kinase signaling, namely, HIV-1 gp120 activates c-Src signaling, while pretreatment with Slit2N significantly reduces these effects (Figure 7A). Moreover, the enhancement or inhibition of Src kinase signaling and the resulting effect on lymphatic permeability is critically dependent on Robo4 levels (Figures 9A, 9B, and 9C).
Robo4 displays unique structure and function, but the relationship between these characteristics is largely unknown [26], [32], [33]. Although the first Ig domains of the Robos are highly conserved and important for Slit binding, direct binding of Slit2 to Robo4 is still debated [37]–[39]. The Robo proteins, including Robo4, contain fibronectin type III domains. Previous studies, which found that the FN domains were required for adhesion to fibronectin, suggest that these domains may play a central role in modulating vascular permeability [38]. Our data strongly support their function in Robo4-mediated lymphatic permeability upon HIV-1 gp120 stimulation, and imply a potential role for Robo4 in fibronectin-associated vasculopathies, such as HIV-associated pulmonary hypertension [63].
Based on our new data and that of others, we propose a hypothetical model for the interactions of HIV-1 gp120, FN, α5β1 integrin, Robo4 and Slit2, and their effect on lymphatic permeability (Figure 10). Robo4 and α5β1 integrin are transmembrane proteins expressed in lymphatic endothelium. We hypothesize that under normal, physiological conditions, soluble FN and Slit2 are expressed at low, basal levels, and they interact with Robo4 via its FN type III domains and Ig domains, respectively. FN also binds to α5β1 integrin, which is expressed on the endothelial cell surface. Under these conditions, the integrity of the lymphatic endothelial barrier is intact, and transmigration through the endothelial barrier is severely restricted. We propose that upon HIV infection, HIV-1 gp120 elevates FN levels significantly and complexes with FN. FN then activates α5β1 integrin, which results in enhanced intracellular signaling through α5β1 integrin, a significantly stronger interaction between FN and Robo4, and enhanced intracellular signaling through Robo4. These changes activate the c-Src signaling pathway and induce hyperpermeability of the lymphatic endothelial barrier. The resulting “leaky” barrier may facilitate the dissemination of HIV-1 and virus-infected cells throughout the body. Furthermore, we propose that elevated levels of Slit2 may protect the lymphatic channels from HIV-induced vasculopathy and HIV spread. We hypothesize that at sufficiently elevated levels, Slit2 will bind strongly to the Ig domains of Robo4 and inhibit c-Src pathway activation and HIV-1 gp120-induced lymphatic hyperpermeability. Slit2 may affect this inhibition by various means. A likely senario is that upon binding, Slit2 alters the protein conformation of Robo4, which may lessen/abolish its ability to interact with FN, alter α5β1 integrin intracellular signaling, and inhibit the activation of c-Src. In addition, the binding of Slit2 may alter also the signaling through Robo4, which may inhibit c-Src pathway activation and lymphatic hyperpermeability. Although our data strongly support this model, further investigation is needed to confirm it, or posit alternative mechanisms for the effects of HIV-1 gp120, FN, α5β1 integrin, Robo4 and Slit2 on lymphatic permeability.
Multiple studies indicate that the lymphatic channels play important roles in the establishment of HIV infection, and its dissemination throughout the host [1]–[4]. HIV-induced lymphadenopathy, including lymphoedema, is commonly seen among HIV-infected individuals with Kaposi's sarcoma, a vascular neoplasm which is derived from lymphatic endothelial cells [64], [65]; however, dysfunction of the lymphatic vasculature and its effects on HIV biology are largely unexplored. We established an in vitro endothelial monolayer model to study the effects of HIV on lymphatic permeability. In this model, HIV-1 gp120 and HIV-1 virions both induced lymphatic hyperpermeability, which was significantly inhibited by Slit2 preincubation (Figure 6B). These results suggest key roles for gp120, FN, and Slit2/Robo4 in HIV-associated lymphatic hyperpermeability, and implicate lymphatic hyperpermeability in HIV infection and spread throughout the body. Future studies to explore the traversion of HIV virions or virus-infected cells through the lymphatic endothelium and its contribution to HIV infection should provide more evidence on HIV-induced lymphatic hyperpermeability and HIV dissemination in a humanized mouse model of HIV infection.
In summary, we found that the balance between HIV-1 gp120/FN/α5β1 integrin-induced signaling and Slit2/Robo4-induced signaling in L-LECs modulates lymphatic monolayer permeability. Targeting these pathways may offer novel approaches to inhibit HIV-induced lymphatic injury, and limit the dissemination of HIV in the host.
Human embryonic kidney cells (293 cells) (Stratagene, La Jolla, CA, USA) were cultured in Dulbecco's modified Eagle's medium with 10% fetal calf serum. Primary human lung lymphatic endothelial cells (L-LECs) and dermal lymphatic endothelial cells (D-LECs) were purchased from Lonza, Inc. (Allendale, NJ, USA) and maintained in EBM-2 medium with EGM-2MV SingleQuots (Lonza, Inc.).
Recombinant human Slit2N (the active fragment of Slit2) was provided by Dr. Dean Li, Department of Oncological Sciences at the University of Utah. The following reagents were obtained through the AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH: recombinant HIV-1Ba-L gp120 (M-gp120) protein and HIV-1 virus (strain HIV-1 Ba-L, which was from Dr. Suzanne Gartner, Dr. Mikulas Popovic and Dr. Robert Gallo). Per NIH data sheet, HIV-1Ba-L was originally isolated from a primary culture of adherent, human, infant lung tissue cells, and amplified in human monocytes/macrophages. Virus was harvested 10 days post-infection. This virus was used in the transendothelial monolayer permeability assays. HIV-1 gp120 LAV (III B) (T-gp120) was purchased from Protein Sciences Corporation (Meriden, CT, USA). HIV-1 gp120CM was purchased from ProSpec-Tany TechnoGene Ltd. (Ness Ziona, Israel). Src inhibitor-1 was purchased from Sigma-Aldrich, Corp. (St. Louis, MO, USA).
Mouse monoclonal antibody to the HIV-1 gp120 was purchased from Advanced Biotechnologies Inc. (Columbia, MD, USA). Anti-integrin β1, anti-phospho-integrin β1 (Tyr783), and neutralizing anti-integrin α5β1 antibodies were purchased from Millipore Corp. (Billerica, MA, USA). Anti-phospho-Src kinase family antibodies were purchased from Cell Signaling Technology, Inc. (Beverly, MA, USA). All other antibodies were purchased from Santa Cruz Biotechnology, Inc. (Santa Cruz, CA, USA).
Total cell RNA was extracted using the RNeasy Mini Kit from Qiagen, Inc. (Valencia, CA, USA). RT-PCR was performed using a one-step RT-PCR kit from Clontech (Mountain View, CA, USA). Specific primers for CXCR4 and CCR5 were purchased from R & D Systems (Minneapolis, MN, USA). The primers for human Slit2 were synthesized by Invitrogen Corp. (Carlsbad, CA, USA). The sequences are: upstream: 5′-GGTGTCCTCTGTGATGAAGAG -3′; downstream: 5′- GTGTTTAGGACACACACCTCG -3′.
Cells cultured in 8-well chamber slides (Thermo Fisher Scientific Inc., Waltham, MA, USA) were fixed with 4% (v/v) paraformaldehyde solution for 1 hour at room temperature, incubated in fresh permeabilization solution (0.1% sodium citrate in 1% Triton X-100 in 1× PBS) for 2 minutes on ice, incubated with 3% BSA/1× PBS on ice for 30 minutes, and then with anti-human CXCR4 rabbit polyclonal antibody or normal rabbit IgG (Millipore Corp.) at 4°C for 1 hour. The slides were then washed 3 times in 1× PBS, and incubated with a FITC-conjugated, goat anti-rabbit IgG antibody (Vector Laboratories, Burlington, CA, USA) at 4 °C for 30 minutes. The slides were washed again 3 times in 1× PBS, and then air dried and mounted with mounting medium (Vector Laboratories).
Lymphatic endothelial cells were seeded in the top chamber of transwell plates, according to the manufacturer's instructions (Millipore Corp.), starved for one hour, and then incubated with different reagents or their respective controls as indicated. Subsequently, FITC-Dextran was added to the top chamber and allowed to permeate through the monolayer to the lower chamber for 5 minutes. The extent of permeability was determined by measuring the fluorescence of the solution in the lower chamber by a standard plate reader (BioTek Instruments, Inc., Vinooski, VT, USA). The gp120 control was prepared by boiling gp120 for 10 minutes to inactivate its protein activity while preserving its inherent endotoxin activity. It was employed here, and in all other experiments that required a gp120 control.
Cells were starved for 2 hours in serum-free media, and then stimulated as indicated. Cells were lysed in RIPA buffer (Cell Signaling Technology, Inc.) after stimulation. Immunoprecipitation and Western blotting were performed as described previously [66]. For quantitation, the ratio of protein expression, phosphorylation, or association vs total protein in each lane was obtained by densitometry with a gel imaging system (Cell Biosciences, Inc., Santa Clara, CA, USA).
The pCMV6 Entry expression plasmid encoding Myc-DDK-tagged Slit2 was purchased from OriGene Technologies, Inc. (Rockville, MD, USA). The expression plasmid encoding RFP-tagged Robo4 was constructed as follows. Robo4 cDNA was amplified from the pCMV-SPORT6 containing Robo4 cDNA (Thermo Fisher Scientific Inc.), using primers purchased from Invitrogen Corp. (upstream sequence: 5′-GAGGCGATCGCATGGGCTCTGGAGACAGCCTCCTG-3′; downstream sequence: 5′-GCGACGCGTGGAGTAATCTACAGGAGAAGCACCAGC-3′). The purified PCR product was digested with Sgf I plus Mlu I, and inserted into the pCMV6-AC-RFP plasmid digested with same restriction endonucleases to create the pCMV6-AC-RFP-Robo4. To make the mutant Robo4 expression plasmid, we designed a pair of primers to amplify a section of the pCMV6-AC-RFP-Robo4 plasmid by PCR. The primers are: 5′-CCCCCCCCGCTAGCTCTAGGCTTGGGGCCCTCTGCAGGATC-3′ and 5′-TTTTTTTTGCTAGCCCTGTCTGCCTCCTTTTAGAGCAGGCC-3′. The PCR product was digested with Nhe1, purified, and ligated with T4 DNA ligase at 16°C overnight. The ligation product was used to transform competent DH5 α cells. Positive clones were screened and confirmed by DNA sequencing.
Specific Robo4 siRNAs and control siRNAs, purchased from Santa Cruz Biotechnology, Inc., were used to transfect L-LECs using HiPerFect transfection reagent from Qiagen, Inc.
Cells were grown to 60% confluence in tissue culture dishes. Transfections were done using Super Effectene transfection reagent according to the manufacturer's instructions (Qiagen, Inc.). At 3 hours post-transfection, cells were washed once with 1× PBS, then cultured in full medium for 48 hours. The transfection efficiency was determined by detection of red fluorescent cells under a fluorescent microscope (Nikon Diaphot 300, Tokyo, Japan).
Cells cultured in 8-well chamber slides (Thermo Fisher Scientific Inc.) were serum starved for 2 hours, and then treated with HIV-1 gp120 as indicated. Subsequently, cells were fixed with 4% (v/v) paraformaldehyde for at least 1 hour at room temperature and permeabilized for 2 minutes on ice. Cells were then incubated with primary antibodies or normal IgG controls overnight at 4°C, and washed 3 times with 1× PBS. Fluorescence-conjugated secondary antibodies were added for 30 minutes at 4°C and the cells were washed 3 times in 1× PBS. Finally, the chambers were removed and coverslips were affixed with mounting medium containing DAPI (Vector Laboratories). Slides were examined under a Leica TCS-NT laser scanning confocal microscope (Leica Microsystems, Bannockburn, IL, USA).
Each experiment was repeated at least 3 times, and representative blots, images, or graphs are shown in the figures. Statistical significance was determined using the ANOVA test (*p<0.05).
|
10.1371/journal.pcbi.1005116 | Decreases in Gap Junction Coupling Recovers Ca2+ and Insulin Secretion in Neonatal Diabetes Mellitus, Dependent on Beta Cell Heterogeneity and Noise | Diabetes is caused by dysfunction to β-cells in the islets of Langerhans, disrupting insulin secretion and glucose homeostasis. Gap junction-mediated electrical coupling between β-cells in the islet plays a major role in coordinating a pulsatile secretory response at elevated glucose and suppressing insulin secretion at basal glucose. Previously, we demonstrated that a critical number of inexcitable cells can rapidly suppress the overall islet response, as a result of gap junction coupling. This was demonstrated in a murine model of Neonatal Diabetes Mellitus (NDM) involving expression of ATP-insensitive KATP channels, and by a multi-cellular computational model of islet electrical activity. Here we examined the mechanisms by which gap junction coupling contributes to islet dysfunction in NDM. We first verified the computational model against [Ca2+] and insulin secretion measurements in islets expressing ATP-insensitive KATP channels under different levels of gap junction coupling. We then applied this model to predict how different KATP channel mutations found in NDM suppress [Ca2+], and the role of gap junction coupling in this suppression. We further extended the model to account for stochastic noise and insulin secretion dynamics. We found experimentally and in the islet model that reductions in gap junction coupling allow progressively greater glucose-stimulated [Ca2+] and insulin secretion following expression of ATP-insensitive KATP channels. The model demonstrated good correspondence between suppression of [Ca2+] and clinical presentation of different NDM mutations. Significant recoveries in [Ca2+] and insulin secretion were predicted for many mutations upon reductions in gap junction coupling, where stochastic noise played a significant role in the recoveries. These findings provide new understanding how the islet functions as a multicellular system and for the role of gap junction channels in exacerbating the effects of decreased cellular excitability. They further suggest novel therapeutic options for NDM and other monogenic forms of diabetes.
| Diabetes is a disease reaching a global epidemic, which results from dysfunction to the islets of Langerhans in the pancreas and their ability to secrete the hormone insulin to regulate glucose homeostasis. Islets are multicellular structures that show extensive coupling between heterogeneous cellular units; and central to the causes of diabetes is a dysfunction to these cellular units and their interactions. Understanding the inter-relationship between structure and function is challenging in biological systems, but is crucial to the cause of disease and discovering therapeutic targets. With the goal of further characterizing the islet of Langerhans and its excitable behavior, we examined the role of important channels in the islet where dysfunction is linked to or causes diabetes. Advances in our ability to computationally model perturbations in physiological systems has allowed for the testing of hypothesis quickly, in systems that are not experimentally accessible. Using an experimentally validated model and modeling human mutations, we discover that monogenic forms of diabetes may be remedied by a reduction in electrical coupling between cells; either alone or in conjunction with pharmacological intervention. Knowledge of biological systems in general is also helped by these findings, in that small changes to cellular elements may lead to major disruptions in the overall system. This may then be overcome by allowing the system components to function independently in the presence of dysfunction to individual cells.
| Multi-cellular biological systems are composed of cellular elements with distinct characteristics, which function collectively as a result of dynamic interactions. While the function of a multicellular system is dependent on the characteristics of its constituent cells, understanding such systems is complicated by the action of cellular coupling and system architecture. Furthermore, cellular heterogeneity and noise complicate assessment of the function of individual cells. As a result, changes in the behavior of individual cells can often lead to unexpected changes in the system behavior. Many diseases, both acute and chronic, arise through genetic variations that impact molecular and cellular function. Given the complexities of multi-cellular systems, effectively predicting how molecular and cellular dysfunction lead to tissue and organ dysfunction and cause disease is challenging. One approach to describe dynamic multicellular systems is using network theory, which distinguishes network structure and cellular behavior to understand how distinct functions can emerge from coupled systems [1,2].
Islets of Langerhans located in the pancreas exhibit complex multicellular behavior. Islets are small (~1000 cells) micro-organs, where the primary cellular elements are insulin secreting β-cells. Death or dysfunction to β-cells and a reduction or absence of insulin secretion is the main cause of diabetes. β-cells are excitable, where glucose-stimulated insulin secretion is driven by electrical activity. The increased metabolism of glucose following blood glucose elevation increases ATP/ADP and inhibits ATP-sensitive K+ (KATP) channels. The resulting membrane depolarization activates bursts of action potentials and elevates intracellular free-calcium activity ([Ca2+]) in the form of oscillations which triggers pulses of insulin granule exocytosis [3–5]. β-cells in the islet are electrically coupled by Connexin36 (Cx36) gap junction channels [6–8]. As a result of electrical coupling, [Ca2+] oscillations are coordinated under elevated glucose and uniformly silent under basal glucose [9–12].
Neonatal diabetes mellitus (NDM) is a monogenic form of diabetes that arises in the first 6 months of life. It can have a transient phenotype (TNDM); be permanent (PNDM) where the patient requires lifelong treatment with insulin or sulfonylureas; and may display neurological features (DEND) [13]. The majority of NDM cases have mutations on the KCNJ11 and ABCC8 genes [13–15], which encode the Kir6.2 and SUR1 subunits of the KATP channel. Mutations to these subunits often result in reduced sensitivity to ATP and increased open channel stability [14,16–22]; thereby limiting electrical activity and insulin secretion. This was shown in a murine model of NDM [23,24], where expression of ATP insensitive KATP channels led to blunted glucose-stimulated [Ca2+], insulin secretion and sharply elevated blood glucose levels. Sulfonylurea therapy to inhibit KATP channel is effective in many NDM patients with KCNJ11 and ABCC8 mutations [25], especially if treated early [26]. However some NDM patients with specific KCNJ11 and ABCC8 mutations do not respond to sulfonylureas and require life-long insulin therapy [25].
Our prior studies have shown that as a result of electrical coupling, small changes in β-cell excitability (e.g. upon glucose stimulation) leads to a rapid change in global islet behavior [27]. Similar behavior, termed critical behavior, has been observed in other systems, including cardiac pacemaker cells and GnRH secretion [28,29]. In the islet, critical behavior also leads to large decreases in electrical activity following small changes in the expression of ATP-insensitive mutant KATP channels. This occurs as a result of the suppressive effect subpopulations of inexcitable cells have in the islet [27]. Specifically a threshold number of inexcitable cells can be tolerated, beyond which electrical activity and insulin secretion is severely diminished. Importantly similar behavior occurs following uniform KATP-channel activation with diazoxide, where some cells are less-excitable and some more excitable due to endogenous heterogeneity [30]. Such behavior was predicted by both a network model of islet cellular coupling [27,31], and a dynamic islet electrophysiological model [27,32], where in each case the overall islet electrical response is determined by the excitability of the constituent cellular population and the level of coupling. Consistent with predictions from these models, when Cx36 expression was reduced in the presence of ATP-insensitive mutant KATP expression, or diazoxide-mediated KATP activation, a recovery in islet [Ca2+] and insulin secretion was observed. This also prevented hyperglycemia and diabetes as a result of ATP-insensitive mutant KATP expression [33]. These findings imply that gap junction electrical coupling aids in the coordinated suppression of islet [Ca2+] and insulin release caused by mutations in KCNJ11 and ABCC8. Thus a reduction of gap junction coupling in the islet may recover islet function and blunt diabetes caused by mutations in KCNJ11 and ABCC8.
Another property of coupled multi-cellular systems is a lack of stochastic noise. At the single cell level stochastic behavior is observed in many cellular processes including gene transcription [34,35], and ion channel activity [36,37]. As discussed above, individual β-cells show heterogeneous function [30], with some cells showing reduced metabolic activity, excitability and insulin secretion compared to others [38–40]. However, isolated β-cells also show noisy, irregular fluctuations in membrane potential and [Ca2+] [41,42]. Ion channels show stochastic behavior where the channel switches rapidly between and open and closed states [36,43]. Previous work has demonstrated a strong effect of stochastic noise on [Ca2+] oscillations in uncoupled β-cells, but negligible effect in the presence of coupling [44]. In the absence of gap junction coupling stochastic behavior has been suggested to elevate [Ca2+] [33], and may therefore contribute to improved islet function in the presence of ATP-insensitive KATP channels.
In this study, we compared a computational model of islet electrical activity and insulin secretion against experimental measurements upon varied mutant KATP channel expression in the presence of different levels of gap junction coupling. Using this computational model, we then examined the role of gap junction coupling in exacerbating islet dysfunction in the presence of specific KCNJ11 and ABCC8 that cause NDM. We further examined the degree to which islet dysfunction can be overcome by reducing gap junction coupling, and the relative role of cellular heterogeneity and stochastic channel noise in mediating this.
β-cell specific expression of Kir6.2[ΔN30,K185Q] under inducible CreER control well-models human NDM [23,24,45,46]. To quantify how gap junction coupling exacerbates suppression of islet electrical activity following Kir6.2[ΔN30,K185Q] expression we imaged [Ca2+] in islets of mice expressing variable levels of Kir6.2[ΔN30,K185Q], as indicated by co-expression of GFP, for reduced levels of Cx36 (Fig 1A and 1B). Upon low levels of Kir6.2[ΔN30,K185Q] expression, [Ca2+] elevation was observed across the islet, where the majority of β-cells showed frequent oscillations both in the presence and absence of Cx36 (Fig 1A). Upon high levels of Kir6.2[ΔN30,K185Q] expression, few [Ca2+] transients were observed in the presence of Cx36, but these transients were more frequent, in more cells, in the absence of Cx36 (Fig 1B). Given heterogeneous cell responses, we quantified the proportion of cells showing significant elevations in [Ca2+] (which we now refer to as ‘[Ca2+]’). In the presence of normal gap junction coupling (Cx36+/+) a sharp transition was observed between high [Ca2+] (many cells showing elevations) and low [Ca2+] (few cells showing elevations) as Kir6.2[ΔN30,K185Q] expression increased (Fig 1C) [27]. In the presence of ~45% coupling (Cx36+/-) the decline was less sharp, and right-shifted to higher Kir6.2[ΔN30,K185Q] expression (Fig 1D). In the absence of coupling (Cx36-/-) the decline in [Ca2+] was gradual (Fig 1E). Regression analysis also showed the decline in [Ca2+] with increasing Kir6.2[ΔN30,K185Q] become more gradual with reduced or absent gap junction coupling.
Previously, a computational model of islet electrical activity showed good agreement with experimental measurements of [Ca2+] upon increased Kir6.2[ΔN30,K185Q] expression [27]. To further validate this model against our experimental measurements, we included the presence of variable Kir6.2[ΔN30,K185Q]-expressing cells and reduced electrical coupling (Fig 1A and 1B). Cx36+/+ (100% coupling) was represented by 120pS; Cx36+/- (~45% coupling) was represented by 50pS; and Cx36-/- (~0% coupling) was represented by 0pS [10,47]. In all cases there was good agreement between simulations and experimental measurements (Fig 1C–1E): Simulations representing Cx36+/- showed a less rapid and right-shifted transition and simulations representing Cx36-/- showed a linear decline, although the simulations generally showed a sharper transition than experimental measurements. Simulations also agreed with experimental data that above ~20% Kir6.2[ΔN30,K185Q] expression near silent behavior was observed with normal levels of gap junction coupling, and a progressive increase in [Ca2+] occurred as gap junction coupling was reduced.
We next examined whether the gap junction dependence of Kir6.2[ΔN30,K185Q]-induced suppression of [Ca2+] was linked to altered insulin secretion and glucose homoeostasis (S1 Fig). Measurements of [Ca2+], insulin secretion, plasma insulin and glucose were grouped according to the level of Kir6.2[ΔN30,K185Q] expression. Upon low (<20%) Kir6.2[ΔN30,K185Q] expression little decrease in [Ca2+] was observed for all Cx36 levels (Fig 2A), whereas upon high (>20%) Kir6.2[ΔN30,K185Q] expression the reduction in [Ca2+] was significantly less for reduced Cx36 levels (Cx36+/-) or absent Cx36 (Cx36-/-) compared to normal Cx36 levels (Cx36+/+) (Fig 2A).
Insulin secretion and plasma insulin showed general agreement with [Ca2+] measurements (Fig 2B and 2C). Upon low (<20%) Kir6.2[ΔN30,K185Q] expression little decrease was observed for all Cx36 levels, and upon high (>20%) Kir6.2[ΔN30,K185Q] expression there was no reduction in insulin secretion and plasma insulin for Cx36-/-, compared to large reductions for Cx36+/+ (Fig 2B and 2C). However large reductions in plasma insulin and insulin secretion were also observed for Cx36+/- (Fig 2B and 2C), although the variability was greater than Cx36+/+ (S1 Fig). Blood glucose showed good correspondence to plasma insulin and insulin secretion measurements (Fig 2D). Upon low (<20%) Kir6.2[ΔN30,K185Q] expression, euglycemia was observed for all Cx36 levels. Upon high (>20%) Kir6.2[ΔN30,K185Q] expression euglycemia was observed for Cx36-/- but substantial hyperglycemia was observed for Cx36+/+ and Cx36+/-, albeit slightly less for Cx36+/-. Regression analysis showed similar strong variations in insulin secretion, plasma insulin levels and blood glucose with increasing Kir6.2[ΔN30,K185Q] expression for Cx36+/+ and Cx36+/-, but less or absent variations for Cx36-/- (Fig 2B–2D).
While Kir6.2[ΔN30,K185Q] expression well-models NDM, mosaic expression of mutant KATP channels is unlikely in human disease. However a similar rapid onset of suppression also occurs upon increasing diazoxide activation of KATP channels across all β-cells, due to endogenous β-cell heterogeneity [27]. Similar behavior also occurred when simulating both a progressive increase in mosaic over-active KATP channels (Pmut, modelling Kir6.2[ΔN30,K185Q]) and uniform KATP activation (α, modelling diazoixde) (Fig 3A and 3B). A sharp transition occurred in each case in the presence of full electrical coupling, whereas with reduced electrical coupling a right-shifted transition occurred, which was less sharp for increasing Pmut. In the absence of electrical coupling a further right-shifted and more gradual transition occurred in each case. Therefore similar behavior occurs for different ways in which reductions in excitability are introduced to the islet.
We predicted similar gap junction dependence would occur in the presence of over-active KATP channels resulting from mutations to KCNJ11 and ABCC8 that cause NDM. We simulated how altered ATP-inhibition kinetics (Eq 7) impact [Ca2+]. Several KCNJ11 and ABCC8 mutations have been reported to show a residual current at saturating ATP concentrations (α) [20,22] and all mutations are reported to show a shift in the half maximal ATP concentration for KATP inhibition (k’1/2) [14,16–21,48–56]. The decline in [Ca2+] upon increasing residual current (α) is equivalent to that in Fig 3B. Upon increasing the half maximal ATP concentration (k’1/2) there was also a decline in [Ca2+] (Fig 3C–3E). In the presence of full electrical coupling, all cells remained active until k’1/2 increased by ~4.3, at which point a sharp transition to full suppression occurred (Fig 3F). In the presence of progressively reduced electrical coupling, similar sharp transitions occurred but at greater k’1/2 values, which became less sharp at <15% electrical coupling (Fig 3F). In the absence of electrical coupling, a gradual transition was observed where many cells remained active for k’1/2 values up to ~10 (Fig 3F). The [Ca2+] duty-cycle showed similar behavior (Fig 3G), although no decrease occurred for absent electrical coupling at low k’1/2 values. Simulations at full electrical coupling were well-repeatable, but showed variable shifts in activity at reduced electrical coupling (Fig 3H). As such for <45% electrical coupling the transition became more gradual averaged over an ensemble of islets.
We then characterized the β-cells that showed elevated [Ca2+] in the absence of electrical coupling (Fig 3I). For increased Pmut (Fig 3A), cells with elevated [Ca2+] lacked Kir6.2[ΔN30,K185Q] expression and were not distinguished by differences in other parameters. For increased α (Fig 3B), cells with elevated [Ca2+] had reduced gKATP. For increased k’1/2 (Fig 3F), cells with elevated [Ca2+] had reduced gKATP and increased kglyc. Thus in the absence of coupling, cells that remained active had increased excitability, determined by KATP density or function, and metabolic activity.
Other parameters characterizing changes to KATP ATP-inhibition kinetics are also sometimes reported for KCNJ11 and ABCC8 mutations. The open channel probability in the absence of ATP (p’0) increases for a number of NDM mutations. We characterized how changes in both p’0 and k’1/2 act in combination, in the presence and absence of electrical coupling (Fig 4A and 4B). Increases in p’0 alone had little effect for any level of coupling (Fig 4C). Upon higher k’1/2 values, increases in p’0 led to a sharp transition to [Ca2+] suppression with normal electrical coupling, a shift in this transition with reduced electrical coupling, and a gradual decline in activity even at high p’0 values in the absence of coupling (Fig 4D). At high k’1/2 values where only a few cells are active in the absence of coupling, a decrease in p’0 elevated [Ca2+], but only in the absence of electrical coupling (Fig 4E and 4F). A change in the steepness of ATP inhibition (hill coefficient, H) is sometimes reported. Increasing H did not affect the dependence of the transition on electrical coupling (Fig 4G and 4H), although in each case the transition occurred at higher k’1/2 values. Increasing H lowers poK(ATP) within the range of cellular ATP levels, which are substantially higher than the half maximal ATP concentration for KATP inhibition.
These results demonstrate that similar behavior occurs irrespective of the way in which KATP over-activity is increased in the islet (S2 Fig): a sharp decline in activity in the presence of electrical coupling; but a shifted and/or more gradual decline in the absence of electrical coupling.
We next simulated specific KCNJ11 and ABCC8 mutations that cause NDM, and correlated the suppression of [Ca2+] with the clinical severity of the mutation (type2 diabetes; transient or permanent NDM; DEND characteristics) that indicates the level of islet dysfunction associated with each mutation. In each mutation we further tested whether a recovery in [Ca2+] could be achieved by eliminating electrical coupling. All mutations in which ATP-sensitivity is characterized (Eq 7) report shifts in k’1/2. p’0 when reported is also elevated, and occasionally α is characterized and reported as non-zero (S1 Table). However p’0 and α are often not reported, or are assumed. We first examined a set of KCNJ11 and ABCC8 mutations, as characterized by reported k’1/2 and α values for channels composed of mixed mutant and wild-type Kir6.2 or SUR1 subunits. A review of NDM mutations has reported KATP currents at elevated, but not saturating ATP levels [57]. Given reported k’1/2 values, we then estimated a residual open probability (αest, S1 Table). Therefore we also examined a set of KCNJ11 and ABCC8 mutations as characterized by reported k’1/2 and estimated αest values.
When including reported k’1/2 and reported α values, simulated islets showed reduced [Ca2+] in the presence of electrical coupling (Fig 5A and 5B, S3 Fig). The majority of mutations showed all cells with [Ca2+] elevations (Fig 5A and S3 Fig), but reduced [Ca2+] duty cycle (Fig 5B and S3 Fig). The [Ca2+] duty cycle was reduced to a greater level in all NDM mutations compared to the type2 diabetes mutation, owing to the greater k’1/2. However, in the majority of mutations the [Ca2+] duty cycle was still >50% of that in control islets. Complete suppression of [Ca2+] was observed for only a subset of mutations across all clinical characterizations of NDM. In the absence of electrical coupling, recovery of [Ca2+] was observed in the majority of these cases. This included sulfonylurea-insensitive mutations, such as I296L that had no reported residual current (α).
When αest values were included, most NDM mutations showed complete suppression of [Ca2+] (Fig 5C and 5D, S4 Fig), albeit with some exceptions (e.g. DEND mutation V64L). In the absence of coupling, many mutations showed a partial recovery in [Ca2+]. This included the majority of transient NDM mutations, but also several permanent NDM mutations (E229G, H46Y, F35V) and those with DEND features (F132L) (Fig 5C and S4 Fig). Thus accounting for a residual open probability (α) is important to describe the action of KCNJ11 and ABCC8 mutations; and in many cases a recovery in islet function is predicted following a modulation of gap junction electrical coupling.
We next accounted for the stochastic nature of ion channels by first simulating islets with normal KATP function and testing the effects of stochastic channel noise at basal and stimulatory glucose. In the presence of electrical coupling, including noise resulted in no change in [Ca2+] at basal and elevated glucose (Fig 6A and 6B). In the absence of electrical coupling, including noise resulted in more cells showing [Ca2+] elevations at basal glucose (Fig 6C). Minor increases in cells showing [Ca2+] elevations were also observed at elevated glucose (Fig 6D), with small changes in [Ca2+] oscillation shape and duty cycle. At ~8% electrical coupling including noise also resulted in more cells showing [Ca2+] elevations at basal glucose (Fig 6E). [Ca2+] elevations at basal glucose were characterized by short transient peaks (Fig 6C), similar to those experimentally observed in Kir6.2[ΔN30,K185Q] positive cells lacking gap junction coupling [33]. Transient KATP closures were observed in the presence and absence of coupling, but only in the absence of electrical coupling did these coincide with significant depolarization and [Ca2+] elevations (Fig 6F). Thus stochastic noise leads to increases in [Ca2+] under basal glucose, upon reduced or absent electrical coupling.
We next examined the effect of stochastic noise in islets with altered KATP ATP-inhibition kinetics, focusing on increased k’1/2 and α values. In the presence of electrical coupling noise did not change how [Ca2+] depended on increasing k’1/2 (Fig 7A). However in the absence of electrical coupling noise enhanced [Ca2+] elevations at increasing k’1/2 (Fig 7B) and increasing α (Fig 7C), and elevated [Ca2+] up to higher k’1/2 or α values. We then asked if recoveries in [Ca2+] would be enhanced for the KCNJ11 and ABCC8 mutations we previously examined. For mutations described by reported k’1/2 and reported α where a recovery was predicted, in the presence of electrical coupling [Ca2+] remained fully suppressed when noise was included (Fig 7D). In the absence of electrical coupling [Ca2+] elevations were much greater when noise was included (Fig 7E). These elevations were again characterized by frequent, low duty cycle [Ca2+] transients (Fig 7F). Similarly, for those mutations described by reported k’1/2 and estimated α where a recovery was predicted, in the absence of electrical coupling [Ca2+] elevations were much greater when noise was included (S5 Fig). Thus after accounting for stochastic noise a much greater recovery in islet function is predicted following a reduction in gap junction coupling.
To predict how a reduction in gap junction coupling may recover insulin secretion in the presence of KCNJ11 or ABCC8 mutations, we included an insulin-release component to the islet model [58]. In control islets a step elevation of glucose or [Ca2+] generated biphasic secretion dynamics (S6 Fig) consistent with experimental observations [59–61]. To verify the islet model can accurately predict gap junction recovery of insulin release (Fig 2B) we first simulated islets with increasing Kir6.2[ΔN30,K185Q] expression. Time-averaged insulin secretion was similar for control islets in the presence and absence of electrical coupling (Fig 8A). Upon 10% over-active KATP channels there was a small decline in insulin release in each case, restricted to the first phase, t = 0-5min (Fig 8B). Upon 50% over-active KATP channels there was a substantial decline in insulin release in the presence of electrical coupling, close to residual levels. However, in the absence of electrical coupling the decline was less, consistent with observations here (Fig 2B and S1 Fig) and published elsewhere [33].
We next examined if a reduction in electrical coupling recovered insulin secretion in the presence of the KCNJ11 and ABCC8 mutations that we predicted would show [Ca2+] recovery (Figs 5 and 7, S5 Fig). When considering mutations described by reported k’1/2 and reported α (if any), those mutations that showed recoveries in [Ca2+] (Figs 5A, 7B, 7D and 7E) also showed recoveries in insulin secretion (Fig 8C). In the presence of electrical coupling the suppression in insulin secretion was much greater in the first phase (t = 0-5min) than the second phase (t>5min) (S7 Fig). The recovery of insulin secretion in the absence of coupling was also greater in the second phase (Fig 8D and S7 Fig), primarily due to the slow [Ca2+] elevation limiting significant [Ca2+] and insulin secretion in the first phase (S7 Fig). Stochastic noise had a greater effect in promoting recovery for more severe mutations, e.g. I296L (Fig 8E and 8F). When considering mutations described by both reported k’1/2 and estimated α, those mutations that showed recoveries in [Ca2+] (Fig 5C and 5D, S5 Fig) also showed recoveries in insulin secretion (Fig 8G). A sub-set of these mutations also showed recoveries in [Ca2+] and insulin secretion at ~8% electrical coupling (Fig 8H and S5 Fig).
Sulfonylureas can achieve near-full inhibition of most mutant KATP channels, albeit with complex inhibition kinetics. By reducing p’0 we modeled the effect of sulfonylurea action and applied this to KCNJ11 and ABCC8 mutations described by reported k’1/2 and estimated α that showed little or no [Ca2+] recovery (S8 Fig). In the presence of electrical coupling no increase in [Ca2+] occurred upon p’0 reduction for all mutations examined (Fig 8I). In the absence of electrical coupling a significant recovery in [Ca2+] (Fig 8I) and insulin secretion (Fig 8J) occurred upon p’0 reduction for several of the mutations examined. Thus further recoveries are predicted for severe NDM mutations, with decreased or absent gap junction coupling upon sulfonylurea therapy.
The islet of Langerhans shows global coordinated behavior as a result of electrical coupling between β-cells [9,10]. β-cells are intrinsically heterogeneous in their function and previous work demonstrated how a threshold number of inactive cells can suppress the activity in cells that otherwise would be active, mediated by electrical coupling [27]. Factors that reduce β-cell excitability through increasing KATP channel activity will lead to there being a greater number of inactive β-cells in the islet and thus a greater likelihood of overall islet suppression. Thus reducing electrical coupling may prevent this suppression by limiting it to just those inactive cells. In this study we used experimental and computational methods to examine how electrical coupling between β-cells can exacerbate islet dysfunction following changes to KATP channel activity due to KCNJ11 and ABCC8 mutations that cause NDM; and that modulating electrical coupling may recover function.
We previously predicted with a mathematical model how the sharp transition to global suppression that occurs upon small increases in the number of inexcitable cells would become more gradual upon reduced electrical coupling [27]. Here we experimentally demonstrate this with good quantitative agreement (Fig 1): With ~45% gap junction coupling a less sharp and right-shifted transition is observed, and in the absence of gap junction coupling there is a near-linear relationship where only excitable cells are active. With high proportions of inexcitable cells there is a progressive increase in electrical activity with reduced coupling. This increase arises from the remaining excitable cells being no longer suppressed via gap junction coupling. However, as we discuss below stochastic noise also plays a role in the absence of electrical coupling. We also show that these model predictions and experimental results extend to insulin secretion, which demonstrates how this behavior is physiologically important.
Through computational modeling we find similar sharp transitions to global suppression when parameters that describe the kinetics of KATP channel inhibition are uniformly changed across the islet (Figs 3 and 4). Previously similar behavior was shown computationally and experimentally upon uniform KATP activation with diazoxide [27]. These results support that irrespective of how inexcitability arises, electrical coupling will mediate the suppression of global activity; and that gradual reductions in gap junction coupling will increase activity. Furthermore the point at which the rapid transition to suppression occurs is consistent: For Pmut (% cells inactive), α (residual current), k’1/2 (ATP concentration at 50% KATP inhibition), p’0 (open channel probability), the transition in the presence of coupling occurs where 20–40% of cells are inactive in the absence of coupling.
We do find a greater variability in the transition position when simulating intermediate levels of electrical coupling compared to normal levels of coupling (Fig 3H). Further, in experimental measurements upon higher levels of Kir6.2[ΔN30,K185Q] expression there was more variability between Cx36+/- animals compared to Cx36+/+ animals (Fig 1D). This variability between islets may explain the less sharp transition we observed experimentally compared to simulations (Fig 1C–1E). This variability in the transition is also similar to the substantial variability in oscillation synchronization at stimulatory glucose previously observed experimentally in islets with ~50% gap junction coupling [8]. We speculate this variability in the transition depends on the link between heterogeneity in electrical coupling and cellular excitability, and whether the two correlate or not. Interestingly we observed further shifts in the transition at lower (<25%) levels of electrical coupling (Fig 3F), suggesting that a few remaining excitable cells can still ‘recruit’ other cells to become active at these low levels of coupling. Examining the role of cellular heterogeneity and understanding the divergence in islet function at lower levels of electrical coupling will be a goal for future work.
Studies here have only considered mouse islets. However there are some differences between mouse and human islets that need considering to fully translate these findings to human islet function. The architecture of human islets is different from that of mouse islets [62–64], and the cellular regulation of excitability has some differences [65]. Despite these differences the level of electrical coupling between β-cells is similar to that in mouse islets [66,67], and mediated by the same gap junction connexins [7]. In 3D β-cell structures there is also little size-dependence to suppression [31]. Human islets have been suggested to be made up of folded ‘sheets’ of β-cells [64], but large 2D structures show similar suppression to 3D structures [31]. Thus we predict similar behavior would be observed in human islets. This prediction is supported by consistency between simulations of KCNJ11 and ABCC8 mutations (with αest) and clinical diabetes severity (discussed below). However, determining whether phase transitions occur upon increasing KATP over-activity in human islets and their dependence on gap junction coupling, as well as testing whether a human β-cell based islet model can recapitulate these findings is needed.
Connectivity is crucial for dictating the overall dynamics of the islet and we predict that the concepts found in the islet can be applied to other systems relying on signaling by means of gap junction channels. For example, in brain injuries, gap junction coupling can disrupt [Ca2+] dynamics and increase cell death in interconnected cells [68]. The absence of gap junctions in these situations does not lead to [Ca2+] dysregulation and reduces cell death [68]. This is similar to results presented here, where gap junction removal specifically alleviates the consequences of diabetes causing mutations. In another example, pulsatile release of GnRH from remodeled GnRH neurons initiates puberty [69,70]. GnRH neurons depolarize in response to GABA before maturation, but switch during maturation to a hyperpolarizing response. This switch depends on the opening of the GABA channel [28] and is analogous to KATP channel closure observed in the β cell. We therefore speculate electrical coupling is important for this sharp transition, and therefore subtle changes in channel kinetics induces critical-like behavior in coupled systems.
A major goal of this study was to predict the role gap junction coupling plays in the presence of KCNJ11 and ABCC8 mutations that cause NDM. Our initial computational model results show good agreement with experimental results presented here and published elsewhere regarding the suppression of both [Ca2+] and insulin upon elevated Kir6.2[ΔN30,K185Q] expression, and its recovery upon reduced gap junction coupling (Figs 1C–1E, 2A–2C, and 8A)[27,33]. There is also good agreement between computational model results and published experimental measurements of diazoxide-induced suppression upon reduced gap junction coupling (Fig 3B) [27]. This validation of the model gives us confidence that robust predictions can be made for the role of gap junction coupling upon altered KATP activity caused by KCNJ11 and ABCC8 mutations. A summary of predictions over all conditions tested is shown in S2 Table.
Including published kinetics of ATP-inhibition for KCNJ11 and ABCC8 mutations in the model showed disruptions to [Ca2+] and insulin secretion (Figs 5 and 8). k'1/2 was important to describe this disruption and is widely reported. However few studies reported the residual current α, and [Ca2+] was highly dependent on this parameter (Fig 3F). Indeed without inclusion of αest the decrease in [Ca2+] was minor and similar across all clinical phenotypes (Fig 5A and 5B), therefore is unlikely to be a full description of islet dysfunction in NDM. With estimates for α included [Ca2+] was generally suppressed across NDM mutations irrespective of clinical characteristics, but not for type2 diabetes mutations (Fig 5C and 5D). This also indicates the importance for characterizing and reporting the residual ATP-independent current when interpreting channel mutation characteristics. While the hill factor was also important for [Ca2+] suppression, it did not correlate with NDM and is not often reported. Despite the good agreement with estimates for α included, there were some NDM exceptions where altered ATP inhibition was insufficient to generate a full suppression of [Ca2+] and insulin that would cause diabetes. KCNJ11 mutations that alter channel trafficking can cause hyperinsulinism [71], and trafficking alterations may be involved in NDM mutations. Altered expression of Kir6.2 and SUR1 subunits has also been observed for some NDM mutations, and may also contribute to altered islet function [72]. In recombinant systems where mutant channels are characterized, expression and trafficking can be variable [73]. Therefore expression of mixed mutant and WT channel subunits may not fully replicate in-vivo conditions, where increased expression of the mutant subunit may lead to increased KATP over-activity than in recombinant systems.
For several KCNJ11 and ABCC8 mutations we predicted significant recovery in [Ca2+] and insulin secretion with an elimination of electrical coupling (Fig 5). Given the agreement between predicted [Ca2+], insulin and clinical phenotype (Figs 5C, 5D and 8G), together with the model validation (Figs 1, 2 and 8A), we can have confidence that a reduction in gap junction coupling would be effective in recovering islet function for these KCNJ11 and ABCC8 mutations. The agreement between simulations of Kir6.2[ΔN30,K185Q] expression, and α or k’1/2 increases (Fig 3); and experimentally measured recovery in insulin secretion and glucose homeostasis upon Kir6.2[ΔN30,K185Q] expression with reduced gap junction coupling (Fig 2C and 2D), further supports these predictions. Recoveries were observed in most transient-NDM mutations, several permanent-NDM mutations, and few DEND mutations; although the latter heavily depends on whether αest is included. For instance the sulfonylurea-insensitive I296L mutation showed no residual current in its initial characterization [21], yet based on other studies we estimate a significant residual current. The recovery in this and other DEND cases may be under- or over-estimated without determining whether there is a significant residual current. Nevertheless, these results indicate the importance of gap junction coupling in mediating the suppression of [Ca2+] and insulin secretion in NDM, and suggest reducing gap junction coupling may partially recover islet function and blunt NDM.
Most mutations required a complete reduction in electrical coupling to show recoveries in [Ca2+] and insulin secretion, although some mutations showed recovery with ~90% reduction. Achieving such large reductions in islet gap junction coupling has not been demonstrated using specific inhibitors, although strategies exist for robustly inhibiting gap junctions formed from other connexins [74]. Furthermore, reducing gap junction coupling will abolish coordinated dynamics of insulin release [60]. Despite increased [Ca2+] and insulin secretion, lack of pulsatility will likely reduce insulin action [75]. The absence of substantial first phase secretion recovery (S7 Fig) may have a similar impact. However chronic sulfonylurea delivery causes glucose intolerance [76], therefore any defect resulting from reduced gap junction coupling may not be more disruptive than long term sulfonylurea therapy.
We also did not include ‘amplifying’ mechanisms of insulin secretion or incretin (GLP1, GIP) hormone action, which would elevate Ca2+-triggered second-phase insulin secretion [77]. Incretin regulation of electrical activity and [Ca2+] has been simulated [78]; but amplifying mechanisms are poorly understood [61,79]. We predict for mutations where a minor second phase recovery occurs that a greater recovery will occur if we consider these mechanisms.
We predicted further recovery would occur in some permanent-NDM and DEND mutations with a combination of gap junction reduction and sulfonylurea treatment. However, for some mutations where no recovery was predicted, effective sulfonylurea therapy does occur in patients. We did not model sulfonylurea inhibition of the residual current α, which can explain this discrepancy. The kinetics of sulfonylurea inhibition of ATP-insensitive KATP channels is complex and not fully understood [80,81]. For example in-vitro studies have shown that the ‘sulfonylurea-insensitive’ I296L mutation is inhibited >50% under conditions of ~95% WT KATP inhibition. Whether the remaining current is ATP sensitive is unknown and will likely determine whether gap junction mediated recovery is possible: here we model it to be ATP-insensitive and thus present the minimal effectiveness expected. Therefore, while applicable to a rare patient population we predict that gap junction reduction may assist in sulfonylurea therapy for some insensitive or partially sensitive NDM mutations.
Another key finding is the role of stochastic noise to elevate electrical activity, [Ca2+] and insulin secretion in electrically isolated cells. While stochastic noise affects β-cell [Ca2+] oscillations [43,44], its role in basal [Ca2+] has not been examined significantly. We demonstrated noise can affect [Ca2+] levels (Fig 6). Including stochastic noise increased [Ca2+] only in the absence of electrical coupling at basal glucose, and upon increased KATP activity due to increased α and k’1/2. This suggests its broad importance in increasing activity in cells that lack electrical coupling. While we predict that stochastic noise will improve the recovery for several mutations, we note that for normal islets noise will be detrimental by inappropriately elevating [Ca2+] under basal glucose conditions following reduced or absent gap junction coupling.
We only included stochastic noise in the KATP channel, yet inclusion of noise from other channels will likely increase [Ca2+] further. The level of noise in a single β-cell is also not well characterized and depends on the number of channels. We estimated noise based on ~420 channels in a β-cell, which is consistent with reported channel populations [21,82]. The [Ca2+] predicted in normal islets in the absence of electrical coupling also matches experimental results [9,10]. However precisely examining noise characteristics in β-cells and their link to channel numbers will be important to accurately estimate [Ca2+] increases, especially in the presence of altered KATP kinetics.
We applied computational and experimental approaches to examine the role of gap junction coupling in islet dysfunction caused by KCNJ11 and ABCC8 mutations. Gap junction electrical coupling strongly mediates the suppression of islet [Ca2+] and insulin secretion in the presence of ATP-insensitive KATP channels, and significant recovery in [Ca2+] can be achieved upon reduced electrical coupling. Following experimental validation of our computational model, we made firm predictions that such recovery could be achieved for many KCNJ11 and ABCC8 mutations that cause NDM. Increased [Ca2+] occurs through reducing the suppression of more excitable cells in the islet, allowing them to regain excitability. However stochastic noise also likely plays a role in elevating [Ca2+]. These results gain further insight into how the dysfunction to the islet of Langerhans can occur in disease, and suggests potential for therapeutic treatments for NDM where sulfonylureas are ineffective. Further the principles discovered that govern how connectivity of excitable and inexcitable β-cells dictate overall function in the islet may be applicable to other multicellular systems.
All experiments were performed in compliance with the relevant laws and institutional guidelines, and were approved by the University of Colorado Institutional Biosafety Committee (IBC) and Institutional Animal Care and Use Committee (IACUC).
The Kir6.2 subunit mutation with GFP tag (Rosa26-Kir6.2[∆N30,K185Q]); β-cell specific, inducible Cre (Pdx-CreER); and Connexin36 knockout (Cx36-/-) mouse models have been previously described [45,83,84]. Pdx-CreER and Rosa26-Kir6.2[∆N30,K185Q] mice were crossed with Cx36 knockout mice to yield all combinations of mice studied. Daily injections of tamoxifen (1–5 doses, 50 mg/kg body weight per dose) administered IP in 8–16 week old mice induced the variable Kir6.2[∆N30,K185Q] expression. Mice lacking Pdx-CreER or Rosa26- Kir6.2[∆N30,K185Q] were used as 0% expression controls.
Blood glucose was measured daily after tamoxifen induction using a glucometer (Ascensia Contour, Bayer). Reported levels were averaged over days 27–29 after tamoxifen induction. Plasma insulin was measured at day 29 after tamoxifen induction from blood samples, centrifuged for 15 minutes at 13,900rpm, then assayed using mouse ultrasensitive insulin ELISA (Alpco).
Islets were isolated from mice under Ketamine/Xylazine anesthesia by collagenase injection through the pancreatic duct, and animals euthanized via exsanguination and cervical dislocation. Islets were handpicked after the pancreas was harvested and digested, and were maintained in RPMI medium at 11mM glucose plus 10% FBS, 100 U/ml penicillin, 100 μg/ml streptomycin, at 37°C under humidified 5% CO2 for 24–48 hours prior to study.
Islets (5/batch, duplicate) were incubated first at 2mM glucose in Krebs-Ringer Buffer (128.8mM NaCl, 5mM NaHCO3, 5.8mM KCl, 1.2mM KH2PO4, 2.5mM CaCl2, 1.2mM MgSO4, 10mM HEPES, 0.1% BSA, pH 7.4), and then for 60 minutes at 2mM or 20mM glucose. The medium was sampled for secretion, and islets sampled for content by lysing in 1% TritonX-100 and frozen overnight at -20C. Samples were assayed using mouse ultrasensitive ELISA.
Isolated islets were loaded with 3μM Rhod-2 (Invitrogen), in imaging medium (125mM NaCl, 5.7mM KCl, 2.5mM CaCl2, 1.2mM MgCl2, 10mM Hepes, 2mM glucose, and 0.1% BSA, pH 7.4) for 45 minutes at room temperature, and were held in polymdimethylsiloxane PDMS microfluidic devices [85] maintained at 37°C. Rhod-2 fluorescence was imaged on a spinning disk confocal microscope (Marianas, 3I), excited at 561nm using an OPSL sapphire laser, with a 580-655nm band-pass filter for emission; or imaged on a confocal microscope (LSM780, Zeiss) excited at 561nm using a diode-pumped solid-state laser, with a 570-645nm band-pass selection for emission. A small sub-set of isolated islets were loaded with 4μM FuraRed (Invitrogen) for 90 minutes at room temperature, and imaged on a spinning disk confocal microscope, excited at 488nm using a diode-pumped solid-state laser, with a 580-655nm band-pass filter for emission. GFP fluorescence was excited at 488nm using a Ar+ laser line (LSM780) or diode-pumped solid-state laser (Marianas), with a 495-555nm band-pass selection for emission. Images were acquired 1/sec, 10 minutes after elevating glucose concentration (2-20mM). Microscope settings (integration time, scan time, gain, laser power) were constant for all images collected within the same day.
The model is modified from that described previously [27], and based on the Cha-Noma β-cell model [86] with cell-cell coupling and altered KATP channel function. Model code and associated files are included as supporting files. All model code was written in C or C++ and run on the University of Colorado JANUS supercomputer.
The membrane potential (Vi) of each β-cell i is related to the total transmembrane current (Ii), which is composed of individual currents described in [86]; using parameters in S3 Table:
−CVi′=ICav+ITRPM+ISOC+IbNSC+IKDr+IKCa(SK)+IK(ATP)+INaK+INaCa+IPMCA
(1)
Gap junction coupling is modeled by assigning a coupling current between neighboring cells (i,j). A sphere packing algorithm was used to assemble cells within the cluster (mean number of cell-cell connections = 5.3) [27,87]
−CVi′=Ii+∑igcoupi,j(Vi−Vj)
(2)
Heterogeneity in coupling was included by randomly assigning gi,jcoup according to a distribution from previously published data with SD/mean = 70% [47]. Endogenous heterogeneity was modeled by randomizing all parameters indicated in S3 Table between cells about a mean value according to a Gaussian distribution with SD/mean as indicated.
The KATP channel current was described as:
IK(ATP)=gK(ATP)∙p0K(ATP)∙(V−VK)
(3)
where the open channel probability poK(ATP) is given by:
poK(ATP)=.08(1+2[ADP].01)+.89([ADP].01)2(1+[ADP].01)2(1+.45[ADP].026+([ATP].05))
(4)
Expression of Kir6.2[ΔN30,K185Q] in rodent islets was modeled as previously and in accordance with experimental data [24,27], by modifying the open probability poK(ATP) in a fraction (Pmut) of cells, according to:
poK(ATP)[ΔN30,K185Q]=γ(poK(ATP))+(1−γ)
(5)
where γ = 0.5 and Pmut increases with number of GFP+ Kir6.2[ΔN30,K185Q] -expressing cells. Simulations for Pmut were run with glucose elevated to 20mM.
Diazoxide application was modeled as previously [27] by modifying the open probability poK(ATP) in all cells according to:
poK(ATP)Diaz=α+(1−α)*poK(ATP)
(6)
where α is variable and increases with diazoxide treatment concentration. Simulations for α were run with glucose elevated to 11mM and Pmut = 100%.
KCNJ11 and ABCC8 mutations were modeled by modifying the open probability poK(ATP) in all cells according to:
poK(ATP)=(1−α)po′*0.08(1+2[ADP].01)+.89([ADP].01)2(1+[ADP].01)2(1+.45[ADP].026+([ATP]k1/2′*0.05)H)+α
(7)
where k’1/2 represents the relative increase in half maximal ATP concentration, and H is the Hill coefficient. Unless noted H = 1 and was not varied when modeling mutations in order to maintain consistency. p’0 represents the relative increase/decrease in open channel conductance. α represents the fraction of current remaining at saturating ATP concentrations, equivalent to α in Eq 6; and was included if reported otherwise set to zero. To generate estimated α values (αest) for a mutation, data reporting the fractional KATP current remaining at 3mM ATP [57] was compared to the poK(ATP) at 3mM ATP when including the reported k’1/2. The difference between these values represents the estimated ATP-independent residual current (αest); where negative values were set to zero. Sources and parameter values for each mutation examined are summarized in S1 Table. All mutations were modeled from data reporting mixed mutant and wildtype subunit expression in heterologous expression systems. Simulations were run with glucose elevated to 11mM and Pmut = 100%.
Stochastic noise was applied to the KATP current by including a time varying noise component, as previously reported [44]. Eq 3 was adjusted:
IKATP=gkATP∙[pok(ATP)∙(1+S)]∙(VM−Vk)
(8)
where S represents the time varying noise component which fluctuates with mean≈0 and standard deviation≈0.049. The time varying noise component S was modeled as:
S′=−Sτ−(Sτ)+ξ
(9)
where τ = 500ms, and ξ represents a noise factor generated according to a random number sequence, such that S follows a normal distribution. The standard deviation of ξ was adjusted such that the standard deviation of S(t) (Sσ = 1/√NK(ATP)) was equivalent to ~420 channels per cell [44]. Similar noise was observed over different simulations (S9 Fig).
The general form of this component was adapted from a previously published insulin secretion model [58]. Insulin granules are designated in distinct pools, with rates of exchange leading up to a secretion event. The granule fusion step was modeled with [Ca2+] dependence to give a first phase release of ~20 granules/min per β-cell, and second phase rate of ~5 granules/min per β-cell [58,61]. Other rates were adjusted from those previously reported [58] to account for simplifications made to incorporate this component with our model. Specific rates, initial conditions and other parameters can be found in S4 Table.
Custom MATLAB scripts were used to analyze all images acquired from calcium imaging [31]. First, images were smoothed using a 5x5 average filter. To calculate the fraction of cells showing elevated [Ca2+], a quiescent reference cell was selected manually from an area where no significant intensity fluctuations occurred over the duration of the experiment. The variance for each pixel time-course of the image was calculated to examine which pixels displayed the greatest fluctuations in time. The variance of the manually selected quiescent reference cell was used to generate a threshold to compare intensity fluctuations of all other pixels: any pixels having a time-course variance > 2 standard deviations above that of the quiescent reference cell were counted as ‘active’. Time courses with significant motion artifacts were excluded from analysis, and photobleaching was handled by applying a linear fit. The percentage of cells active is calculated based on the number of pixels calculated to be ‘active’, normalized to the number of pixels in the area of the islet. To calculate GFP+ regions, the mean fluorescence intensity was calculated in GFP- control islets cells and used to generate an intensity threshold: any pixel having a GFP intensity greater than this threshold was considered GFP+. GFP+ area was expressed as a % of the total islet area. In islets where lack of nuclear GFP was observed, nuclear areas were manually included in the GFP+ area. For each level of Cx36 expression, the average activity over all control islets was used to normalize the maximum activity of simulation fits.
All analysis was performed with custom MATLAB routines on the output for each simulation. For all time courses, the first ~200 time points were excluded. To calculate the fraction of cells showing elevated [Ca2+], a threshold of 0.165μM was applied, and cells showing [Ca2+] fluctuations that exceeded this threshold were considered active. The [Ca2+] duty cycle for each cell was represented by the fraction of time the [Ca2+] time course exceeded this threshold. The duty cycle was reported as the mean duty cycle over all cells in the islet, where a silent cell has a duty cycle of 0. The same analysis methods were used when noise was added to the model.
The total number of granules secreted was calculated according to Eq 17 over 30 minutes of the simulation time course. First phase secretion was calculated according to Eq 17 over the first 5 minutes after insulin elevated. Second phase secretion was calculated from t = 5 min to t = 30 minutes after insulin elevated. All values were normalized to those of control islet simulations ran using the same heterogeneity and coupling distributions.
Student’s t-test was utilized to test for significant differences between simulation results; [Ca2+] imaging results; or insulin secretion, plasma insulin and blood glucose results. Linear regression between [Ca2+] imaging results, insulin secretion, plasma insulin and blood glucose against % GFP was grouped across Cx36+/+,Cx36+/-,Cx36-/- conditions. All statistical analysis was performed in Prism (Graphpad).
|
10.1371/journal.pgen.1004362 | Meiotic Drive Impacts Expression and Evolution of X-Linked Genes in Stalk-Eyed Flies | Although sex chromosome meiotic drive has been observed in a variety of species for over 50 years, the genes causing drive are only known in a few cases, and none of these cases cause distorted sex-ratios in nature. In stalk-eyed flies (Teleopsis dalmanni), driving X chromosomes are commonly found at frequencies approaching 30% in the wild, but the genetic basis of drive has remained elusive due to reduced recombination between driving and non-driving X chromosomes. Here, we used RNAseq to identify transcripts that are differentially expressed between males carrying either a driving X (XSR) or a standard X chromosome (XST), and found hundreds of these, the majority of which are X-linked. Drive-associated transcripts show increased levels of sequence divergence (dN/dS) compared to a control set, and are predominantly expressed either in testes or in the gonads of both sexes. Finally, we confirmed that XSR and XST are highly divergent by estimating sequence differentiation between the RNAseq pools. We found that X-linked transcripts were often strongly differentiated (whereas most autosomal transcripts were not), supporting the presence of a relatively large region of recombination suppression on XSR presumably caused by one or more inversions. We have identified a group of genes that are good candidates for further study into the causes and consequences of sex-chromosome drive, and demonstrated that meiotic drive has had a profound effect on sequence evolution and gene expression of X-linked genes in this species.
| Sex chromosome meiotic drive causes changes in the sex-ratios of natural populations, and may even lead to extinction if the driving element reaches high frequency. However, very little is known about the genes that cause sex-ratio drive, and no causal gene has been identified in a species that consistently exhibits distorted sex ratios in natural populations. Several species of stalk-eyed flies in southeast Asia – genus Teleopsis – express X chromosome drive, but the genes underlying drive have been difficult to locate due to reduced recombination between drive and standard X chromosomes presumably caused by the presence of a large inversion. Here, we use high throughput RNA sequencing to identify over 500 transcripts that are differentially expressed in the testes due to the effects of a driving X chromosome (XSR) in T. dalmanni. Most of these are X-linked, evolve more rapidly than control genes, and exhibit elevated expression in the gonads. Finally, XSR has become genetically differentiated from standard X chromosomes – using the RNA sequence data, we found nearly 1000 sites in X-linked transcripts and only a handful in autosomal transcripts where there was a fixed nucleotide difference. We conclude that XSR has led to widespread sequence and expression divergence on the X chromosome in T. dalmanni.
| Meiosis typically results in an equal transmission probability of each allele from parent to gamete. This seemingly cooperative outcome masks an inherent genetic conflict. Alleles on any one chromosome would increase in frequency more rapidly if that chromosome passed to all, instead of half, of the gametes produced. Such selfish alleles cause meiotic drive and would be expected to sweep quickly to high frequency or even fix. Detecting autosomal drive is difficult because distorted segregation patterns of chromosomal markers must be observed. However, when a drive allele is on a sex chromosome, the sex ratio of offspring is distorted. As a driving sex chromosome increases in frequency, the sex ratio in the population will become increasingly biased. If a drive allele nears fixation, population extinction due to absence of the rare sex is expected [1]. Alternatively, because the rare sex will have a fitness advantage [2], alleles which act to restore the sex ratio to equality will be favored. Potential mechanisms to counter fixation of drive alleles include sexual selection in which standard males outcompete drive males in mating or sperm competition [3]–[5], selection acting on female XSR carriers [6], and the evolution of loci on the other sex chromosome or autosomes that suppress drive [1].
Genomic conflicts in general, and meiotic drive in particular, can create dynamic evolutionary systems that influence patterns of molecular evolution and the evolution of gene expression. Drive loci have a strong local fitness advantage, but decrease fitness of the population because selection cannot act efficiently to remove low-fitness drive carriers [7]. In addition, many examples of suppressed or “cryptic” drive systems have been uncovered in Drosophila in which either autosomal or Y-linked suppressors mask the phenotypic expression of the drive allele in extant populations [8]–[13]. Like the drive locus, at the time they arose, these loci would have been strongly selected, whether or not they provided any benefit to the organism [14]. Furthermore, the inherent fitness advantage of drive and suppressor alleles is expected to lead to strong effects on linked neutral polymorphism as these alleles increase in frequency - as has been documented for both autosomal [15], [16] and sex-ratio drive [17] in Drosophila species. In fact, the theoretical effects of meiotic drive on the genome are so extreme that it has been invoked as a possible cause for fundamental phenomena [18], such as homologous recombination [19] and Haldane's Rule [20]–[22], with some experimental evidence of the latter [23], [24].
Although meiotic drive has been observed in many different species, particularly dipterans (reviewed in [14]), the genetic basis of drive is known in only a few cases, and of these, none distort sex ratios appreciably in natural populations. Partly, this is due to the tendency of actively driving loci to be found on sex-ratio X chromosomes that do not recombine with standard X chromosomes due to the presence of one or more inversions. This has prevented fine mapping of the drive loci in most cases [25], [26]. Intriguingly, both cases of X-chromosome drive that have been mapped to the gene level are associated with copy number variants and occur in Drosophila simulans. The “Paris” sex-ratio drive system (also known as XSR6) recombines freely and populations polymorphic for both the suppressor and the driver exist [8], allowing genetic dissection using interpopulation crosses. The element that causes drive has been mapped to a segmental duplication of six genes on the X chromosome, with associated changes in gene expression for some of the duplicated genes [27]. The “Winters” drive system is caused by an X chromosome drive gene, Dox, which is an imperfect duplication of a previously existing gene, MDox, and is suppressed by an autosomal retroduplicated gene, Nmy, which functions to silence Dox through an RNAi-like mechanism [10], [11].
In the stalk-eyed fly, Teleopsis dalmanni, males carrying a meiotic drive X chromosome (XSR-1 or XSR-2) [28] parent mostly daughters [29]. Drive chromosomes are present in natural populations but appear not to recombine with standard X chromosomes in laboratory crosses [26]. The X chromosome gene content in T. dalmanni is mostly orthologous to Muller element B, i.e. chromosome arm 2L in Drosophila melanogaster [30]. Thus, the genetic context for X chromosome drive in Teleopsis is distinct from that found in the Drosophila systems described above. Furthermore, meiotic drive associates with a number of characters that influence male reproductive success, including eye-stalk length [26], [31], sperm precedence [32] and sperm morphology [33]. The fate of the drive allele may be influenced by sexual selection acting against the drive X chromosome [34], which also causes males to have shorter than average relative eye-stalk length [26], [35]. Conversely, females carrying a drive X chromosome may have elevated fecundity, providing a possible explanation for why drive X chromosomes are not lost or suppressed [28]. However, the genetic basis for most of these traits - and meiotic drive itself - is unknown apart from the association with XSR. Like many chromosomes carrying meiotic drive loci, XSR – or at least the portion of XSR that causes both meiotic drive and associated phenotypic differences – does not recombine with standard X chromosomes [26], making identification of causal loci difficult.
To identify genes that are involved in sex-ratio drive and associated phenotypes in T. dalmanni, we performed RNAseq on replicate pools of testes carrying meiotic drive (XSR) and standard (XST) X chromosomes. We aligned these reads to a transcriptome assembled de novo, identified hundreds of transcripts differentially expressed between XSR and XST testes, identified their expression patterns, determined whether they were X-linked, Y-linked or autosomal, and identified fixed differences between the two samples. We found that drive-associated transcripts were more likely to be X-linked and to have elevated expression in testes (as expected) as well as in both testes and ovaries. These transcripts were also more rapidly evolving than a control set and included a number of interesting candidate genes with Drosophila orthologs involved in potentially relevant molecular and biological processes. Finally, we found that hundreds of X-linked transcripts carry fixed differences between XSR and XST samples while only a handful of such differences were found in autosomal transcripts. Our data support previous studies [36], [37] suggesting that the XSR haplotype is evolving independently from XST, and reveal a group of candidate genes that will be useful targets for future studies of meiotic drive in this species.
We sequenced RNA collected from replicate pools of testes dissected from T. dalmanni – Gombak males that carried the sex-ratio meiotic drive X (XSR) or the standard X (XST) chromosome (XSR and XST status was determined by microsatellite haplotype following Wright 2004 [36] and Wilkinson 2006 [28]). Reads were aligned to the T. dalmanni transcriptome (see methods) with bwa [38] and raw read counts were corrected using RSEM [39] to account for hits to multiple isoforms (contigs) making up the same transcript. We then used DESeq [40] with the corrected read counts to find transcripts that were differentially expressed between XST and XSR testes using a FDR<0.001 cutoff and after removing transcripts which had no expression in any of the four samples. We found a total of 513 transcripts to be significantly differentially expressed between transcriptomes from XSR and XST testes. As a group, we refer to these as “drive-associated transcripts”. Among them, 233 were expressed at a higher level in XSR males and 280 were expressed at a lower level in XSR males (Table 1). A total of 113 transcripts exhibited more than 10-fold differential expression between XSR and XST. For technical reasons, transcripts that are significantly differentially expressed are more likely than other genes to be expressed at a high level. In order to prevent weakly expressed transcripts from biasing our results, we defined a control gene set from among the remaining transcripts by removing the most weakly expressed genes from consideration (see Methods). We next aligned predicted proteins to the Drosophila proteome to identify putative Drosophila orthologs (Table 1). Among the drive-associated transcripts, 28.2% had putative Drosophila orthologs (18.4% among control genes). Of the remaining transcripts, 239 contain a long open reading frame, and may be Teleopsis-specific proteins, whereas 129 had short (<50 AA) open reading frames and may be noncoding RNA genes. Compared to drive-associated transcripts, a larger proportion of the control transcripts had short (<50 AA) open reading frames (52.3% vs. 25.1%, χ2 = 56.88, P<4.625 e-14, Table 1). Given that noncoding RNA genes are thought to be more narrowly and weakly expressed compared to protein-coding genes [41], [42], we speculate that an excess of presumptive noncoding RNA genes in the control gene set may be caused by the observation that drive-associated transcripts tend to be expressed more strongly than the average transcripts. Alternatively, protein-coding genes may be more likely to become drive-associated than noncoding RNA genes. Finally, we used quantitative RT-PCR to confirm differential expression of drive-associated transcripts (Table S1). After excluding weakly expressed samples, 11 of 11 transcripts replicated the qualitative pattern observed in the RNA-seq data (i.e. differentially expressed in the same direction).
We performed a multi-tissue expression analysis of drive-associated and control transcripts using RNAseq from six T. dalmanni tissues using tools provided on the trinity website (trinityrnaseq.sourceforge.net [43]). We clustered differentially expressed transcripts to identify the eight most common patterns of gene expression and compared the number of transcripts assigned to each cluster for drive-associated and control transcripts (Figure 1A, Figure S1). Control transcripts were more likely than drive-associated transcripts to have no significant pattern of differential expression (“Not differentially expressed”), possibly because many of them could be housekeeping genes. Testes-associated clusters were enriched among drive-associated transcripts compared to controls (χ2 = 737.3, P<2.2 e-16). We also assessed testes-specificity in drive and control transcripts by calculating the Tau metric [44] and found more drive-associated than control transcripts were testes-specific (57.5% vs 16.9%, χ2 = 622.7, P<2.2 e-16). This is not surprising considering the comparison was between testes from XSR and XST males. However, we also expect that a subset of drive-associated transcripts are likely to be involved directly in various aspects of spermatogenesis, given that meiotic drive affects sperm development in T. dalmanni [26] and a closely related species [45]. Among the other expression categories, a cluster showing elevated expression in the gonads of both sexes was also enriched (Figure 1B, χ2 = 30.5, P<2.2 e-16), raising the possibility that genes with pleiotropic effects on female reproduction may be differentially expressed on XSR. Early models of sex chromosome drive predicted that drive loci could be maintained if they also cause increased fitness in heterozygous females [6], [46]. Given that gonad expressed genes are often tissue specific, it has been thought unlikely that a single gene would do both, but given that an excess of drive associated genes show elevated expression in both ovary and testis, perhaps some of these genes are involved in both increased female fecundity (see [28]) and meiotic drive.
The presence of an XSR haplotype in a male T. dalmanni is sufficient to cause him to parent strongly female-biased broods, regardless of his genetic background [26]. We determined whether this strong X effect extended to the level of gene expression by comparing the chromosomal linkage of drive-associated transcripts and a control set using data from a comparative genomic hybridization experiment. We found that drive-associated transcripts were strongly enriched on the X chromosome compared to the control set (78% vs 18%, χ2 = 256, P<2.2 e-16), suggesting that the majority of downstream effects of XSR on gene expression are in cis rather than in trans (Figure 2). While the previous observation [30] that the D. melanogaster 2L, i.e. Muller element B, is orthologous to the T. dalmanni X generally holds (across all transcripts, 9.3% violate this rule), a large proportion of drive-associated transcripts (21.6%) have moved onto the X chromosome, in contrast to only 3.3% of controls (Figure 2). In D. melanogaster, male-specific genes have a tendency to move off of the X [47], [48], though young male-biased genes may be enriched on the X [49]. As the X chromosome in T. dalmanni is distinct from the D. melanogaster X, it is unclear whether the same pattern would be expected. While the number of moving drive-associated transcripts appears to be in large excess, drive-associated transcripts are more likely to be on the X chromosome than are controls, and much of the movement can be explained by the effect of linkage in that more genes are moving onto the X chromosome – relative to D. melanogaster - in T. dalmanni than are moving onto the autosomes (19.2% of controls and 29.7% of drive-associated transcripts have moved onto the X in T. dalmanni, relative to D. melanogaster). In addition, we recently found that in T. dalmanni, an excess of testes-specific transcripts have moved onto the X chromosome (unpublished data), and an excess of drive-associated transcripts are testes-specific. Indeed, among testes-specific transcripts, 21.1% of controls and 56.8% of drive-associated transcripts have moved onto the X chromosome from Muller elements other than B (Figure 2). Given these factors may be confounding, we fit nominal logistic models to predict gene movement by chromosome linkage (A or X), drive association (drive-associated/control), tissue source (testes or other) and interactions among these three factors for 7,150 transcripts. We compared three models with different interaction terms and chose the model with the lowest AICc score (Table S2, 4-parameter model). The best-fitting model explained 20.7% of the variation in gene movement (χ2 = 916, d.f. = 4, P<0.0001) with strong effects of X-linkage, tissue, and the interaction between X-linkage and tissue (all P<0.0001, Table S2) but no significant effect of drive-association (P = 0.1745). Therefore, while the large proportion of drive-associated transcripts moving onto the X is striking, this is most likely not due to the effect of drive per se. Instead, we conclude that most of the effect of XSR on expression is due to genes on the X chromosome, regardless of whether they moved there recently or have persisted on Muller element B since the divergence of genus Drosophila and Teleopsis. In addition, a group of five drive-associated transcripts was found to be Y-linked (Table 2). While the number of Y-linked genes does not exceed expectation, they are of interest as potential targets of sex-chromosome drive. During spermatogenesis in drive-carrying T. dalmanni, the Y-bearing sperm do not complete elongation. While the genetic cause of this is unknown, in other cases of X chromosome drive the Y chromosome is the direct target of drive. For example, in the Slx/Sly system in mice, expression of an array of Y-linked genes is modified by the presence of a driving X chromosome [50]. Currently, we have very little information about these Y-linked transcripts. They lack D. melanogaster orthologs, though two of the genes appear to be protein-coding and have orthologs in the sister species T. whitei.
Because only a small proportion of drive-associated transcripts had Drosophila orthologs, and because the taxa diverged ∼70 MYA [51], to assess protein sequence evolution we used T. whitei, a closely related species of stalk-eyed fly (∼1.8–3.5 MY since most recent common ancestor [52]) for comparison. RNA was extracted from testes collected from a T. whitei lab population derived from flies collected in Chiang Mai, Thailand, sequenced using Illumina Hi-Seq paired-end reads and assembled de novo using Trinity. Proteins were predicted from the T. whitei transcripts and aligned to T. dalmanni predicted proteins. Because a larger proportion of drive-associated than control transcripts are expressed in testes and have moved between the X and autosomes (see above), we used a generalized linear model with an exponential distribution and reciprocal link function to determine if drive association, expression (testis-specific or not), transcript movement, or chromosome location influence protein evolution (Table S3). In the best fitting model, significant factors positively affecting dN/dS included expression (P<0.0001), transcript movement (P = 0.0042), and drive-association (P = 0.0434). Testis-specific genes have elevated dN/dS compared to genes with expression in other tissues (median dN/dS = 0.338 vs 0.155, P<2.2e-16 Mann-Whitney U test). Genes inferred to have moved have lower dN/dS than genes that have not moved (median dN/dS = 0.125 vs 0.136, P = 0.004, Mann-Whitney U test). Drive-associated transcripts show higher dN/dS than controls (Figure 3A, median dN/dS = 0.307 vs 0.199, P = 3.8e-7, Mann-Whitney U test) and this holds true when only testis-specific genes are compared (Figure 3B, dN/dS = 0.379 vs 0.336 P = 0.0284 Mann-Whitney U test).
We conclude that (as expected [53]) testis-specificity influences much of the variation in dN/dS, but in addition, drive-associated transcripts are more likely to be evolving rapidly even after accounting for testis specificity. It is possible that a lack of constraint rather than positive selection is causing the increase in dN/dS among drive-associated transcripts, i.e. weakly deleterious alleles are expected to fix more rapidly when recombination is suppressed, as appears to be the case for large portions of XSR in T. dalmanni [37]. If the accumulation of deleterious alleles among X-linked genes due to a lack of recombination, i.e. Muller's ratchet [54], was the main cause of elevation in dN/dS, we would expect X-linked genes to have higher dN/dS than autosomal genes. However, X-linkage did not affect dN/dS in any model (Table S3). The most likely explanation is, therefore, that recent expression divergence in drive-associated transcripts coincides with divergence at the sequence level.
Recombination is suppressed between XSR and XST in T. dalmanni [26]. This is a common feature in several extant drive systems (see [25]) and may have evolved as a way to prevent recombination breaking up suites of genes that are beneficial to individuals carrying drive loci [14]. Recombination suppression leads to accumulation of genetic differences between the suppressed regions and is thought to be the primary mechanism leading to the degeneration of the Y chromosome [54], [55]. We hypothesized that it should therefore be possible to identify fixed genetic differences between the suppressed regions on XSR and XST using RNAseq data. Conversely, in a freely breeding population there should be very few fixed differences between autosomal genes in XSR and XST males. Indeed, we found 955 fixed differences in X-linked transcripts but only 11 fixed differences between XSR and XST males in autosomal transcripts (Table 2). Even more remarkably, roughly one-fourth of X-linked transcripts contain at least one fixed difference. Given the large number of individuals sampled (∼60 for each drive and standard individuals, see methods), this excess of X-linked fixed differences cannot be explained by the fact that the X chromosome was sampled at half the depth of the autosomes (see Figure S2). If the entire X chromosome were nonrecombining, a simple null expectation would be that fixed differences should be randomly distributed across the transcripts based only on their length. We performed a simulation to test this hypothesis. Based on the observed per-basepair frequency of fixed differences in X-linked (6.35 e-04) and autosomal transcripts (2.06 e-06) and the known lengths of all transcripts used in this study, we performed 10,000 draws from the binomial distribution to determine the expected number of genes carrying one or more fixed differences on each type of chromosome. We found more genes with no fixed differences (Figure 4), and more genes with six or more fixed differences (Figure 4, inset) than the X expectation. These data could be interpreted in one of two ways. First, this increased “clustering” of X-linked fixed differences could be due to repeated selection on multiple sites in certain transcripts. Indeed, theory predicts that genes modifying drive would be under positive selection after drive arose [56]. Alternatively, the excess of genes with no fixed differences could be due to free recombination on a relatively large portion of the X chromosome, either currently or historically. For example, if the drive X chromosome slowly accumulated multiple inversions in order to become fully non-recombining, then that could explain the presence of fewer fixed differences if some of the inversions are more recent than others.
Observed differences in transcription may be the direct result of genetic changes responsible for meiotic drive, or may impact other functions through linkage to the drive locus. While many drive-associated transcripts are expressed in the testis and hence may be directly involved in drive, others have higher expression in other tissues. To further understand what functions drive-associated transcripts might have, we first used the DAVID functional analysis tool [57] to determine whether drive-associated transcripts with Drosophila orthologs were enriched for any gene ontology (GO) terms. We found that at a 5% FDR cutoff, four ion binding GO terms (GO:0008270, GO:0043169, GO:0043167, and GO:0046872) were enriched among drive-associated transcripts (Table S4). None of these terms were enriched in the control gene set, despite the fact that the control genes are a much larger sample giving increased power. The genes in these GO categories were functionally diverse, and included a cytochrome P450, calmodulin, chiffon (an eggshell protein), and many others. In total, 37 drive-associated transcripts had at least one significant GO term (Table S4).
In T. dalmanni sperm bundles from drive males contain approximately 50% arrested sperm [33]. The molecular mechanism leading to arrest is not known, but inspection of spermatid bundles indicates that Y-bearing sperm fail to complete elongation in drive males. In one example of sex chromosome meiotic drive in Drosophila melanogaster the Y-sister chromatids fail to segregate during meiosis II, ultimately leading to arrest of Y sperm development prior to elongation [58]. It may also be the case that Y-bearing sperm undergo apoptosis or another form of regulated cell death. Among drive-associated transcripts, we found several with Drosophila orthologs involved in centrosome function, meiosis, mitosis, fertility, and apoptosis (Table S5). These genes may be causal to drive, or they may be misregulated due to the action of upstream drive genes. We also found several genes that are important to male and female fertility in Drosophila. Fs(1)N causes sterility in females when lost [59], Tom7 [60] and Hexo1 [61] are involved in sperm transfer and spermatogenesis respectively, and the loss of tj (traffic jam) causes sterility in both sexes [62]. Interestingly, one group of drive-associated transcripts are known to affect brain and eye development in D. melanogaster. Misregulation of these genes – if extended to development - could underlie some of the traits associated with drive [28], such as changes in behavior and eye span (Table S5). Previously, we identified a group of genes differentially expressed in T. dalmanni males selected for longer and shorter eyespan [63]. Two of these genes, chiffon, and CG4598 were also drive-associated and may be involved in the genetic link between shorter eyespan and meiotic drive [64]. Chiffon has a variety of functions, one of which is exon guidance in photoreceptors [65], CG4598 is a member of the Crotonase subfamily and is of unknown function.
Finally, genes that have differences in expression may be good candidates for the proximal causes of meiotic drive and associated phenotypes, but a heritable difference in sequence is required to trigger drive. As a first attempt to identify possible candidate genes, we identified a subset of X-linked, drive-associated transcripts that contained fixed differences between XSR and XST. We determined whether these fixed differences fell into the protein-coding regions or the UTRs of genes, and whether they were synonymous or nonsynonymous if protein-coding. We identified 24 drive-associated transcripts (of 46 drive-associated transcripts carrying fixed differences) that carried at least one nonsynonymous fixed difference between XST and XSR (Table 3). Many of these genes are also evolving fairly rapidly between T. dalmanni and T. whitei, with dN/dS values well above the average for all genes, though not necessarily due to positive selection (i.e. dN/dS is not >1). While most of these genes are testes specific (Tau is >0.95), six of 24 fall into the ovary and testis expression category, implying they could function in both male and female reproduction. A gene called klarsicht also contains two nonsynonymous fixed differences and reduced expression in XSR testes. This gene – a transport regulator - has been linked to a variety of functions including eye development [66]. Interestingly, it was recently discovered that klar mutants affect nonrandom segregation of sister chromatids in germline stem cells of the testis [67]. While klar mutants did not affect segregation of chromosome pairs, the association with nonrandom chromosome segregation is intriguing and worthy of future investigation.
Although distortion of sex ratio due to meiotic drive has been observed in a variety of species for over 50 years [46], [68]–[73], the genetic causes of sex chromosome drive remain obscure in the vast majority of cases. Sex chromosome meiotic drive is notoriously recalcitrant to traditional genetic dissection due to its tendency to associate with chromosomal inversions, presumably as a result of meiotic drive involving the combined action of multiple loci [4], [69], [74]. In addition, X-chromosome drive is predicted to have consequences for processes ranging from sexual selection to the evolution of the genome. As populations become biased towards one sex or the other, inter- and intra- sexual selective pressures diverge. As females become increasingly common and if male reproduction is at all costly, males may become choosy [75]. Meanwhile, females employing strategies that increase their chances of mating with a standard male would benefit, as more of their offspring would be the rare (male) sex [2]. This might occur through female preference for a linked trait [34] or multiple mating [76]. Meanwhile, sex-ratio meiotic drive is expected to favor Y-linked and autosomal alleles that suppress drive, subjecting the genome to strong local selection pressures. Fixation of alleles causing or modifying drive may be nonadaptive or even maladaptive.
To gain insight into the genetic differences between nonrecombining drive and standard X chromosomes, we used RNAseq to measure differences in expression between drive and standard testes from a species, T. dalmanni, with high frequencies of unsuppressed X chromosome meiotic drive and a wealth of biological data associated with the drive system. We sequenced testes from males carrying XSR and standard X chromosomes and identified a group of genes that are significantly differentially expressed, including a number of candidate genes whose D. melanogaster orthologs have been associated with male sterility and chromosomal nondisjunction during mitotic and meiotic divisions, and some of which carried fixed differences. While some of these genes may have diverged in expression due to neutral processes associated with sequence divergence of XSR from standard X chromosomes, others may either impact, or be impacted by meiotic drive directly. Interestingly, some of the genes whose expression changed in XSR males are also strongly expressed in other tissues and may be involved in other observed phenotypic differences between drive and non-drive males, including genes that may be involved in differences between drive and standard males in the sexually selected exaggerated eye stalk phenotype [64]. XSR males are generally at a reproductive disadvantage as they are less able to directly compete with other males for matings due to reduced ornament size [64] and for fertilizations after copulation due to weaker sperm competitive ability [32]. Conversely, heterozygous female carriers of XSR have higher fecundity than their XST sisters [28]. It has been suggested that the overdominant effect of XSR on female fecundity may be one reason why drive is still expressed in the population, rather than being suppressed as in many Drosophila spp. We found that drive-associated transcripts were enriched for genes showing elevated expression in both the testis and ovary. If loci impact both drive in males and fecundity in females, natural selection may select against suppression of the activity of these genes. In fact, models of drive demonstrate that in the absence of frequency dependent selection, a stable drive polymorphism may still be maintained when female fecundity and drive are impacted by the same locus, or tightly linked loci [6], [46]. Due to the relative scarcity of genes that impact both male and female reproduction [77], it has been thought unlikely that the same locus would impact both traits [14]. The excess of drive-associated genes expressed in both tissues provides a counter example that warrants further investigation.
In addition to identifying specific candidate genes that may be involved in meiotic drive in T. dalmanni, we identified a number of patterns associated with genes that are differentially expressed between XSR and XST testes. First, we found that the X chromosome carried a majority (∼80%) of the genes whose expression differed between XSR and XST testes. In addition, we found that there was a large excess of gene movement from the autosomes to the X chromosome relative to Drosophila, especially among testes-specific genes, though this type of movement was enriched in control genes as well as drive-associated transcripts (Figure 2). Inheritance of XSR is generally sufficient to induce drive regardless of the genetic background, implying that segregating suppressors of XSR are absent or rare in nature. However, although the XSR chromosome has a strong genetic effect to induce meiotic drive, it is not necessarily obvious that changes in expression should be limited to the X chromosome. Meiotic drive genes on XSR could in principle act as “triggers” that alter expression of genes in trans across the genome. Alternatively, cis regulatory mutations, copy number changes, and the accumulation of null alleles [54] could affect the expression of genes on XSR directly. Our finding of a large X effect on drive-associated expression, along with the accumulation of many fixed genetic differences between XSR and XST genes, suggests that cis effects dominate trans effects in the case of sex chromosome meiotic drive. This is consistent with the hypothesis that stable persistence of a sex chromosome drive polymorphism requires that a suite of co-adapted genes be inherited together, often in the form of a large inversion or series of inversions [14]. Another possibility is that a meiotic drive trigger gene could impact expression preferentially on the X chromosome (chromosome-specific gene regulation). This is seen in the Slx/Sly system in mice, although in that case sex-linked genes are either up- or down-regulated by SLX or SLY respectively rather than causing a variety of expression changes (Coquet et al 2012).
We also found that these genes are evolving more rapidly at the protein level (dN/dS), and this increased evolutionary rate could not be entirely explained by a tendency of these genes to be testes-specific or linked to the X chromosome. By virtue of violating Mendelian inheritance, drive alleles produce a strong local fitness advantage, and if not suppressed, are expected to increase in frequency in the population, both removing polymorphism and bringing hitchhiking variants with them [1]. It is possible that much of the acceleration in the rate of protein evolution we observe is due to relaxed purifying selection during such a sweep (see [78], [79]). Alternatively, as XSR reaches higher frequency in the population, other genes in the genome may begin to evolve to adapt to the new genetic context. Theory predicts, for example, rapid evolution of modifier and suppressor loci should occur [56], [80]. Although we have not previously identified these loci, it is plausible that some differentially expressed loci may be modifiers of drive.
Because the testes we collected were from an outbred population, we were able to use natural variants in the XSR and XST individuals to confirm that XSR almost certainly contains at least one inversion that prevents genetic exchange between the XSR and XST chromosomes. Nearly 1,000 variants have become fixed between XSR and XST, whereas only 11 such differences exist between autosomes carried by XSR and XST males. It would be difficult to explain this discrepancy in any way other than a lack of genetic exchange between XSR and XST – it is highly unlikely that freely recombining chromosomes would pick up any fixed differences, whether X-linked or autosomal (Figure S2). A simple simulation (Figure 4) demonstrates that there are more genes carrying zero fixed differences than expected if recombination was suppressed uniformly across the X chromosome and affected all genes equally. The apparent clustering of fixed differences could be due to some proportion of the drive X chromosome continuing to recombine normally with standard X chromosomes. Alternatively, the fixed differences may cluster due to selection acting on certain genes differently between drive and standard individuals, even when the entire drive X is failing to recombine with standard X chromosomes. Further genetic analysis will be needed to discover which regions of the XSR chromosome recombine and which do not. A number of these fixed differences caused nonsynonymous changes in proteins, some of which were drive-associated (Table 3). These genes may be good initial targets for future analysis.
Finally, a large number of genetically isolated populations of T. dalmanni – as well as the closely related species T. whitei - can be found in southeast Asia and the valleys neighboring the Gombak valley from which the flies used for this study were collected. Many of these populations express sex chromosome drive ([81] and unpublished data). Although reverse genetic dissection is difficult in this species, these flies represent a potential natural laboratory for the study of gene expression and meiotic drive. Sex chromosome drive has persisted as a stable polymorphism in T. dalmanni for many generations – possibly for millions of years, given that it exists in the sexually dimorphic sister species in the same genus. Within such a long timescale, drive X chromosomes may have arisen once, or they might be evolving constantly through arms races between suppressors and drivers. In either case, further study of teleopsid populations and species will advance our understanding of how meiotic drive can impact gene structure and function when it is a constant evolutionary companion.
Testes for RNAseq were dissected from mature T. dalmanni males derived from an outbred lab population established in 1999 (cf. [28]). This population was founded from ∼100 flies caught in the Gombak valley in Malaysia and was maintained for approximately 30–40 overlapping generations at that size. After dissection, testes were transferred to RNAlater and stored at −20°C, and remaining tissue was used to extract DNA using Chelex [82]. We determined XSR/XST status using three X-linked microsatellite markers previously associated with meiotic drive [26], [28]. These three markers span the X, and previously a “drive” haplotype including these three loci was diagnostic for drive [37]. The frequency of drive in the 1999 population was estimated to be 24% (15/62 phenotypically screened flies) in 2003 [28] and 18% (22/122 genotypically screened flies) in 2010. Multiplexed PCR was performed using three fluorescently labeled primers and PCR products were genotyped on an ABI 3730×l DNA analyzer (Applied Biosystems). Products were sized using Rox500 and scored with GeneMapper 4.0 software (Applied Biosystems). Two replicate samples each were pooled for individuals carrying XSR or XST as follows: sample XSR-1, 35 testes pairs; XSR-2, 30 testes pairs; sample XST-1, 38 testes pairs; and sample XST-2, 30 testes pairs. RNA was extracted using the mirVana RNA Isolation Kit (Invitrogen) according to manufacturer's protocols for extracting mRNA. Samples were sent to Cofactor Genomics (St. Louis, MO) for bar-coding and library preparation and 51.5 million 60 bp paired-end reads were obtained by sequencing all four libraries across two lanes in an Illumina Genome Analyzer run.
A T. dalmanni draft transcriptome assembly was generated using 100 bp paired end Illumina HiSeq reads from five T. dalmanni tissues (ovaries, testes, gonadectomized females, gonadectomized males, and third instar larvae), 84 bp paired end Illumina GA reads from female and male heads, and the 60 bp XST and XSR testes paired reads described above. Together, these samples produced ∼308.5 million reads and ∼55.5 Gbp of sequence. All reads were assembled into a single transcriptome using Trinity (paired end mode, –CPU 24, –kmer_method inchworm –max_memory 190G). The resulting transcriptome assembly and associated raw read data can be obtained from NCBI as BioProject accession PRJNA240197. In order to be compliant with NCBI's TSA (transcriptome sequence assembly) database, a small number of the contigs in the original assembly were trimmed to remove potential vector contaminants, and a handful of contigs were shorter than the minimum 200 bp required for TSA and could not be uploaded. These sequences are available from the authors by request. Details of sequencing and assembly can be found in Table S6.
We used a modified version of bwa [38] that allows multiple mapping (available as part of the Trinity RNAseq software bundle, [83] trinityrnaseq.sourceforge.net) to align the left end reads back to the transcriptome. We chose to align the left end only because 1) the right end is not independent from the left and therefore adds no additional power to the analysis and 2) the first read is typically higher quality than the second [84]. The T. dalmanni transcriptome assembly contains many genes that are represented by multiple transcripts - often, these are multiple isoforms of the same gene. After alignment with bwa, expression was quantified using RSEM to correct for hits to multiple isoforms of the same gene. Genes were defined as those transcripts derived from the same Trinity component (see trinityrnaseq.sourceforge.net), and read counts were corrected at the gene/component and isoform/seq level based on the share of reads derived from each isoform. Corrected gene-level read counts were used with DESeq [40] to identify significantly differentially expressed genes between the two XSR and two XST samples using a 0.001 FDR cutoff and using DESeq's independent filtering option to improve power. The highest expressed isoform for each gene/component was identified and used for subsequent analyses. To ensure our results were independent of the statistical method, we also used edgeR [85] to identify significantly differentially expressed genes and obtained qualitatively similar results (Table S7). Only DESeq results are presented henceforth.
The expression patterns of T. dalmanni genes were assessed using transcriptome sequencing from six T. dalmanni tissues (ovaries, testes, gonadectomized females, gonadectomized males, adult heads, and third instar larvae). With the exception of the heads, each of these tissues included two biological replicates. For the heads, one sample was from females and the other was from males. These were treated as biological replicates for the analysis of expression across tissues as we were more interested in differences between tissues than between the sexes per se. Corrected read counts for each sample were obtained as described for the XSR versus XST comparison above. Normalized gene-level expression values (FPKM) were determined and expression profiles were assessed using tools provided with the trinity RNAseq package [83] as described on trinityrnaseq.sourceforge.net, “Identifying Differentially Expressed Transcripts” (see also, [43]). A 0.001 FDR cutoff was used to identify genes that were significantly differentially expressed between samples. The significantly differentially expressed genes were then grouped by similarity of their expression patterns using Euclidean complete clustering. Next, we used k-means clustering to define distinct expression pattern groupings from among the differentially expressed genes (see Figure S1). We tried a range of K values (6 to 12) and assessed the number of genes and the expression profile for each cluster. We chose K = 8 for further analysis, as this number of clusters provided the maximal number of qualitatively different expression patterns. Increasing the cluster number to 9 added a cluster with different expression levels but the same expression pattern as already represented by previous clusters. In addition to the gene expression pattern analysis presented above, we also calculated a measure of tissue specificity – Tau [44] – for each gene using the average of the two FPKM values for each tissue. Genes with Tau >0.95 were considered to be expressed specifically in the highest expressed tissue. For all subsequent analyses using the sequence of a gene (gene prediction, orthology prediction, linkage, etc.), the highest expressed isoform (Trinity variant) of each gene was used as the representative sequence.
Testes were dissected from a newly collected (August, 2012) population of T. dalmanni Gombak. This population was used because genotyping of ∼150 flies in the 1999 population failed to identify any males carrying the previously defined XSR haplotype [28]. To obtain testes from drive males, we genotyped second-generation T. dalmanni males from the 2012 Gombak population using the three markers described above. A male was defined as carrying a drive haplotype if he carried an ms125 allele <152, an ms244 allele >238, and an ms395 allele >230. Breeding studies using the 2012 population confirm that males with these haplotypes produce drive and reveal that other drive haplotypes exist (unpublished data). To confirm standard status we phenotyped individuals for unbiased progeny sex ratios (between 0.4 and 0.6 proportion sons, 50 + progeny). RNA was extracted from pools of 3 testes pairs using the mirVana kit (Invitrogen AM1560) and first strand cDNA was synthesized using M-MLV reverse transcriptase (Promega M1705). From the list of candidate genes, 18 (11 that were up in XST testes and 7 that were up in XSR testes) were selected for confirmation by quantitative reverse transcription polymerase chain reaction (qRT-PCR). qRT-PCR was conducted on a Bio-RAD CFX real-time PCR machine using SYBR 2× RT-PCR mix (Invitrogen 4472942), 1 uL of cDNA template and gene specific primers. In order for a primer pair to be used, it had to have a Ct value below 32 in both replicates in at least of one of the two conditions, otherwise we discarded it from the analysis. Six genes were excluded using this criterion. A failure to detect expression in RT-PCR could be for a variety of reasons: 1) the number of testes in the pool was much smaller so if there was an expression polymorphism in the original pools it might have been missed, 2) the population sampled for qRT-PCR was 12 years separated from the lab population so differences in expression may be present, and 3) the total amount of RNA was much less. Primers were also tested on genomic DNA to ensure that failure to amplify was not due to primer failure. Expression was quantified relative to a control gene (GAPDH-2), and when all four samples showed robust expression, a t-test was performed on resulting delta Ct values between the two conditions. When data was available from all samples, the log2 expression differential was calculated using the delta-delta Ct method [86] between XSR and XST samples relative to GAPDH-2 (Table S1).
After determination of expression values for all genes (above), we created a control gene list by removing the most weakly expressed genes in the Trinity assembly. This step prevents misinterpretation of results that could arise from inclusion of very weakly expressed transcripts in the control dataset (such weakly expressed transcripts could never be detected as drive-associated due to a lack of statistical power). Therefore, we defined an expression floor using the drive-associated genes. We identified the tissue for each drive-associated gene that had the highest expression level (the maximum expression level for that gene) and ranked genes by this value from lowest to highest. We used the lowest maximum expression level among the drive-associated transcripts (FPKM = 0.86) as the expression cutoff for the control gene set. If the highest expressed sample for a transcript had an FPKM>0.86, it was included in the control gene set. Otherwise, it was removed from further analysis.
Transcripts were annotated as having Drosophila orthologs using blastp. First, proteins were predicted from T. dalmanni transcripts using two methods – the longest start to stop ORF and FrameFinder [87], which can find longer ORFs if a transcript is truncated due to poor assembly. FrameFinder was run with the local (not strict) model using a word probability set generated from the entire T. dalmanni transcriptome using the Fasta2count and wordprob tools included with FrameFinder, and options were set to disallow frameshifts and indels (options: -I −500 –D −500 –F −500 –s False). The proteins generated by these two methods for each transcript were aligned by blastp to the D. melanogaster proteome (Flybase v. 5.50) and the best hit in T. dalmanni was kept for each gene. An e-value cutoff of 0.1 was used, and only hits covering 50% or more of the D. melanogaster protein were kept. The coverage cutoff prevented keeping partial hits due to the assembly incorrectly splitting a gene into two contigs, both contigs hitting different parts of the same ortholog, and being seen, incorrectly, as paralogs. If both the framefinder and the longest orf predictions had qualifying hits, the best hit (by e-value, then by %ID) was kept. In the case of a tie, the FrameFinder hit was kept. These protein hits were annotated using the Flybase batch download tool. The gene family size in T. dalmanni was estimated for genes with D. melanogaster orthologs. The number of occurrences of each Flybase gene ID (Fbgn) among the putative orthologs of the control gene set was used to estimate the T. dalmanni gene family size for each gene.
A T. whitei transcriptome was assembled using Trinity [83] (–max_memory 190G –CPU 24 –kmer_method inchworm, paired-read mode) on RNAseq from a pool of approximately 30 pairs of T. whitei testes (33,753,826 100 bp paired end reads were generated, and 60,650 contigs were assembled). The T. whitei assembly and raw data can be obtained from the NCBI website under the BioProject accession PRJNA241109. As with the T. dalmanni assembly, to be compliant with NCBI's TSA database, a small number of the contigs in the assembly were trimmed to remove potential vector contaminants and sequences shorter than 200 bp were removed. These sequences are available from the authors by request. Proteins were predicted from the resulting transcriptome assembly as with T. dalmanni using both FrameFinder and the longest ORF as described above (the T. whitei transcriptome was used to create a word probability set for FrameFinder prediction). The T. whitei proteins were aligned by blastp to the predicted T. dalmanni proteins and hits with e values <0.1 were kept. For each gene, whichever predicted protein had the best hit to a predicated T. whitei protein was kept. The resulting T. whitei and T. dalmanni protein pairs were aligned using Clustal omega [88], and the consensus transcriptome sequences were mapped onto the protein alignments (after trimming excess sequence). dN/dS was predicted from each alignment greater than 50 amino acids in length using SNAP [89]. Only consensus sequences were used in calculating dN/dS - polymorphism in the RNAseq data was not considered in this analysis, as such data from RNAseq data can be unreliable (high or low levels of coverage due to differences in gene expression may cause over or underestimation of the number of polymorphic sites, respectively).
Gene ontology analysis was performed for Drosophila orthologs using DAVID functional annotation tools (http://david.abcc.ncifcrf.gov, [57], [90]). The list of D. melanogaster orthologs to drive-associated transcripts was compared to the orthologs in the control gene list and to the entire D. melanogaster proteome using DAVID's functional annotation tables tool. Annotations with an FDR<0.05 were considered significant when interpreting the output.
We used data from comparative genomic hybridization (CGH) to determine linkage of genes differentially expressed between drive and standard samples as well as the rest of the Trinity assembly. The CGH data are accessible using accession number GSE55601 from the NCBI Gene Expression Omnibus (GEO). First, the log2 ratio of female to male expression was calculated for each probe on each of four duplicate oligonucleotide Agilent arrays containing 180K probes representing 12,000 unique genes. These values were normalized so that the maximum number of probes had a log2(f/m) ratio of 0, the expected value given the nature of the array (divergence from 0 is caused by small differences in the quality or quantity of genomic DNA from the two sexes applied to the array). The sequences for the probes (see GSE55601) were then aligned with BLAT (multiple matching allowed, perfect hit required) to the Trinity assembly, giving a set of probes matching each transcript. The median of the probe values for a given transcript was calculated for each array, and then a median and a 95% confidence interval (CI) was calculated across the four arrays. Calls for the linkage of each contig were as follows: 1) if the upper bound of the CI was less than −2, the transcript was called Y-linked; 2) if the lower bound of the CI was greater than 0.5, the transcript was called X-linked; 3) if the CI was entirely between −2 and 0.5, the transcript was called autosomal; 4) if the CI overlapped any of these bounds, or if a transcript had only a single probe or a single array informing on it, it was called U. For genes with putative D. melanogaster orthologs, the linkage of each gene was compared relative to D. melanogaster. The X chromosome in T. dalmanni is mostly orthologous to chromosome 2L in D. melanogaster [30]. Therefore, genes that are X-linked in T. dalmanni and on non-2L chromosomes in D. melanogaster have most likely moved relative to one another in one of the two lineages. Likewise, autosomal genes in T. dalmanni that are on 2L in D. melanogaster have most likely moved at some point since they last shared a common ancestor.
After alignment of the RNAseq data to the transcriptome with bwa, SAMtools [91] was used to create pileup files across the T. dalmanni transcriptome. Using the pileup files from XST and XSR testes, we counted the number of sites on each transcript that were fixed as different alleles between the two samples. We ignored sites that were polymorphic in either the XST, XSR, or both samples. In order for a site to be used, it had to have at least 10 reads informing on it in both samples (“10× coverage”), and there had to be at least 100 sites from a given transcript with sufficient read coverage for that transcript to be used. To determine if an excess of fixed differences on the X could be due to the fact that half as many X chromosomes as autosomes are sampled in males, we used fastsimcoal2 [92] to simulate populations of chromosomes with 100,000 SNPs using various values of Ne, a per SNP recombination rate of 10−5 per generation, and a minimum possible derived allele frequency of 10−6. We simulated 100 replicate populations with each parameter set. In each simulation we took two samples of equal size from each set of chromosomes, counted the number of fixed differences between the two samples, and then averaged across each set of parameters. Under all parameter sets, once at least 16 chromosomes were sampled, no fixed differences were observed (Figure S2). To determine if the entire X chromosome is nonrecombining, we used the observed probability of a fixed difference per basepair and performed 10,000 simulated draws from the binomial distribution for each of the transcripts that carried a fixed base using the observed distribution of transcript lengths and the per site rates of fixed differences in X-linked and autosomal genes. We compared the resulting distribution of fixed differences per transcript to observed values to determine if the observed distribution was different from that expected if fixed differences are randomly distributed across the X assuming the X chromosome carried by XSR individuals was entirely nonrecombining.
|
10.1371/journal.pgen.1003630 | Fine Time Course Expression Analysis Identifies Cascades of Activation and Repression and Maps a Putative Regulator of Mammalian Sex Determination | In vertebrates, primary sex determination refers to the decision within a bipotential organ precursor to differentiate as a testis or ovary. Bifurcation of organ fate begins between embryonic day (E) 11.0–E12.0 in mice and likely involves a dynamic transcription network that is poorly understood. To elucidate the first steps of sexual fate specification, we profiled the XX and XY gonad transcriptomes at fine granularity during this period and resolved cascades of gene activation and repression. C57BL/6J (B6) XY gonads showed a consistent ∼5-hour delay in the activation of most male pathway genes and repression of female pathway genes relative to 129S1/SvImJ, which likely explains the sensitivity of the B6 strain to male-to-female sex reversal. Using this fine time course data, we predicted novel regulatory genes underlying expression QTLs (eQTLs) mapped in a previous study. To test predictions, we developed an in vitro gonad primary cell assay and optimized a lentivirus-based shRNA delivery method to silence candidate genes and quantify effects on putative targets. We provide strong evidence that Lmo4 (Lim-domain only 4) is a novel regulator of sex determination upstream of SF1 (Nr5a1), Sox9, Fgf9, and Col9a3. This approach can be readily applied to identify regulatory interactions in other systems.
| The commitment of a bipotential gonad to differentiate as a testis or an ovary is governed by a dynamic transcription network that remains to be elucidated. We profiled genome-wide gene expression at <5 hr resolution during the critical one-day developmental window in XX and XY mouse gonads from the 129S1/SvImJ and C57BL/6J strains. We identified cascades of expression in both strains and show that establishment of dimorphic expression is largely due to activation and repression programs initiated by the testis pathway. Strikingly, comparison of expression differences between the two strains revealed a delay in the male program in C57BL/6J gonads, suggesting an explanation for the increased susceptibility to male-to-female sex reversal exhibited by this strain. Finally, we exploit the predictive power inherent in this temporal dataset to identify a novel candidate gene, Lmo4, underlying an expression QTL identified in a previous study. Confirming our prediction, knockdown of this gene in primary XY gonad cells resulted in the down-regulation of several male program genes. Our results highlight the importance of fine-scale resolution time-course measurement of expression in developmental systems to identify candidate regulatory genes and to understand general properties of the system.
| Fate determination of the bipotential gonad results in differentiation of a testis or ovary and is crucial to the sexual differentiation of the embryo. This binary decision, known as primary sex determination, takes place at mid-gestation in the mouse. The initial pliable nature of the gonad and its rapid progress into one of two divergent, opposing fates makes it a particularly attractive model to investigate transcriptional network dynamics during fate decisions in developmental systems.
The early sexual plasticity of the mammalian gonad appears to result from a balanced, transient transcriptional network state [1]. Many genes associated later with a specific sexual fate are expressed early and at similar levels in both XX and XY gonads, a pattern indicative of lineage priming [2]. Sex determination proceeds by first establishing a bias in the transcription network toward the male (testicular) or female (ovarian) fate. In therian mammals, the Y-chromosome transcription factor, Sry, is the genetic trigger responsible for diverting the bipotential gonad to a testicular fate. Sry is expressed in XY gonads beginning at E10.5 and plays an important role in the up-regulation of Sox9 and Fgf9 [3]–[5]. While several genes are known to be required for adult ovarian fate [6]–[9], much less is known about the initiation of the female pathway. Subsequent to the primary fate decision in both differentiation pathways, feedback mechanisms are activated that canalize the chosen sexual fate and repress genes associated with the alternative fate. Failure to trigger or maintain one sexual fate can result in trans-differentiation to the alternative fate (i.e., sex reversal) [10]–[12].
Several lines of evidence suggest that many more important players in mammalian sex determination await discovery. First, approximately 1,500 genes are already expressed in a sexually-dimorphic pattern at E11.5, when the gonad is morphologically indistinct and still competent to sex-reverse [13]–[18]. Second, the majority of cases of human sex reversal are yet to be explained by any of the genes known to have an impact on sex determination [19].
Some inbred strains appear to be better suited than others to cope with perturbations in the sex determination pathway. For example, C57BL/6J (B6) is sensitive to XY male-to-female sex reversal in response to multiple genetic perturbations (including both Y-linked and autosomal variants), while other strains like 129S1/SvImJ (129S1) and DBA/2J (D2) are resistant to these variants (i.e., develop normal testes) [20], [21]. This differential sensitivity to sex reversal was first exploited by Eicher and colleagues in genetic studies to map regions of the B6 genome correlated with sensitivity of the XY gonad to male-to-female sex reversal [22]. More recently, using an expression QTL (eQTL) mapping approach, we identified multiple genomic regions where segregation between B6 and 129S1 markers was highly correlated with the expression levels of multiple genes of known importance to sex determination [1]. Most of these eQTL intervals harbored no genes with known roles in sex determination, and thus, likely contain novel genes in the network.
To improve our resolution of the transcriptional cascade controlling sex determination, and choose attractive candidates in eQTL intervals, we conducted a fine time course transcriptome analysis of the gonad between E11.0–E12.0, when the bipotential gonad approaches a decision point, initiates the testicular or ovarian pathway, and begins to reinforce the sexual fate decision. We profiled global gene expression at six equally-spaced intervals in XX and XY gonads from the susceptible B6 and resistant 129S1 strains, developed and trained a Hidden Markov Model (HMM) to discern the onset of sexually-dimorphic expression, and identified gene cohorts activated or repressed specifically in the testis or ovary during this brief 24-hour window of development. By comparing the onset profiles of both strains, we found that susceptibility to sex reversal in B6 XY gonads is likely due to the delayed activation of many testis pathway genes and delayed repression of many ovarian pathway genes. We exploited this detailed view of the B6 and 129S1 gonad transcriptomes to prioritize candidate regulatory genes underlying eQTLs mapped in our previous study [1]. Finally, we developed a primary cell validation assay and lentivirus-based shRNA delivery method to artificially silence Lmo4 (Lim-domain only 4), a candidate regulatory gene within an eQTL interval. We provide strong evidence that Lmo4 is a novel regulator of sex determination upstream of many sex-associated genes. This work provides a systematic framework for predicting and testing regulatory genes (eQTGs) underlying eQTLs that is applicable to other systems.
To elucidate the fine temporal dynamics of the gonad transcriptome during the critical fate decision to develop as a testis or ovary, we assayed total transcript abundance in XY and XX gonads at six equally spaced intervals between E11.0–E12.0 (Figure 1A). This 24-hour window captures the gonad from the time it is still bipotential to a point when it has shifted to a testis or ovarian fate. To associate variation in the transcriptome with phenotypic differences, we compared gene expression from gonads of two common inbred strains, 129S1 and B6, that differ in their susceptibility to XY sex reversal.
Total transcript abundance was measured by microarray for individual pairs of gonads for each sex/strain/stage combination (n = 74 total arrays, see Materials and Methods). A total of 9,254 genes (12,213 probes) exhibited significant expression above background in at least two replicates of one sample type, and were included in subsequent analyses (Figure 1B). Next, we fit a linear model accounting for the effects of strain, sex, stage, and two-way (e.g. sex*stage) and three-way (e.g. sex*stage*strain) interactions among these factors (Figure S1). For more than half (n = 4,752 (5,659 probes)) of the genes that passed our filtering criteria, a significant proportion of the observed variation in expression could be attributed to one or more experimental variables (Figure 1B). The individual components of sex (n = 1,172 genes/1,343 probes), stage (n = 2,434 genes/2,805 probes), and strain (n = 3,279 genes/3,879 probes), as well as the interaction effect of sex by stage (n = 659 genes/733 probes), all had significant effects on the expression of hundreds to thousands of genes. For many of these genes, expression was influenced by the additive effects of multiple components (probes in overlap regions in Venn diagram, Figure 1C). Moreover, the sex by stage interaction effect reflects the number of genes that have a sexually-dimorphic pattern of expression that changes over time in the E11.0–E12.0 window. Finally, the large number of genes with a strain effect highlights the extent to which the transcription programs vary in the B6 and 129S1 strains. These data illustrate both the complexity and dynamic nature of the transcriptional program driving sex determination during this brief but critical developmental window.
To obtain a global view of how dimorphism is achieved during the sexual fate specification of the gonad, we calculated the fold change in both XX and XY gonads for each gene between E11.0 and E12.0 in the robust 129S1 strain. We then graphed these sex-specific fold changes for each gene on an X-Y scatter plot, with fold changes in the XY gonad appearing on the Y-axis, and changes in the XX gonad appearing on the X-axis (Figure 2, Dataset S1). Complex refractory or oscillatory patterns were not detected over this relatively short temporal window, and therefore this two-stage comparison accurately characterized overall changes in gene expression. We first note from the scatter plot that the microarrays captured the expected expression patterns of several genes with known roles in sex determination. Some of these exhibiting male enrichment (Sox9, Dhh, and Cbln1 [23], [24]), or female enrichment (Irx3, Wnt4, and Msx1 [25], [26]), are shown adjacent to their locations in the scatter plot. In addition to genes with sexually dimorphic expression patterns, we also identified genes that are identically-repressed or activated by both the male and female programs (genes on the diagonal in Figure 2). We hypothesize that pathways active early in both sexes are associated with a plastic bipotential state. We would expect these genes (e.g. Gfra3 or Hoxa7) to be down-regulated in both sexes as sexual differentiation proceeds. In contrast, genes that are identically-activated in both sexes between E11.0–E12.0 (e.g. Hoxd10 and Hs3st3a1), may be associated with transition from a sexually-primed but plastic transcriptional state to a ovarian or testicular fate regardless of the nature of the fate commitment. A subset of these genes (14, log2(fold change) >0.585 in both sexes) overlaps with the “core adrenogonadal program” previously identified in the related steroidogenic-factor-1-positive cell population [27].
This view of sexually dimorphic expression changes between E11.0 and E12.0 also revealed that higher expression in one sex can result from activation in one sex (e.g. Dhh in Figure 2), repression in the other sex (e.g. Msx1), or a combination of both mechanisms (e.g. Wnt4). From the scatter plot, it is evident that dimorphic expression of most genes (205 genes, log2(fold change) >0.585) expressed higher in the XY gonad occurred primarily through activation, with a small outlier group of genes (25 genes, log2(fold change) <−0.585) showing dimorphism as the result of repression in the XX gonad. Among genes showing higher expression in XX gonads, two principal gene clusters were evident: members of one cluster (77 genes) achieved dimorphism primarily through activation in the XX gonad, and members of the other (148 genes), primarily through repression in the XY gonad. This indicates that the dynamic expression changes observed during gonad fate commitment are a result of the action of activation and repression programs. We designed our following analysis to thoroughly characterize these and other aspects of the male and female transcriptional programs.
To identify ordered cascades of expression and co-regulated genes, we developed a Hidden Markov Model (HMM) (Figure 3). HMMs are well-suited to the task of discerning patterns in time series data [28], [29] because they use correlations between adjacent time points to overcome noise and increase sensitivity. Briefly, the HMM was designed with 18 states, three per time point – a male (i.e. testis-enriched) state, a female (i.e. ovary-enriched) state, and a similar expression state (Figure 3, Figure S2). The fold difference of a gene's expression between XX and XY gonads at each time point was used to train the HMM. After training the model, the Viterbi state path of each gene reflected whether the gene was expressed in a sexually dimorphic fashion, the sex in which it was expressed more highly, and the times at which the gene exhibited dimorphic expression. Importantly, only 22 of a possible 729 state paths through the model were populated (Table S1, Dataset S2), indicating that despite the highly dynamic changes in the transcriptome, there are common expression trajectories by which the expression patterns can be clustered. We note that while the HMM classifies groups of genes as becoming dimorphically expressed at specific times, this is due to the discrete sampling during the window. In reality, expression changes likely occur in a continuum. Nonetheless, grouping the genes by their onset of dimorphism reveals interesting details of the regulatory programs involved.
Out of the 4,752 genes included in the analysis, 1,321 genes exhibited dimorphic expression at one or more time points between E11.0–12.0 in the 129S1 strain and similar numbers (1,037 genes) were dimorphically expressed in the B6 strain (Table S1). Interestingly, for both 129S1 and B6, once a gene established a dimorphic expression pattern, most continued in a state of sexually dimorphic expression until E12.0 (n = 1,254 genes for 129S1, n = 995 genes for B6). We refer to these genes as male- or female-enriched depending on which sex exhibited higher expression. Finally, only three genes (Lefty2, Mcm6, and LOC233529) in 129S1 (and none in B6) switched from being more highly expressed in one sex to the other during the duration of our window.
We used the HMM to cluster male- and female-enriched genes by the time of onset of dimorphic expression from E11.2 to E12.0 for the 129S1 strain (Figure 4A, B). This analysis revealed striking cascades of sexually-dimorphic male and female enrichment with the number of male- and female-enriched genes gradually increasing across time points. For example, for male-enriched genes, a single gene (Sox9) showed higher expression in males starting at E11.2, followed by 30 genes at E11.4, and finally 202 genes that showed sexually dimorphic expression at E12.0 (Table S1). To determine whether these cascades primarily reflected changes in one gonadal cell type or several, we compared our whole gonad data with cell type-specific gene expression data from E11.5 and E12.5 isolated XX and XY supporting cells and germ cells [2]. We found that the overlap with germ cells was low (5%) (Figure S3). In contrast, 58% of genes that became male- or female- enriched in our whole gonad transcriptome prior to E11.8 were specifically dimorphic at E11.5 or E12.5 in supporting cells. After E11.8, the overlap with the supporting cell precursors dropped to 45%. Thus, consistent with previous results, the supporting cell lineage, known to be critical for initiating the sex determination decision, is responsible for a large proportion of the sexually-dimorphic gene expression that arises in the gonad between E11.0–E12.0 [2].
To determine whether activation, repression, or a combination of both was involved in primary establishment of dimorphism for each gene, we compared gene expression in each sex before the initiation of differential expression and the E12.0 stage (Figure 4C). Note that this analysis is more sensitive than the scatter plot (Figure 2) in identifying the cause of dimorphism as it examines the trajectory of expression following the onset of dimorphism as opposed to the starting point of the analysis (E11.0).
For all genes that showed higher expression in the XY gonad by E11.8, 73% of genes (256 genes) were strongly activated (log2(fold change) >0.32, p<0.05) in XY gonads, whereas only 9.5% (34 genes) were repressed in the XX gonad (log2(fold change) <−0.32, p<0.05). In addition 5.6% (20 genes) become dimorphic through a combination of activation in the XY and repression in the XX gonad. This indicates a strong activation program in XY gonads with a much lower contribution from repression of male pathway genes in XX gonads. In striking contrast, enrichment of genes in XX gonads results not from activation in the ovary, but primarily through repression in the testis (Figure 4C, lower panel). Only 16% of probes (61 genes) that are female-enriched by E11.8 are activated in the XX gonad, while 61% (217 genes) are repressed in the XY gonad, with 7.5% (27 genes) becoming dimorphic due to a combination of activation in the XX and repression in the XY gonad. In fact, in several cases, Msx1 for example (Figure 2), female enrichment stems exclusively from repression taking place in the testis. This indicates that in addition to the activation program, a strong repressive program is also present in the testis. While male repression of specific female pathway genes (Wnt4) has been known, the extent of this repressive signature is surprising.
The HMM classified Sox9 as the earliest male-enriched autosomal gene (E11.2), reflecting its position directly downstream of Sry in the testis pathway [5] and affirming the fine temporal resolution of our dataset. As in previous microarray experiments using whole gonad samples [14], Sry was not detected above background levels in the current study. Following the up-regulation of Sox9, several other known crucial downstream genes such as Fgf9, Amh, and Dhh showed increased expression in XY gonads (Figure 4A, top panel). In the female-enriched group, Wnt4, one of the earliest autosomal genes known to act in ovary differentiation [6], [12], was sexually-dimorphic no earlier than E11.4. Other known female pathway genes such as Fst and Axin2 became differentially expressed at the same stage or immediately following the dimorphic expression of Wnt4. All the male-enriched genes that were activated prior to Sox9 and female-enriched genes enriched prior to Wnt4 are Y- and X-linked genes, respectively (Figure S4 and S5). Particularly interesting are the 7 X-linked genes that exhibited higher expression in XX gonads starting at E11.2 (Figure S5). The cell-type specific data indicate that these genes are all highly expressed in germ cells and likely reflect the reactivation of the inactive X chromosome in XX germ cells at this stage [30].
After characterizing the temporal dynamics of XY and XX gonad transcriptomes from the 129S1 strain that is resistant to XY sex reversal, we determined how the transcriptome varied in B6, a strain that is sensitive to XY sex reversal in response to multiple genetic perturbations [20], [31], [32]. While previous studies showed that male-enriched genes were expressed at a higher level and female-enriched genes at a lower level in 129S1 compared to B6 E11.5 XY gonads [1], it was not clear whether this strain difference was a result of the difference in expression levels, time of onset, or a combination of both. To address these questions, we profiled global gene expression in XY and XX gonads from B6 at the same six time points spanning the critical 24-hour window (E11.0–E12.0).
We used the HMM to identify male- and female-enriched genes in B6 mice (Figure S6). In good agreement with the data from 129S1 mice, these genes showed activation and repressive programs in XY gonads, and were strongly biased toward dimorphic expression in supporting cells. We then compared the timing of onset of sexually dimorphic gene expression between 129S1 and B6 (Figure 5). We observed a clear, consistent temporal shift in the onset of sexually-dimorphic expression in many genes in the susceptible B6 strain (Figure 5A, B). Specifically, for male-enriched genes, a comparison of strain onset distribution profiles revealed a statistically significant ∼5-hour delay in B6 relative to 129S1 (Figure 5C, upper panel). For example, 208 probes became enriched in 129S1 XY gonads relative to XX starting at E11.6 (Figure 5B). When the same probes were examined in the sensitive B6 strain, only 37 became dimorphic at the same stage, while a majority (n = 107) became male-enriched ∼5 hours later at E11.8. Another 20 from this set did not become dimorphic in B6 until E12.0, and 35 probes that were male-enriched in 129S1 at E11.6 failed to become sexually-dimorphic in B6 by E12.0. Even in the case of genes that became male-enriched at the same time point in B6 and 129S1, a comparison of XY v. XX fold difference at the onset of dimorphism indicated that a majority of these genes (65.6%) show a higher male v. female fold difference in 129S1 than in B6 (Figure S7A, Binomial test p-value<0.005). This difference may reflect a more robust activation mechanism driving the male differentiation pathway in 129S1, or it could reflect a delay in expression onset in B6 that is less than ∼5 hours and therefore smaller than the minimum resolution threshold of this analysis.
Importantly, this delayed onset pattern of male-activated genes in B6 does not appear to stem from a difference at the top of the cascade in Sry expression level during this critical window. Although Sry was not detected above background in these arrays, there was no significant difference in expression levels between 129S1 and B6 between E11.2–E12.0 by qRT-PCR (Figure S8A). It should be pointed out that the high variability in Sry abundance observed among individual pairs of XY gonads could mask a small but real strain effect for Sry expression levels. In contrast, Sox9 is more robustly up-regulated in 129S1 relative to B6 (Figure S7B), an observation we confirmed by qRT-PCR (Figure S8B). By E11.8 expression of Sox9 in B6 XY gonads has caught up with expression in 129S1. Some of the delay in onset of male-enriched genes at later stages in B6 may be due to this initial deficiency in the robustness of Sox9 activation.
Note that the delayed onset pattern was not observed for every gene that became male-enriched over this 24-hour period. For example, of the 208 probes that were male-enriched starting at E11.6 in 129S1, a few (Gstm2, Etv5, Gas7, and Mybphl) became sexually-dimorphic earlier in B6. Similarly, of the 75 genes that showed male-enrichment in B6 at E11.6, 11% (n = 8, including Schip1, Lpl, Socs2) become male-enriched later in 129S1 or not at all before E12.0. This indicates that although much of the male differentiation pathway is delayed in B6, this pattern is unlikely to be due to a more general delay in gonad differentiation.
Female-enriched genes exhibited a similar significant ∼5 hour delay in B6. However, this strain delay stems from the later repression of female genes in XY gonads (Figure 5B lower panel). As a consequence, B6 XY gonads are exposed to higher levels of female pathway genes for a longer period relative to 129S1 XY gonads. For example, of the 166 probes that become female-enriched in 129S1 starting at E11.6, only six exhibit a similar expression pattern in B6, while 70 become female-enriched ∼5 hours later, another 43 probes become dimorphic at E12.0, and 47 fail to reach a sexually-dimorphic state by E12.0 (Figure 5C, lower panel). Similar to the male-enriched genes, 62.4% of female-enriched genes that become dimorphic at the same time in both strains show a higher fold difference in 129S1 than in B6 (Figure S7C, p<0.005).
In summary, our evidence argues against a general developmental delay in B6 and suggests that the increased sensitivity of B6 XY gonads to sex reversal stems from a delay in the activation of male pathway genes downstream of Sox9, combined with a consequent delay in the repression of female pathway genes.
In addition to providing a more comprehensive view of the global transcription dynamics driving male and female sex determination in the robust 129S1 and sensitive B6 strain backgrounds, this fine temporal expression data provided a means for narrowing eQTLs to identify novel regulators of sex determination. In previously published work, we mapped 19 regions of the genome in an F2 intercross population where genetic variation between B6 and 129S1 was correlated with differences in gene expression for one or more genes associated with sex determination [1]. Eight of these regions were correlated with the expression of multiple genes, yet none of these prominent “trans-band eQTLs” harbored an obvious candidate gene with a known role in the sex determination process. Unfortunately, most of the eQTL regions identified in this initial coarse mapping were too large to functionally test every gene in the interval.
We established filtering criteria based on temporal strain expression and genomic data to prioritize candidate genes within the eight trans-band eQTLs (Table 1). Briefly, protein-coding genes in the interval were considered as candidates if they were expressed at one or more time points between E11.0–E12.0 in XY samples (eQTLs were mapped only in XY samples; therefore, the causative gene underlying an eQTL should be expressed in the XY gonad). Based on this list, we analyzed each candidate within the interval for strain differences in expression levels or time of onset, and prioritized genes with strain-dimorphic patterns. We investigated whether each gene harbored one or more polymorphisms (SNPs, insertion/deletions) that differed between B6 and 129S1 and might affect its expression or function. Only those genes with characterized variation within 10 kb up- and down-stream of the transcription start site (TSS) were prioritized for further analysis. Finally, we interrogated the Mammalian Phenotype (MP) browser to identify any genes in the region with a characterized knockout phenotype affecting sex determination (MP:0002210, abnormal sex determination), or a known relationship with any of the target genes it was predicted to regulate. We tailored our candidate search strategy to each individual eQTL, and exploited prior information about the expression or function of the target genes for that region. Thus, we expected that genes involved in regulating early gonadogenesis genes would be expressed in both sexes at an early stage, whereas those regulating the male or female pathway would be more likely to exhibit sexually dimorphic gene expression.
In total, for the eight prominent trans-band eQTLs mapped in our previous study [1], of which the average interval contains ∼300 genes (range = 60–526 genes), we narrowed each down to, at most, eight promising candidates. Of particular interest, the distal Chr 3 region was strongly associated with the expression of nearly one-third of all the genes in our previous mapping study [1], including known regulators of early gonadogenesis (Fog2/Zfpm2, SF1/Nr5a1, Gata4, Wt1, and Ctnnb1) and both the female (Ctnnb1, Rspo1) and male (Fgf9) differentiation pathways. A total of six genes were identified as candidates based on their strain-dimorphic expression patterns, but only the transcription cofactor Lmo4 (lim domain only 4) exhibited a dynamic pattern consistent with a role early in both pathways and an additional male-specific role downstream of the sexual fate decision. Lmo4 is expressed at similarly high levels in both sexes until E11.4 (Figure 6A), and becomes male-enriched as early as E11.6. Importantly, while Lmo4 was up-regulated in B6 XY gonads at the same time as in 129S1, there was a significant strain effect with expression in 129S1 being higher. This observation is consistent with the observed allelic effect for the eQTL (B6129SF2 gonads that were homozygous for the 129S1 allele exhibited higher expression of target genes). Finally, there is a significant amount of genetic variation in Lmo4 between 129S1 and B6, including an insertion in the 3′ UTR in 129S1 as well as multiple intronic SNPs and indels [33]. Based on these selection criteria, we elected to focus on developing a functional assay for Lmo4.
Historically, moving from a list of candidate genes to a validated quantitative trait gene (QTG) has represented the largest hurdle (in both resources and time) to success in complex trait mapping studies in the mouse. To address this problem, we optimized a lentivirus-mediated shRNA delivery method to artificially silence candidate regulatory genes with high efficiency in dissociated gonad primary cells from E12.5 XY gonads (Figure 6B). As a positive control, we utilized pre-designed and validated shRNA clones (Sigma MISSION) packaged in lentiviral vectors to silence Sox9 expression in primary gonadal cell culture, and quantified the expression of known downstream targets (Figure S9). Lentiviral-mediated knockdown resulted in a nearly 80% reduction in Sox9 expression relative to a nontargeting control sample. Two of the three known targets (direct or indirect) of SOX9, Amh [34] and Fgf9 [4], were down-regulated significantly following Sox9 knockdown. Ptgds, the third known target of SOX9 [35], is not expressed at high levels in cultured gonad primary cells, and we could not detect a change in Ptgds expression following Sox9 knockdown. However, a marker of the female pathway, Fst [36], showed a significant and greater than 2-fold up-regulation in this assay (p<0.016). Thus, this in vitro assay recapitulates well-characterized genetic interactions that occur in the gonad in vivo.
We extended our analysis to test Lmo4 as a candidate regulator underlying the Chr 3 eQTL. We silenced Lmo4 expression to 36% of a nontargeted control (p<0.001) (Figure 6C). Although the degree of knockdown was relatively modest, it was observed consistently in three independent trials and with two shRNA clones. Importantly, silencing Lmo4 expression by two-thirds resulted in the consistent, significant down-regulation (p<0.05) of three of the four putative eQTL targets measured by qRT-PCR. Fgf9, Col9a3, and SF1/Nr5a1 were significantly down-regulated by both shRNAs (Figure 6C). Down-regulation of a fourth target of the Chr 3 eQTL, Wt1, was not statistically significant (p<0.13). Interestingly, although it was not identified as a Chr 3 eQTL target in our original mapping study, Sox9 expression is significantly down-regulated following Lmo4 knockdown by both shRNA clones (p<0.001). Note that the predicted targets (Fgf9, Col9a3, and SF1/Nr5a1) of the Chr 3 eQTL region are those that are affected by the different alleles in B6 and 129S1. Even though Sox9 did not map as a target of the Chr 3 eQTL in our study, it might not be differentially regulated by Lmo4 between B6 and 129S1, yet could still be a target of Lmo4. In total, these experiments provide strong support for Lmo4 as the transcriptional regulator underlying the Chr 3 trans-band eQTL and may reveal additional regulatory interactions that were undetected in the eQTL mapping study.
The transcriptional cascades that control development of multicellular organisms are a central focus in modern biology [37]. It is now evident that transcriptional regulation involves the coordinated action of a cohort of players including transcription factors, chromatin remodelers, non-coding RNAs, and epigenetic modifications. An important step in identifying the specific players in this network and deconvolving their effects on the transcriptional program is a detailed characterization of the transcriptome during the developmental process. Our efforts in this paper were focused on the critical 24-hour window when the gonad begins to transition from a bipotential primordium to a testis or ovarian fate. To that end, we sampled global transcript abundance at 3× finer granularity than previous studies, and in the process discovered multiple temporal cohorts of sexually-dimorphic genes in this brief window. Importantly, we found that sexually-dimorphic gene expression patterns during this period are primarily driven by activation and repression cascades in the XY gonad. Most male-enriched genes are activated in the XY gonad but remain unexpressed or unchanged in the XX gonad. In contrast, female–enriched genes acquire that pattern mostly by a combination of repression in the XY gonad and continued activation in the XX gonad. A ∼5-hour delay in both the activation of the testis pathway and repression of the ovarian pathway likely underlies the sensitivity of the B6 strain to XY sex reversal. We applied this new temporal expression resource to prioritize eQTL intervals mapped in our previous study. Finally, we developed a primary cell-based RNAi assay, and used it to validate a candidate new regulator of sex determination.
Previous microarray studies profiled transcript abundance in whole gonads or isolated cell populations at two or more time points before and after the sex determination decision [2], [13]–[15]. These datasets served as important resources for the field. However, the temporal resolution around the critical stage of sex determination was limited in all but one study to 24-hour intervals (Nef sampled at E10.5, E11.0, and E11.5). It was evident from these earlier studies that the gonad transcriptome changed very little between E10.5–E11.0, that the difference between E11.0–E11.5 was significant, and between E11.5–E12.5 the testis and ovarian transcriptomes are highly sexually dimorphic. We predicted that information about the sequential order of gene activation/repression during the E11.0–E12.0 window would be valuable. Using data from the fine time course, we designed an HMM to precisely separate genes based on their position in the transcriptional cascade. As opposed to other clustering methods such as k-means clustering [38], HMMs are able to both account for the time dependence in the data and exploit this layer of information to identify patterns often obscured by noise prevalent in microarray data. We note that this HMM can be readily extended to time course expression analyses in other systems.
Our analysis identified waves of sexually-dimorphic gene expression in the 24-hour window following the onset of Sry expression, which suggest regulatory cascades. We note that while the HMM identifies dimorphically expressed genes as being dimorphically expressed at distinct time points, this is a result of the sampling times of our transcriptome analysis. Finer sampling in this window is likely to reveal that genes grouped together at a time point show minor differences in timing of the onset of dimorphism. Previous work indicated that the supporting cell lineage is the first lineage in the gonad to show sexually dimorphic expression followed by other gonadal cell lineages after E11.5 [2]. Consistent with this, over half (58%) of the genes we identified that became sexually dimorphic prior to E11.8 could be specifically assigned to the supporting cell lineage prior to E11.8 with 5% showing dimorphism in germ cells. The discrepancies in the overlap are likely due to the increased sensitivity of the HMM to identify dimorphically expressed genes and the conservative measure of dimorphic expression used in the cell-type specific expression study.
As expected, we observed a strong signature of gene activation associated with up-regulation of the testis pathway in XY gonads. However, we were surprised by the extent of the repressive program that silences female pathway-associated genes in the XY gonad following activation of the male pathway. Testing of candidates from our study will be an important step towards identifying the factors responsible for these patterns of expression in the male and female program. Based on their early onset of sexually-dimorphic expression, several genes are promising candidates to play an early regulatory role in the male and female pathways (Table S2, Text S1). Sox13 is a member of the SOX protein family that lacks an activation domain but can repress Wnt signaling by forming a complex with the β-catenin cofactor, TCF1 [39], [40]. Mef2c is activated in the XY gonad at E11.6 and has been shown to interact with Sox9 in chondrocytes [41]. Among genes that showed female-enrichment, Gtf2a1, a general transcription factor that is part of the initiation complex for PolII recruitment [42], became dimorphic at E11.4 in the XX supporting cell lineage and Tcea3, a known PolII elongation factor became dimorphic at E11.6. Interestingly, female-enriched TFs such as Zfp277, Runx1, Lef1, Lhx9 and Msx1 were strongly down-regulated in XY gonads. Conversely, Irx3 showed strong activation in XX gonads, and has been predicted to have a function during ovarian differentiation independent of Foxl2 and Wnt4 [43].
Strain differences in resistance to sex reversal upon perturbation of the sex determination network have been the focus of several studies. The importance of the timing of the antagonistic testis and ovarian programs to B6-associated sex reversal was first proposed by Eicher in 1983 [44], based on the finding that introduction of a Mus domesticus Y chromosome (YDom, or YPOS) onto a B6 genetic background led to sex reversal [20]. Sex reversal in this case was later shown to be associated with the delayed onset of Sry [45]. However, this work did not explain why the B6 strain is more susceptible to sex reversal in cases where a weak allele of Sry is not involved [37].
Here we showed that the onset of sexually dimorphic gene expression was delayed by approximately five hours in the “unperturbed” (i.e. wildtype) B6 strain compared to 129S1. This delay is consistent across the cascade starting from E11.4 with genes that are both up- and down-regulated in XY gonads. Interestingly, we detected no significant difference by qRT-PCR in the level of Sry expression between the strains; however we cannot rule out a difference in the onset of Sry expression prior to the window of our analysis. Nonetheless, Sox9 is up-regulated at the same stage (E11.2) in B6 and 129S1. Despite this agreement, the activation of many downstream genes in the male pathway is delayed in B6. In our previous microarray comparison of B6 and 129S1 testes at E11.5 [1], Sox9 was found to be enriched in B6 relative to 129S1 XY gonads at E11.5, in contrast to the current study, where Sox9 levels are lower in B6 until E11.8–E12.0 (Figure S7B). This discrepancy may stem from small developmental staging differences in the pooled gonads used in the previous study, as Sox9 levels are changing very rapidly at E11.5. However, as this and other recent studies [32], [46] illustrate, the system may be sensitive to minor fluctuations in gene expression between E11.0–E11.5. Thus, the slightly lower level of Sox9 expression that we detected in B6 relative to 129S1 XY gonads might contribute to the delayed onset timing of downstream genes in B6. In addition, our fine time course data here helps explain the previously observed higher expression of female-enriched genes in B6 compared to 129S1. Specifically, the observed difference at E11.5 is a consequence of the delayed repression of female-enriched genes in B6 XY gonads.
Previous transcriptome and genetic mapping studies produced gene lists or large intervals with candidate regulators of sex determination [1], [13]–[16], [18], [22], [47]. The bottleneck in applying the results of these studies has been the inability to prioritize between several dozen candidates and then test these candidates in a manner that is inexpensive and efficient. We have addressed both these deficiencies in the current study. We used our fine time course dataset in conjunction with a previous eQTL study and cell-type expression data to identify candidate regulators of sex determination.
To overcome the hurdle of testing candidate regulators, we developed an RNAi assay to silence the expression of candidate genes, and then monitored the expression of putative downstream target genes after knockdown. As predicted, shRNA-mediated silencing of the transcription cofactor Lmo4 resulted in the down-regulation of known important regulators of early gonadogenesis (SF1/Nr5a1) and the male pathway (Sox9 and Fgf9). This provides strong evidence that in addition to previously characterized roles during development in the neural tube [48]–[50], neural crest [51], cortex [52], and thymus [53], Lmo4 is also a regulator of sex determination in the gonad. However, this does not preclude the possibility that other genes on distal Chr 3 have roles during sex determination and control the expression of one or more of the 16 eQTL target genes. To point, four of the other candidate regulators identified in this region (Gbp1/2/3, and Ccbl2) are expressed at similar high levels in both sexes before E11.4, and then become down-regulated specifically in XY gonads at or after E11.6. This pattern predicts a role for these genes in the female differentiation pathway. Future assays to overexpress these candidates in XY primary cells or silence them in XX primary cells will assess their potential as regulators for one or more of the Chr 3 eQTL target genes.
In closing, our fine temporal analysis of the gonad transcriptome revealed multiple cascades of sexually-dimorphic gene activation and repression during the critical first 24 hours of sex determination. This information provides a valuable resource for future experiments to identify novel genes, pathways, and network motifs associated with sex determination in particular, but also organ differentiation in general. We replicated this analysis in a strain that exhibits a unique sensitivity to sex reversal, and showed that the compromised capacity to buffer genetic perturbation in B6 is most likely due to a consistent ∼5-hour delay in the activation of a large portion of the male pathway and subsequent down-regulation of much of the female pathway. We integrated the temporal strain expression data with genetic mapping data that identified regions associated with gene expression in the gonad at E11.5, and in so doing were able to narrow down large intervals to a small set of the best candidate genes. Finally, we optimized lentivirus-mediated RNAi knockdown in cultured gonad primary cells, and used this assay to validate Lmo4 as a novel sex determination gene. Importantly, this validation strategy is easily scalable, and we expect that this assay will be a valuable first step to test potential regulators and in assembling a transcriptional network of sex determination.
All animals were maintained and experiments were conducted according to the Institutional Animal Care and Use Committee of the Duke University Medical Center and NIH guidelines (Permit Number: A168-11-07).
For the time course microarray study, C57BL/6J (stock no. 000664) and 129S1/SvImJ (stock no. 002448) mice were obtained from The Jackson Laboratory. CD-1 outbred mice were used (strain code 022, Charles River) in the gonad primary cell assays.
Timed matings were established for B6 and 129S1, and embryos were collected from dams between embryonic day (E) 11.0–12.0. Embryos were individually staged by counting tail somites (ts) distal to the hindlimbs: E11.0, E11.2, E11.4, E11.6, E11.8, and E12.0 corresponds to 13, 15, 17, 19, 21, and 23 ts, respectively [1], [54]. For each strain at each time point, three individual pairs of XX and XY gonads from at least two separate litters were collected. The chromosomal sex of each embryo was determined by PCR on head DNA using primers to detect Kdm5c/Kdm5d (5′-TGAAGCTTTTGGCTTTGAG-3′ and 5′-CCGCTGCCAAATTCTTTGG-3′). Gonads were dissected away from mesonephroi in sterile PBS (Gibco/Invitrogen, cat no. 1490-144) and stored in RNAlater RNA stabilization solution (Ambion, cat no. AM7024) at −20C until all samples were collected. To minimize contamination and RNA degradation, all surgical instruments and surfaces were treated with RNaseZAP RNase decontamination fluid (Ambion, cat no. AM9780), followed by 70% EtOH in DEPC-treated water, before and during the dissection procedure.
For the microarray analyses, at least three biological replicate samples were profiled for each strain/stage/sex (n = 74 total arrays), with one exception (n = 2 replicates for 129S1 E12.0 XY). Total RNA was first extracted from individual pairs of E11.0–E12.0 XX and XY gonads (separated from mesonephroi) with the RNeasy Micro kit with on-column DNase digestion (QIAGEN, cat no. 74004) following the manufacturer's protocol. Total RNA was eluted in 14 ul RNase-free water (not DEPC- treated), and 2 ul were used to quantify RNA concentration on a NanoDrop ND-2000 (Thermo Scientific). Only samples with >100 ng of total RNA and an A280∶A260 ratio of >1.6 were included in the expression analyses.
From each total RNA sample, mRNA was selectively reverse transcribed with oligo(dT) primers to T7-labelled cDNA, and then amplified by in vitro transcription (IVT) to produce biotinylated cRNA using the Illumina TotalPrep Amplification Kit (Ambion/Life Technologies, cat no. AMIL1791) according to manufacturer's instructions. cRNA concentration was quantitated on the NanoDrop ND-2000, and as necessary, individual samples were concentrated in a vacuum centrifuge. 750 ng of biotinylated cRNA (in ∼10 ul volume) were hybridized to Illumina MouseRef-8 v2.0 BeadChips (Illumina, cat no. BD-202-0202) according to Illumina protocols, and array intensity was measured on an iScan scanner (Illumina). To minimize potential for batch effects to confound analysis, individual samples were assigned to 8-sample BeadChips using a balanced design.
Microarray data files were imported into GenomeStudio software (Illumina, V2010.1), and raw expression values for each sample extracted. Expression values were quantile normalized and log2 transformed using the R package Beadarray [55]. Probes that had a detection p<0.005 in at least two replicates for any sample type were used for analysis. Data are publicly accessible in GEO (accession number GSE41948).
The ANOVA analysis was conducted using the R package Limma [56]. A sex by strain by stage factorial analysis was conducted as outlined in [57]. The model included the sex, strain, and stage variables, the sex*strain, sex*stage and strain*stage two-way interaction terms, and a three-way interaction term sex*strain*stage. The model was fit for all the probes that had reliable expression (detection p<0.005) in at least two replicates of any one sample using the lmFit function in the Limma package. The statistical significance of each of the terms was evaluated using the eBayes function in Limma. Probes that did not have a significant difference (Benjamini-Hochberg adjusted p<0.05) for at least one of the variables were excluded from further analysis.
Hidden Markov Models (HMMs) are generative probabilistic models that explicitly model the observed data as being emitted by a ‘hidden’ biological state (here, male or female enrichment). Further, transition probabilities between states capture the time dependencies in data between adjacent time points. Inference algorithms allow for computing the most probable state paths that give rise to the observed data, and accounts for noise inherent in observed data. The modeling of time dependencies between biological states, and accounting for noisy observations, makes HMMs particularly well suited to analyze time course microarray data [28], [29].
We designed a left-to-right HMM with three states per time point (Figure 3). The three states correspond to male state (with higher expression in males), female state (with higher expression in females) and similar expression state (with no difference in expression between the two sexes). The observed data on which the model was trained and clustered was the quantized Fold Difference (FD) of the log2 normalized values between XX and XY gonads at each time point. Note that a fold change of expression of 1.25 corresponds to an FD of 0.3219, fold change of 1.5 to an FD of 0.585 and a fold change of 2 to an FD of 1. Limma was used to calculate FD between XX and XY gonads at each time point for each strain. If a specific comparison did not have a p-value<0.05, or |FD|>0.3219 then the FD for that comparison was set to 0. The FD was then quantized into symbols as follows:
Symbol s - Similar expression in XX and XY gonads [−0.3219 < FD < 0.3219].
Symbol m1 - Higher expression in XY gonads [0.3219 < FD < 0.5850].
Symbol m2 - Higher expression in XY gonads [0.5850 < FD < 1].
Symbol m3 - Higher expression in XY gonads [1 < FD]
Symbol f1 - Higher expression in XX gonads [0.3219 > FD > −0.5850].
Symbol f2 - Higher expression in XX gonads [0.5850 > FD > −1].
Symbol f3 - Higher expression in XX gonads [1 > FD].
The symbols m1, m2, m3, and f1, f2, f3 indicate varying levels of confidence in the differential expression between XX and XY gonads.
For each gene, for each strain there were 6 symbols indicating the FD between XX and XY gonads across the time window. For example, for Sox9 in the 129S1 strain, the following FDs were observed at the 6 time points – 0, 0.77, 1.41, 2.41, 2.30, and 1.97. Following the rules listed above, this was quantized as s, m2, m3, m3, m3, m3.
The emission probabilities of the HMM were initialized as shown in Figure S2B to reflect the intuitive meaning of the states and the possible observed symbols from each state. Note that all probabilities were initialized as being non-zero. After training was completed, emission probabilities still reflected the intuitive meaning of the states (Figure S2B).
The transition probabilities between states were initialized as follows (Figure S2A). Observed symbols were first classified into male, female and similarly expressed – symbols m1, m2, m3 into state M, symbols f1, f2, f3 into state F, symbol s to state S. Transitions between all combinations of states in adjacent time points were counted and normalized to make transition probabilities from each node sum to 1. A pseudocount of 1 was added to all possible transitions (transition between states in adjacent time points) to initialize all probabilities as non-zero.
The HMM was trained using the Baum-Welch algorithm for 200 iterations with data from both strains for all the probes that passed the filtering criteria and were shown to have a significant effect for at least one variable in the ANOVA analysis. The state path for the observed FDs for each of the probes was computed using the Viterbi algorithm. Probes with the same state paths were clustered together.
Pre-validated gene-specific MISSION shRNA clones (Sigma Aldrich; Sox9 pLKO.1 clones: TRCN0000086165, TRCN0000086167; Lmo4 pLKO.1 clones: TRCN0000084373, TRCN0000084375; Nontargeting Controls – TurboGFP shRNA SHC004, eGFP shRNA SHC005) and lentiviral packaging and envelope plasmids (Addgene; pCMV-dR8.2 dvpr ID# 8455, pMD2.G ID# 12259) were purchased as bacterial stocks, and high quality plasmid DNA was isolated from overnight liquid LB cultures with a Maxiprep kit (QIAGEN, cat no. 12162) following manufacturer's instructions and quantitated on a NanoDrop ND-2000.
Lentivirus production followed the Addgene 4-day protocol with slight modifications (www.addgene.org/tools/protocols/pLKO/) [58]. All work with lentiviruses was performed in a BSL2+ hood following approved biosafety procedures. On Day 1, for each sample, 5×106 HEK-293T/17 cells (ATCC cat no. CRL-11268) were suspended in 10 ml of Dulbecco's Modified Eagle Medium (DMEM, Gibco cat no. 11995) +10% Fetal Bovine Serum (FBS) without antibiotics, plated to 10 cm cell culture plates, and incubated at 37°C, 5% CO2 overnight. Late in the afternoon of Day 2, 10 ug of pLKO.1 shRNA plasmid, 7.5 ug of pCMV-dR8.74 dvpr packaging plasmid, and 2.5 ug of pMD2.G envelope plasmid DNA were suspended in Opti-mem serum-free medium with 60 ul X-tremeGENE HP DNA transfection reagent (Roche, cat no. 06 366 236 001) in a 3∶1 ratio to a total volume of 600 ul, incubated at 25°C for 20 minutes, then applied drop-wise to the 10 cm plate containing HEK-293T/17 cells at 60–80% confluency, swirled gently to disperse evenly but not dislodge cells from the plate, and incubated at 37°C, 5% CO2 overnight (12–18 hours). On day 3, media containing the transfection reagent was removed carefully and decontaminated in >10% bleach. Next, 5.5 ml of fresh viral growth medium (vGM, containing Neurobasal medium (Gibco, cat no. 21103-049) supplemented with 10% FBS, 0.5 mM L-glutamine (Gibco, cat no. 25030-149), and 1× Antibiotic-Antimycotic (Gibco cat no. 15240-062)) was added carefully to the side of the plate so as not to disturb the transfected virus-producing cells, and incubated at 37°C, 5% CO2 overnight. Late in the afternoon of day 4, the virus-containing vGM was harvested with a 10 ml syringe, and filtered through a 0.45 um PES syringe filter (Whatman, cat no. 6780-2504) into sterile 2.0 ml polypropylene cryo-vials. Viral media was stored at 4°C for use within 5 days, or at −80°C for long-term storage. All laboratory materials that came into contact with viral particles were treated as biohazardous waste and autoclaved according to BSL2+ safety practices.
The effect of silencing candidate regulatory genes was assayed in dissociated gonad primary cell cultures. Timed matings were established for CD-1 mice, and embryos were collected from dams at E12.5. Gonads were dissected away from the attached mesonephroi, sexed by visual inspection for testis cords, and XY gonads from a litter were counted and pooled. Pooled XY gonads were then dissociated in 0.25% Trypsin-EDTA (1×, Gibco cat no. 25200-056) for 15 minutes at 37°C with slight agitation, followed by centrifugation at 4000 rpm for 5 minutes, washed once with DMEM (Gibco cat no. 11965) followed by centrifugation, and suspended in Opti-Mem (Gibco cat no. 11058-021) supplemented with 1% FBS. Cells from one pair of XY gonads were determined to be sufficient for one well of a 24-well culture plate, and the amount of suspension liquid was calculated by multiplying 250 ul by the number of pairs of XY gonads in the pooled sample.
Following the dissociation and wash steps, 250 ul gonad primary cells were immediately added to individual wells of a 24-well cell culture plate, and 250 ul of the appropriate lentivirus-containing vGM was added to each well in a BSL2+ hood. In addition to wells designated to assay target gene shRNA-mediated knockdown, separate wells containing XY gonad primary cells from the same litter were infected with the non-targeting eGFP shRNA (SHC005) and/or TurboGFP shRNA (SHC004) controls. Plates were incubated at 37°C 5% CO2 for 68–72 hours and cell viability was monitored daily with a light microscope. Virus production could be monitored visually for the TurboGFP control infected cells using a fluorescence microscope.
Following 68–72 hours incubation, total RNA was isolated from shRNA lentivirus-infected gonad primary cells using Trizol reagent (Life Technologies, cat no. 15596-018). Briefly, lentivirus-containing culture media was first removed from each well and disposed in bleach. Next, 400 ul of Trizol was added to the adherent cells in each well, allowed to sit at room temperature for 3–5 minutes, after which the lysate was transferred to 1.5 ml microcentrifuge tubes. Subsequent RNA isolation steps follow Munger et al. [1]. Total RNA was quantified on a NanoDrop ND-2000, treated with DNaseI (Life Technologies, cat no. 18068-015), and converted to cDNA using the iScript cDNA synthesis kit (Bio-Rad, cat no. 170-8891) following manufacturer's instructions.
Gene expression was quantified by quantitative RT-PCR (qRT-PCR) on a StepOnePlus Real-time PCR System (Life Technologies). For qRT-PCR, each analysis was performed in technical triplicate in a total volume of 20 ul reaction mix containing 2 ul cDNA template, 4 ul 1 uM gene-specific forward and reverse primers, 10 ul 2× Quantace SensiMix SYBR (Bioline, cat no. QT615-02), and 4 ul RNase-free water. The list of qRT-PCR primers can be found in Table S3; most have been previously published [1], [12]. All primer sets were tested for efficiency and found to work optimally with the ΔCt method [59]. Within a sample, target gene Ct thresholds value were determined and normalized to Gapdh. Differences between target gene shRNA and non-targeting control shRNA samples were compared using the ΔΔCt method as described previously [59]. Significance of expression differences between samples was assessed using a t-test.
|
10.1371/journal.pbio.3000077 | Catalytic mechanism of the tyrosinase reaction toward the Tyr98 residue in the caddie protein | Tyrosinase (EC 1.14.18.1), a copper-containing monooxygenase, catalyzes the conversion of phenol to the corresponding ortho-quinone. The Streptomyces tyrosinase is generated as a complex with a “caddie” protein that facilitates the transport of two copper ions into the active center. In our previous study, the Tyr98 residue in the caddie protein, which is accommodated in the pocket of active center of tyrosinase, has been found to be converted to a reactive quinone through the formations of the μ-η2:η2-peroxo-dicopper(II) and Cu(II)-dopasemiquinone intermediates. Until now—despite extensive studies for the tyrosinase reaction based on the crystallographic analysis, low-molecular-weight models, and computer simulations—the catalytic mechanism has been unable to be made clear at an atomic level. To make the catalytic mechanism of tyrosinase clear, in the present study, the cryo-trapped crystal structures were determined at very high resolutions (1.16–1.70 Å). The structures suggest the existence of an important step for the tyrosinase reaction that has not yet been found: that is, the hydroxylation reaction is triggered by the movement of CuA, which induces the syn-to-anti rearrangement of the copper ligands after the formation of μ-η2:η2-peroxo-dicopper(II) core. By the rearrangement, the hydroxyl group of the substrate is placed in an equatorial position, allowing the electrophilic attack to the aromatic ring by the Cu2O2 oxidant.
| Tyrosinase is an enzyme that controls a rate-limiting reaction of melanogenesis: it catalyzes the conversion of a phenol to the corresponding ortho-quinone. Streptomyces tyrosinase is formed as a complex, with a “caddie” protein that assists with the transport of the two copper ions into the enzyme’s active center. In our previous study, we showed that the Tyr98 residue in the caddie protein, which is accommodated in the pocket of active center of tyrosinase, is converted to a reactive quinone through the formations of the μ-η2:η2-peroxo-dicopper(II) and Cu(II)-dopasemiquinone intermediates. Until now—despite extensive studies of the tyrosinase reaction based on the crystallographic analysis, low-molecular-weight model systems, and computer simulations—the catalytic mechanism was unclear at an atomic level. To understand the catalytic mechanism of tyrosinase in detail, we determined the cryo-trapped crystal structures at very high resolutions, which suggest an important new step for the tyrosinase reaction: the hydroxylation reaction triggered by the movement of CuA, which induces the syn-to-anti rearrangement of the copper ligands after the formation of μ-η2:η2-peroxo-dicopper(II) core.
| Tyrosinase (EC 1.14.18.1), which has an active center formed by dinuclear copper, catalyzes the conversion of phenol to the corresponding ortho-quinone through the hydroxylation and subsequent oxidation reactions, together with the oxidation of catechol to the quinone [1–6] (Fig 1). The quinone product is a reactive precursor to synthesize melanin. A series of reactions is coupled with reduction of dioxygen to water.
Tyrosinase is a type 3 copper protein family including catechol oxidase [7] and hemocyanin [8]. Although the former enzyme oxidizes catechol to the corresponding quinone, it lacks a monooxygenase activity. On the other hand, hemocyanin acts as a dioxygen carrier in arthropods and mollusks. The dicopper center of the type 3 copper protein takes three redox forms [1–6]. The deoxy form [Cu(I)–Cu(I)] has two cuprous ions into the active center, which binds dioxygen to yield the oxy form. In the oxy form, dioxygen binds as a peroxide ion in a μ-η2:η2 side-on bridging mode [Cu(II)–O22-–Cu(II)]. The met form [Cu(II)–Cu(II)] denotes a state in which copper atoms at the active site are oxidized, but dioxygen is not bound to the copper atoms. As for tyrosinase, the met form is a resting enzymatic form, in which two cupric ions are bridged with one or two small ligands, such as water molecules or hydroxide ions. The oxy form catalyzes the conversion of the phenol and catechol substrates to ortho-quinones, whereas the met form does not catalyze the former reaction containing an oxygenation step [1].
Our group has previously cloned a melanin-synthesizing gene cluster from the Streptomyces (S.) castaneoglobisporus HUT6202, which produces a large amount of melanin [9]. The cluster is composed of two cistrons: one is an open reading frame consisting of 378 nucleotides and designated as orf378. The other gene designated tyrC, which is located just downstream of orf378, encodes tyrosinase. Because the orf378 gene product facilitates the incorporation of copper ions to the apo-tyrosinase, we named it as a “caddie” protein. As observed in the case of the S. antibioticus tyrosinase and its partner protein, MelC1 [10], the Cu(II)-free tyrosinase forms a complex with the caddie protein [11]. Although the tyrosinase is not activated by copper added from the outside, the addition of copper to the complex facilitates the incorporation of two copper ions into tyrosinase. Furthermore, the resulting Cu(II)-bound tyrosinase is liberated from the complex, whereas the released caddie protein is not detectable in a solubilized fraction, suggesting that the released caddie molecules form aggregation.
We have determined the tertiary structure of the S. castaneoglobisporus tyrosinase in a complex with the caddie protein at very high resolutions [12] (Fig 2). This is the first determination of the crystal structure of tyrosinase, demonstrating its structural similarity with the catechol oxidase previously determined (Protein Data Bank [PDB] ID: 1BT1) [13]. The crystal structure of the Cu(II)-free tyrosinase in the complex with the caddie (PDB ID: 1WXC) was determined at 1.20-Å resolution. We have obtained the met forms of Cu(II)-bound tyrosinase complexed with the caddie protein by soaking the native crystals in a CuSO4-containing solution. At the active center of tyrosinase, each of two closely positioned copper ions (CuA and CuB) is surrounded by three histidine residues through the Nε nitrogen atoms. The met1 form (PDB ID: 1WX3 at 1.33-Å resolution), which was obtained by soaking for about 40 h, has a molecule containing one oxygen atom between the copper ions, whereas the met2 form (PDB ID: 2AHK at 1.71-Å resolution), which was obtained by soaking for longer than 80 h, has two molecules. In the crystal structure of the complex [12], the Tyr98 residue of the caddie protein is present in the active-site pocket of tyrosinase (Fig 2). In recent years, several crystal structures of tyrosinase-related enzymes have been elucidated, e.g., prophenoloxidases from the insect Manduca sexta (PDB ID: 3HHS) [14] and Anopheles gambiae (PDB ID: 4YZW) [15], tyrosinases from the bacterium Bacillus megaterium (PDB ID: 3NM8) [16] and the fungus Agaricus bisporus (PDB IDs: 2Y9X and 4OUA) [17,18], and protyrosinase (MelB, PDB ID: 3W6Q) [19] and small catechol oxidase (PDB ID: 4J3P) [20] from the fungus Aspergillus oryzae. Additionally, the crustacean and plant enzymes with tyrosinase-like activity have been crystallographically characterized (PDB IDs: 2P3X, 3WKY, 4Z0Y, 5CE9) [21–24], as has the X-ray structure of human tyrosinase-related protein 1 (TRP1) (PDB ID: 5M8T) [25].
At the dicopper center of tyrosinase from S. castaneoglobisporus, CuA is surrounded by His38, His54, and His63 residues, whereas CuB is surrounded by His190, His194, and His216 residues [12] (Fig 2A and 2B). In the absence of copper ions, the His54 residue takes two conformations that are suggested to be important for the copper acquisition [26]. When the side chain of His54 is oriented toward the CuA-binding site, seven water molecules may be present in the active center (Fig 2C). Four of the water molecules (Wat4–Wat7) are aligned between the side chain of Asn191 and the main-chain carbonyl of Asp45. The Wat4 molecule forms hydrogen bonds with the side-chain atoms of Glu182 and Asn191. The Glu182 residue is well conserved among tyrosinase enzymes, whereas the Asn191 is less conserved. A recent example is provided by two tyrosinases from Malus domestica that present an alanine or a glycine at the position corresponding to Asn191 [27]. However, it was reported that the replacement of the residue corresponding to Asn191 to glycine significantly reduced the tyrosinase activity [28]. The side chain of the caddie Tyr98 residue forms hydrogen bonds with Wat2 and Wat3. Wat1 exists between the side chain of His38 and the main-chain carbonyl of Gly204. On the other hand, when the side chain of His54 protrudes toward that of the surface residue in the caddie protein, six water molecules are present in the active center (Fig 2D). In detail, two waters (Wat5 and Wat6) are removed to avoid the close contact with the side chain of His54. Instead, Wat8 is introduced between Wat3 and Wat4. In the met2 form, two copper ions are present at the CuA-2- and CuB-2-binding sites at a distance of 3.4 Å, with two bridging molecules. The bridging molecules (presumably two hydroxide ions) are positioned at the Wat3 and Wat8 sites. The His54 residue is disordered even in the met2 form, probably because of the steric hindrance between His54 and Wat8 [26]. In addition, the Wat1 and Wat2 molecules were found to disappear from the active center, and the side chains of His38 and caddie Tyr98 were altered to interact directly with Gly204 and Ser206, respectively. Although the functional meaning of the disappearance of water molecules, which is coupled with the introduction of copper ions, is currently unclear, it may be an advantage in the entropic energy term.
In the previous study, we have discussed the transferring mechanism of Cu(II) ions to the active center of tyrosinase, which is assisted by the caddie protein, on the basis of the kinetic and crystallographic studies [26]. The binding sites for the additional copper ions (CuC, CuD, and CuE) in the caddie protein and the hydrogen-bonding network around the tyrosinase active site were found to be important for the effective transfer of Cu(II). Our group has recently demonstrated that the incorporation of copper ions into tyrosinase and the following release of copper-bound tyrosinase progress more quickly in the presence of NH2OH, which can reduce the met form to the deoxy form, but not the oxy form, under aerobic conditions [29]. Cu(I), but not Cu(II), must be suitable species to be incorporated into the active center of tyrosinase. Furthermore, the mass spectroscopic analysis has indicated that the Tyr98 residue in the caddie protein is converted to the reactive dopaquinone, which stimulates the aggregation of the caddie protein and the dissociation of tyrosinase from the complex. The dopaquinone must be formed as a result of the catalytic activity of the oxy-tyrosinase. The ultraviolet-visible (UV-vis) and resonance Raman spectroscopic analyses indicated that the Tyr98 residue is converted to dopaquinone through the formations of μ-η2:η2-peroxo-dicopper(II) and Cu(II)-dopasemiquinone intermediates [29], although the formation of dopaquinone is a speculation from the fact that the modified caddie is easily aggregated. Reaction intermediates were able to be trapped under the conditions at which the aggregation of the caddie was inhibited.
Until now—despite extensive studies based on the crystal structures of tyrosinase [12,15,23,24,28,30–33], low-molecular-weight model systems [4–6,34–39], or computer simulations [40–42]—its catalytic mechanism has not yet been clearly understood at an atomic level. For instance, we need to understand about the oxidation states of copper ions, the bases in the tyrosinase reaction, and the lack of tyrosine hydroxylase activity in catechol oxidases. To understand the catalytic mechanism, in the present study, we analyzed time-resolved X-ray crystal structures of the complex between tyrosinase and caddie after the addition of a reducing agent under aerobic conditions.
In the present study, the deoxy-tyrosinase complexed with the caddie protein (ST1) was obtained by soaking the met2-form crystal in a purged solution containing NH2OH for 2 h anaerobically at 25 °C (Table 1). The use of synchrotron radiation improved the resolution, when compared with the deoxy form reported previously (PDB ID: 2AHL at 1.60-Å resolution) [12]. Similar to the results obtained previously, two copper ions are at the CuA-1- and CuB-1-binding sites at a distance of 4.3 Å in ST1 (Fig 3A). The bridging molecule, which may be a water molecule, is positioned at the Wat3 site. Both CuA-1 and CuB-1 take a trigonal coordination with three Nε atoms from the histidine residues in each, rather than a tetrahedral one, since the bridging molecule is somewhat distant from both CuA-1 and CuB-1 (S1A Fig).
The structure of the oxy-tyrosinase was determined using the crystal of tyrosinase complexed with the caddie Y98F mutant, in which the Tyr98 residue is replaced with phenylalanine (ST2 in Table 1). Prior to the data collection, the crystal was soaked in a CuSO4-containing solution for 80 h and then in a NH2OH-containing solution for 2 h under aerobic conditions at 25 °C. In the structure, electron densities for both CuA and CuB are elongated (Fig 3B). In addition, an elongated density, which can be assigned as a peroxide ion, was observed between CuA and CuB. These findings suggest that the structure represents the mixture of deoxy form (65%)—in which two copper ions are positioned at the CuA-1 and CuB-1 sites—and oxy form (35%)—in which two copper ions are positioned at the CuA-2 and CuB-2 sites (Table 2). We have previously reported an oxy-form structure (PDB ID: 1WX2 at 1.80-Å resolution), which had been prepared by the addition of H2O2 to the met2-form crystal [12]. However, the structure could not be determined at a high resolution, probably because the crystal had been seriously injured by the reagent. Changing the method for preparation of the oxy form and using the synchrotron radiation made it possible to determine the structure at a high resolution, although the occupancy of peroxide was low. Both CuA-2 and CuB-2 take the monopyramidal tetragonal coordination preferred by Cu(II), similar to the met2 form. The oxy form is in a syn arrangement, in which axial ligands of CuA-2 and CuB-2 are His63 and His216, respectively. The His63 residue is weakly associated to CuA-2, with a long coordination bond distance (2.5 Å).
In most of the synthetic models of μ-η2:η2-peroxo-dicopper(II), the Cu2O2 core is planar. In contrast, the current oxy form exhibits a bent-butterfly structure in the Cu2O2 core, where the midpoint between two peroxide oxygen atoms is above the midpoint of CuA-2 and CuB-2 (Fig 3B). Considering that Cu(II) prefers the planar coordination, the bent-butterfly Cu2O2 structure is less solid than the planar one [3]. The bent structure, which was also observed in the oxy-form structure reported previously [12], has been suggested to be formed by the hydrogen bond interaction between the hydroxyl of the caddie Tyr98 and the bridging peroxide. However, the hydroxyl group is absent in the current structure. The tyrosinase may be suitable to take a flexible bent-butterfly structure in the oxy form to allow the conformational change during the reaction, as proposed by the other research group [24].
To visualize the process in the tyrosinase reaction toward the caddie Tyr98 residue, we determined the crystal structures, each of which was cryo-trapped after the aerobic soaking of the crystal of met2-tyrosinase complexed with the caddie protein in a buffer containing CuSO4 and NH2OH for a given time at 25 °C (ST3 to ST5 in Table 1). After soaking the crystal for 10 min at 25 °C (ST3), the electron density map indicates the coexistence of the deoxy and the oxy forms in the crystal (Fig 3C), as observed in ST2. The anomalous difference Fourier map suggests another copper-binding site, which is approximately equidistant from CuA-2 and the hydroxyl of the caddie Tyr98, although the electron density is very weak (Fig 3C). This is in striking contrast to the results obtained from ST2, at which the anomalous difference Fourier map and Fo-Fc map did not show any signal at that position. The density at this site becomes strong in the other structures, as described below. Additionally, using the diffraction data of another crystal collected at the wavelengths of 1.35 and 1.40 Å, we confirmed that the position was occupied by the copper atom. Hereafter, we refer to copper observed at the new position as CuA-3. The information on the important distances in the two possible oxy-form structures is shown in S1B and S1C Fig. At the CuA-3 position, a new coordination bond is formed with the hydroxyl of the caddie Tyr98 residue, whereas the coordination bond with the His63 residue is completely lost. In ST3, occupancies of CuA-1, CuA-2, and CuA-3 were calculated to be about 0.6, 0.3, and 0.1, respectively, whereas occupancies of CuB-1 and CuB-2 were about 0.6 and 0.4, respectively (Table 2).
In the crystal structures obtained at 20 min (ST4) and 2 h (ST5) after the aerobic addition of NH2OH, the electron density at the CuA-3 site is stronger than that obtained at 10 min (Fig 3D and 3E). The high occupancy of CuA-3 was also suggested by the anomalous difference Fourier map (Fig 4A). The electron density maps also indicate that the side chains of the His38 and His54 residues clearly take two different conformations. One conformation is suitable for the coordination to CuA-1 and CuA-2 and, the other is suitable for the coordination to CuA-3. Although the flexibility of the His54 residue was recognized by early studies [12,26], the His38 residue also has the flexibility to adapt to the movement of CuA. The flexibility of the His38 residue is enabled by the removal of Wat1.
In ST5, occupancies of CuA-1, CuA-2, and CuA-3 were calculated to be about 0.2, 0.4, and 0.4, respectively, whereas occupancies of CuB-1 and CuB-2 were about 0.2 and 0.8, respectively (Table 2). This result implies that, when CuA occupies the CuA-3 site, CuB is positioned at the CuB-2 site. The electron density at the bridging position between the two copper ions is also elongated (Fig 3E), indicating the heterogeneity at this site. The major bridging molecule (80%) seems to correspond with a molecule containing one oxygen atom (water or hydroxide ion) positioned at the Wat3 site. The minor molecule (20%) seems to be peroxide, although the binding mode is different from that in the above-mentioned oxy forms (Fig 3B to 3D). In detail, one oxygen atom in the peroxide exists at a different position, where it can form a hydrogen bond with the Nε atom of the His63 residue (S1D Fig).
Furthermore, in ST5, clear electron densities were found around the Cε2 atom of the caddie Tyr98 residue (Figs 3E and 4B), indicating that the reaction proceeded even in the crystalline state. The density in this case corresponds to an oxygen atom with the occupancy of about 0.6 (Table 2). The newly added oxygen (Oζ2) is within the coordination bond distance from CuA-3 and near to CuB-2 (S1E Fig). The complex between CuA-3 and the oxygenated Tyr98 may correspond with the Cu(II)-bound dopasemiquinone observed in the solution state [29]. Specifically, CuA-3 is in a bipyramidal trigonal coordination cage, in which the axial ligands are the Oζ2 atom added to the caddie Tyr98 and the Nε atom of His38, and equatorial ligands are the Oη atom of the caddie Tyr98, the Nε atom of His54, and the bridging oxygen atom at the Wat3 site. The Nε atom of His63 is not coordinated with CuA, but it is within the hydrogen-bonding distance of one of the peroxide oxygens (2.8 Å) and Wat3 (3.4 Å) (S1D and S1E Fig). The structural refinement suggests that the occupancies of CuA-2 and CuB-2 are higher than the values obtained at the earlier times (ST3 and ST4), whereas the occupancy of peroxide is lower (Table 2). Therefore, in ST5, a large part of CuB-2, as well as a part of CuA-2, seems to take tetrahedral coordination, which is preferred by Cu(I), with three Nε atoms from the histidine residues and one oxygen molecule at the Wat3 site in each (CuB-2 in S1E Fig and both CuA-2 and CuB-2 in S1F Fig). Atomic models of the active site in ST3 and ST5 are shown in Fig 5C and 5D, respectively, together with those in the Cu(II)-free form (Fig 5A) and the met2 form (Fig 5B).
In crystallography, refinements of both the occupancy and the temperature factors of atoms are difficult. In the present case, each crystal is considered to contain intermediates in a different ratio. Therefore, occupancies were refined using the following restraints. When CuA occupies the CuA-1, CuA-2, and CuA-3 sites, CuB is likely positioned at the CuB-1, CuB-2, and CuB-2 sites, respectively. In addition, when CuA occupies the CuA-3 site, the His38 and His54 residues seem to take the minor conformations. Therefore, the occupancy of CuA-1 was set to equal that of CuB-1, and the sum of occupancies of CuA-2 and CuA-3 was set to equal that of CuB-2. The occupancies of the major and minor conformations of the His38 and His54 residues were set to equal the sum of occupancies of CuA-1 and CuA-2 and the occupancy of CuA-3, respectively. The hydroxylation reaction must proceed after the binding of oxygen to the deoxy form, where the Wat3 atom is positioned between the CuA-1 and CuB-1 sites. However, after the oxygenation, one of the peroxide oxygens is attached to the Cε2 atom of the Tyr98 residue, whereas the other oxygen seems to occupy the Wat3 site. Therefore, the sum of occupancies of peroxide and the Oζ2 atom added to Tyr98 was also set to equal that of CuB-2, and the sum of occupancies of peroxide and one oxygen atom at the Wat3 site was set to equal 1. Temperature factors of a copper ion and its ligands were refined to become similar values using DELU and SIMU restraints in the SHELXL-97 program [43]. Refined occupancies and equivalent B-factors are shown in Table 2 and S1 Table, respectively. In ST3 and ST4, the occupancy of the Oζ2 atom is lower than that of CuA-3, whereas in ST5, the occupancy of the Oζ2 atom is higher than that of CuA-3. These observations suggest that CuA moves to the CuA-3 site prior to the oxygenation reaction, whereas CuA moves back to the CuA-2 site after the reaction.
To investigate whether the movement of CuA between the CuA-2 and the CuA-3 sites is important for the catalytic reaction, a complex between the H63F-mutated tyrosinase, in which the His63 residue is replaced with phenylalanine, and the caddie protein was prepared. The protein complex did not form the aggregate of the caddie protein after the addition of CuSO4 and NH2OH. However, the spectroscopic analysis suggests that the Cu(II)-bound semiquinone complex is formed under the alkaline condition at pH 9, but the formation rate is slow. In this case, the met2 form was unable to be generated in the crystal even after longer incubation with Cu(II), which is sufficient for the wild-type crystal to generate the met2 form, probably because of the defect at the CuA-binding site. Soaking in a buffer containing CuSO4 and NH2OH was necessary to generate the dicopper center. The crystal structure of the H63F-mutated complex was obtained by using a crystal aerobically soaked in a buffer containing CuSO4 and NH2OH for 24 h at 25 °C (ST6 in Table 1). At the active center in the H63F-mutated tyrosinase, CuA is found at one site. The distances between the site in ST6 and CuA-1, CuA-2, or CuA-3 sites in ST5 are 2.7, 1.9 or 0.2 Å, respectively. CuA is localized at the CuA-3 site probably because of the mutation at the His63 residue, which is a ligand of CuA-1 and CuA-2. Based on the temperature factors (S1 Table), the CuA-3 atom and the His54 residue are unstable. However, CuB is found at either the CuB-1 or the CuB-2 site. In contrast to the wild-type complex, the position of CuB seems to be variable when CuA occupies the CuA-3 site. The positional uncertainty of CuB may depend on the structural uncertainty at the Tyr98 residue, although the correct reason is unknown at this time. The sum of occupancies of CuB-1 and CuB-2 was calculated to be approximately 1.0, whereas the occupancy of CuA-3 was approximately 0.9 (Table 2), suggesting the slight incompleteness of the copper uptake. On the other hand, although a part of the caddie Tyr98 residues seems to be converted to dopasemiquinone, the occupancy of the Oζ2 atom was calculated to be 0.30 (Table 2). In this case, a part of CuA-3 is ligated to the oxygenated Tyr98 residue, whereas the rest is ligated to the unmodified one. This is similar to the results obtained using the wild-type crystal at the early stage, as the occupancy of the Oζ2 atom is lower than that of CuA-3, indicating that the movement of CuA to the CuA-3 site occurs prior to the hydroxylation reaction. In addition, it is thought that CuA was unable to move to the CuA-2 site after the reaction because of the impairment of the site, which might inhibit the progress of the reaction toward dopaquinone and thereby inhibit the aggregation of the caddie protein.
In the previous study [29], we demonstrated that the addition of NH2OH stimulates the caddie proteins to aggregate, resulting in the release of tyrosinase from the complex. The aggregation is likely triggered by the formation of reactive dopaquinone on the caddie Tyr98 residue. The UV-vis and resonance Raman spectroscopic analyses indicate that the Tyr98 residue is converted to a reactive quinone through the formations of the μ-η2:η2-peroxo-dicopper(II) and Cu(II)-dopasemiquinone intermediates. It is important to note that intermediates after the μ-η2:η2-peroxo-dicopper(II) generation have not been trapped when adding the poor substrate (3,5-difluorophenol) to the S. antibioticus tyrosinase [44], which shares high similarities with the S. castaneoglobisporus tyrosinase used in the present study. The caddie Tyr98 residue may be a poorer substrate than 3,5-difluorophenol, enabling detection of the intermediates.
After the crystal structure of tyrosinase was shown by our group [12], its catalytic mechanism has been actively discussed by other groups [4,6,15,23,24,27,28,30–33,40–42]. Since tyrosinase can react with the Tyr98 residue of the caddie protein [29], the Tyr98 residue is expected to adopt a similar binding position of L-tyrosine as a genuine substrate of tyrosinase. However, free L-tyrosine, which lacks the structural restraints as compared with the Tyr98 residue as a part of the caddie protein, may be bound deeply into the active-site pocket. The hydroxyl group of the caddie Tyr98 residue interacts with the hydroxyl group of Ser206, and the phenol ring has a stacking interaction with the imidazole ring of His194 (Fig 2B). Therefore, when a genuine substrate is bound to the active center of the Streptomyces tyrosinase, it must interact with Ser206 and His194. The computer simulation analysis also suggests that the interactions with Ser206 and His194 are important for the binding of kojic acid, which is a tyrosinase inhibitor, to the active center of the Streptomyces tyrosinase [42]. However, the Ser206 residue is not conserved in tyrosinases from other microorganisms.
As the first step in the hydroxylation reaction of tyrosinase, in general, the substrate hydroxyl was assumed to bind directly to one of the two copper ions in the oxy form. However, since the hydroxyl group of the Tyr98 residue has no direct interaction with the copper atoms in the starting oxy form, the movement of the Tyr98 residue and/or the structural change of the active center of tyrosinase must occur prior to the hydroxylation reaction. Another research group has already anticipated that genuine substrate binds to tyrosinase in a manner similar to the caddie Tyr98 residue [4,30]. They insisted that the substrate must shift toward CuA-2 from the position of the caddie Tyr98 residue to form a coordination bond. In the crystal structure of the Bacillus tyrosinase complexed with tyrosol (PDB ID: 4P6T) [31], the substrate was found at the position about 2 Å from that of the caddie Tyr98 residue, and the hydroxyl is bound at the axial position of CuA. Additionally, in the case of the B subunit, CuA has a very long coordination distance with the residue corresponding to His63 upon the binding of the substrate. Because of positional restriction by the surrounding residues, the binding position of the caddie Tyr98 residue may be slightly different from that of the small substrate, resulting in emphasis of the importance of the CuA movement. In addition, our crystallographic results (ST3, ST4, and ST6) indicated that the movement of CuA to the CuA-3 site occurs prior to the hydroxylation reaction. In total, CuA-3 seems to be the functional site but not the artifact one occupied only in the product-bound state.
From the current data, we propose a catalytic mechanism of tyrosinase toward the caddie Tyr98 residue as shown in Fig 6. In the deoxy form as a starting point (Fig 6A), two copper ions are located at the CuA-1 and CuB-1 sites. When the dioxygen is bound to the deoxy form, CuA and CuB move toward the CuA-2 and CuB-2 sites, respectively. The movement is triggered by the bonding interactions between dioxygen and Cu(I) atoms, together with the change in oxidation states from Cu(I) (which prefers trigonal or tetrahedral coordination) to Cu(II) (which prefers tetragonal or monopyramidal tetragonal coordination). In the oxy form, two oxygens of dioxygen are positioned at the Wat3 and Wat8 sites, resulting in the formation of a μ-η2:η2-peroxo-dicopper(II) with a bent-butterfly structure (Fig 6B).
Deprotonation of the substrate hydroxyl is thought to be important for the hydroxylation reaction of tyrosinase [34]. Based on the observation that hydroxylation of the caddie Tyr98 does occur in the crystalline state at pH 6.5, a base must be present in proximity to the hydroxyl group. The base that deprotonates the hydroxyl has been under debate [4,6,12,15,23,24,27,28,31,33,40]. Although the recent study suggests that Wat4, which forms hydrogen bonds with the Glu182 and Asn191 residues, acts as a base [6,15,23,28,31], it is far away from the hydroxyl oxygen of the caddie Tyr98 residue (5.9 Å) (S1B Fig). Additionally, in contrast to the cases of large tyrosinases, since the substrate-binding pocket of the small Streptomyces tyrosinase is directly exposed to the solvent region in the absence of the caddie protein (Fig 2A), there may be no different binding positions for the substrate. Therefore, deprotonation of the substrate is unlikely to occur at the entrance or during the preorientation by the second-shell residues located near the active site, as proposed by other groups [15,23,33]. In the crystal structure of oxy-tyrosinase complexed with the caddie protein, the hydroxyl group is positioned near the two peroxide oxygens within a distance of 3.5 Å (S1B Fig). This indicates that the hydroxyl proton can easily move to the peroxide. Studies using model systems with low molecular weight [4,34] suggest that the neutral substrate is difficult to be hydroxylated. The difficulty is explained as follows: after the binding of the neutral substrate to the dicopper center and the subsequent transfer of a proton from the substrate to peroxide, one electron is transferred from the substrate to one of the two copper ions, leading to the formation of C–C coupled dimer products, like a tyrosine dimer. Considering from a different perspective, protonation of the dicopper center may diminish or largely decrease the hydroxylation activity, which affords the side reaction to generate the C–C bond. On the other hand, tyrosinase has not been reported to generate the C–C coupled dimer, probably because of the high reaction rate of the enzyme or of the situation of the substrate in the active-site pocket, which prevents the dimer formation.
It should be noted that a coordination bond between CuA and His63 is completely lost after the movement of CuA to the CuA-3 site and that the distances between the peroxide oxygens and the Nε atom of His63 are in the range of 3.5 and 4.0 Å (S1B Fig). This histidine flexibility opens the opportunity for the imidazole to serve as a base to deprotonate the phenol substrate. Additionally, a recent study using a small-molecule model suggests that the copper ligand acts as an internal base for the substrate hydroxyl, since the addition of excess copper ligand enables the oxygenase reaction toward protonated phenol [45]. Therefore, we assume that a proton from the Tyr98 hydroxyl moves to the His63 residue via the peroxide. Although there is another possibility that the proton moved to Wat4 via the peroxide, the distance between one of the peroxide oxygens and the Wat4 atom is slightly larger (4.2 Å) than the distance to the Nε atom of His63. The significance of the proton transfer step is partially supported by the results using the H63F-mutated complex. The mutant protein does not actually have reactivity toward the small substrate (L-tyrosine), even under the condition in which the dicopper center could be formed, and the reaction toward the caddie Tyr98 residue was arrested at the Cu(II)-dopasemiquinone intermediate, suggesting that the reaction catalyzed by the mutant is halted at the first turnover. In ST6, the occupancy of the oxygen atom added to the Tyr98 residue is significantly lower than that of CuA-3. In this case, probably because of the lack of the internal base, a large part of the hydroperoxide ion, which was formed between two coppers after movement of the proton from the Tyr98 hydroxyl, would be replaced by a water or hydroxide ion prior to the deprotonation to produce peroxide. However, when the Tyr98 residue was deprotonated beforehand, hydroxylation reaction could proceed. This hypothesis is supported by the observation that the alkaline pH conditions stimulate the change in the UV-vis spectrum, which indicates the accumulation of Cu(II)-dopasemiquinone intermediate. In addition, this result may exclude the possibility that (μ-oxo)(μ-hydroxo)-dicopper(II,III) acts as an active species for the hydroxylation, which was proposed from the simulation analysis [40], at least in our system.
The deprotonation of the hydroxyl seems to have two roles. At first, since the ortho-carbon has a partial negative charge after the deprotonation of the hydroxyl, the atom comes to exhibit a high nucleophilicity. An electrophilic aromatic substitution reaction by the oxy form has been suggested for the hydroxylation mechanism on tyrosinase [34]. The deprotonation of the substrate hydroxyl may also play a role in generating an electrostatic interaction between CuA and the hydroxylate, which induces the movement of CuA to the CuA-3 site (about 1.7 Å). The large movement is not surprising given that the large structural changes are observed to create the side-on peroxide species from the reaction of the reduced enzyme and dioxygen (1.0 and 0.5 Å for CuA and CuB, respectively). The location of CuB is better conserved than that of CuA, which is in agreement with the previous observations [12,26,31] as well as with the recently elucidated crystal structure of Zn(II)-bound TRP1 [25]. The energetic driving force for the movement may be placing the strongest sigma-donating ligand (phenolate) into an equatorial position. Together with the movement of CuA, the side chains of His38 and His54 change conformation to maintain the coordination bond with CuA. The movement of CuA is also found in the crystal structure of the Bacillus tyrosinase complexed with tyrosol [31] and in the simulated structure of the Streptomyces tyrosinase complexed with kojic acid [42], although the movement lengths are shorter than that observed in the present study.
In accordance with the movement of CuA, peroxide must move to a new position. Although the crystal structure is absent, we propose a hypothesis that the peroxide is arranged keeping the μ-η2:η2-binding mode (Fig 6E), which may be useful to destabilize the O−O bond for the reaction. In this putative intermediate, two copper ions are positioned at the CuA-3 and CuB-2 sites, one of two peroxide oxygens (Oproximal) is near the ortho-position of the caddie Tyr98 residue, and the other (Odistal) is at the Wat3 site. The Cu2O2 core lies on a plane created by the Oη atom of the caddie Tyr98 and the Nε atoms of His38, His190, and His216. This intermediate is in an anti-arrangement in which the axial ligands of CuA and CuB are His54 and His194, respectively. Our previous crystallographic studies [12,26] have indicated the flexibility of His54. In addition, the distance from CuB-2 to the Nε atom of His216 is comparable with that to the Nε atom of His194 and is longer than that to the Nε atom of His190 (S1B to S1F Fig). These observations suggest that the axial-to-equatorial exchange of His216 and the equatorial-to-axial exchanges of His54 and His194 would occur. As possible intermediates between Fig 6B and 6E, we propose two structures based on the crystal structures (S1C and S1D Fig), since the different binding modes of peroxide are present in the early stage (ST3 and ST4) and the late stage (ST5). In the first structure (Fig 6C), two copper ions are positioned at the CuA-3 and CuB-2 sites, and peroxide is positioned at the original site, leading to the formation of reverse butterfly structure. The proton from the caddie Tyr98 residue may be attached to peroxide. In the second one (Fig 6D), two copper ions are positioned at the CuA-3 and CuB-2 sites, and peroxide is positioned at the site observed in the later stage. The proton attached to peroxide may interact with the Nε atom of the His63 residue.
Orientation of the substrate to the dicopper center is crucially important for the tyrosinase reactivity. For the hydroxylation reaction, the σ* orbital of the bridging peroxide ligand must overlap with the π orbitals of the substrate. To this end, rotation of the peroxide [4,30], rotation of the substrate [31], and axial-to-equatorial interchanges of the His54 and His63 residues [41] were proposed to occur after the binding of the substrate hydroxylate to CuA. In the present study, we propose that the orientation of the substrate is adjusted by the movement of CuA, which accompanies the syn-to-anti rearrangement of the copper ligands. By the rearrangement, the hydroxyl group of the Tyr98 residue was changed to an equatorial ligand of CuA. After the adjustment, Oproximal attacks the ortho-carbon of the caddie Tyr98 residue electrophilically. The reaction would be followed by the cleavage of the O–O bond and the proton transfer from His63 to the oxide ion derived from Odistal, resulting in the formation of a dienone intermediate with a nonplanar ring (Fig 6F).
The axial-to-equatorial exchange of the substrate hydroxylate was previously proposed for the reaction mechanism of the tyrosinase model with a low molecular weight, in which the μ-η2:η2-peroxo-dicopper(II) catalyzes the conversion of a deprotonated phenol to the Cu(II)-bound semiquinone [35,36,39]. The current study presents the first evidence that the axial-to-equatorial exchange actually occurs in the macromolecular system. The research group demonstrated that an axial-to-equatorial reorientation of the substrate hydroxylate induces cleavage of the O–O bond prior to the oxygenation reaction, resulting in the generation of bis(μ-oxo)-dicopper(III), which has an absorption peak at about 400 nm because of the oxide-to-Cu(III) charge transfer transition [35,36,39]. However, our previous study of the solution state indicated that after the development of the oxy form with a μ-η2:η2-peroxo-dicopper(II) core, the Tyr98 residue of the caddie is converted to dopaquinone via Cu(II)-bound dopasemiquinone [29]. Since Cu(II)-dopasemiquinone and intact dopaquinone also have an absorption peak at about 400 nm, it is difficult to confirm the generation of the cleaved species from the UV-vis spectrum. Furthermore, we could not detect any cleaved species, such as bis(μ-oxo)-dicopper(III), although resonance Raman analysis using a 413-nm laser was conducted intensively under the different temperature conditions [29]. However, since the lifetime of the cleaved species may be too short for detection, we cannot exclude the possibility that O–O bond scission occurs prior to the reaction.
Based on the results obtained in the solution state [29], the crystal structures prepared by soaking for a longer time may contain Cu(II)-dopasemiquinone or dopaquinone in high ratios. In the proposed dopasemiquinone-bound structure (Fig 6G), Cu(II) and Cu(I) are present at the CuA-3 and CuB-2 sites, respectively. Hereafter, we refer to the structure bound with dopasemiquinone as a half-met form. CuA-3 takes the bipyramidal trigonal coordination preferred by Cu(II) next to the tetragonal or monopyramidal tetragonal coordination, whereas CuB-2 takes the tetrahedral coordination preferred by Cu(I). The bipyramidal trigonal coordination was also detected in early studies using a half-met form of the Neurospora tyrosinase complexed with the tyrosinase inhibitors such as mimosine or benzoic acid [2,46,47], although the active site of the Neurospora tyrosinase is different from that of the Streptomyces tyrosinase because of the presence of the cysteine–histidine thioether bond. The η2-semiquinone-bound half-met form may be formed via the η2:η1-catecholate-bound met form with a one-electron reduction of CuB. The η2:η1-catecholate-bound dicopper(II) was previously proposed as an intermediate of the tyrosinase reaction [35–37,48–50]. However, if the η2:η1-catecholate-bound intermediate was formed in our system, the ring of the caddie Tyr98 would become unparallel to the ring of the tyrosinase His194, generating the steric hindrance. Therefore, the deprotonation of the ortho-carbon in the nonplanar intermediate (Fig 6F to 6G) may be coupled with the one-electron reduction of CuB to avoid the formation of the undesirable η2:η1-catecholate-bound intermediate.
For the formation of the G intermediate, another base to abstract a proton from the ortho-carbon of the substrate is needed. Although the deprotonation of the substrate hydroxyl is recognized as an important step, the significance of the base at the second deprotonation step seems to be underestimated by many research groups. As a consensus, the proton is thought to finally move to oxide or hydroxide ion derived from Odistal, resulting in the conversion to a hydroxide ion or a water molecule, respectively. Direct movement of the ortho-carbon proton to the Odistal-derived oxygen is unlikely, since the Odistal-derived atom and the atom added to the substrate are on the same side with respect to the phenol ring. This step has been proposed to be mediated by a base such as His54 [40] or a solvent atom [36]. Since the Nε atom of His54 is distant from the Cε2 atom of the caddie Tyr98 (4.2 Å, S1E Fig), the residue is unlikely to act as a base. Similarly, Wat4, which was recently considered as a hopeful base by other groups [6,15,23,28,31], is also distant from the Cε2 atom of the caddie Tyr98 (4.1 Å). However, adjacent Wat5 is closer to the Cε2 atom (3.6 Å). Therefore, we now consider Wat5 as a candidate base at this step. The generated hydroxonium ion at the Wat5 site may be stabilized by the hydrogen-bonding network to the well-conserved Glu182 residue via Wat4.
Our crystallographic results indicate that CuA moves back to the CuA-2 position after the oxygenation reaction. The movement may be coupled with the one-electron transfer from semiquinone to CuA, resulting in the putative complex between the deoxy form and dopaquinone (Fig 6H). Hereafter, we refer to the structure bound with dopaquinone as a deoxy2 form, in which two cuprous ions are positioned at the CuA-2 and CuB-2 sites. Both CuA-2 and CuB-2 take the tetrahedral coordination preferred by Cu(I) (S1F Fig). The deoxy2 form has a shorter Cu–Cu distance than the starting deoxy form (Fig 6A). This may be due to the difference in bridging molecule. That is, deoxy and deoxy2 forms have a water molecule and a hydroxide ion at the bridging position, respectively. Movement of the proton from Wat5 to the Odistal-derived oxygen may result in the formation of deoxy form (Fig 6I).
The aggregation of the caddie protein seems to progress through the generation of intermolecular linkage between the liberated molecules after the quinone formation. As shown in the previous study [29], the aggregation of the caddie protein was stimulated under the acidic pH conditions probably because of the induction of the release of the caddie from the complex. Since the acidic pH condition rather weakens the linking reaction, the release of the caddie from the complex may be the rate-limiting step for the aggregation. The conversion from the dopaquinone-bound deoxy2 form (Fig 6H) to the deoxy form (Fig 6I) would be stimulated under the acidic pH conditions. The resulting deoxy form may interact with the dopaquinone residue more weakly than the deoxy2 form because of the shift of the copper positions, inducing the release of the quinone-containing caddie protein.
In the catalytic mechanism of tyrosinase generally accepted, the product quinone is separated from the deoxy form, and a new substrate enters the binding pocket. However, the reaction mechanism of tyrosinase in the catalytic cycle may be different from that in the first cycle, as suggested by another group [4]. In this case, it might be unnecessary to develop the deoxy form. That is, the product quinone bound to the hydroxide-bridged deoxy2 form (Fig 6H) is replaced by the next substrate. Then, the bridged hydroxide deprotonates the substrate hydroxyl. The deprotonated substrate is bound to the active center of tyrosinase (Fig 6J) prior to the creation of the oxy form (Fig 6K). If so, the hydroxylation reaction may progress quickly without bases (Fig 6L to 6F), which are currently assumed to be peroxide and His63. In the present proposition, the hydroxonium ion at the Wat5 site, generated after the second deprotonation step (Fig 6F to 6G), may be stabilized by the interaction with the conserved Glu182 residue via Wat4. The stabilization effect may inhibit the proton transfer to the Odistal-derived hydroxide ion and thereby the development of the deoxy form, although it might be not enough to inhibit the transfer to the oxide ion (Fig 6L to 6F). The basicity of Wat4 has recently been considered to be a structural factor to distinguish tyrosinase from catechol oxidase [6,15,23,28,31] rather than the accessibility of the substrate to the active center [2,4,12,51,52]. The stabilization of the H intermediate containing a hydroxonium ion may be an important factor to distinguish tyrosinase from catechol oxidase. That is, tyrosinase can adopt the fast catalytic cycle via the deoxy2 form, whereas catechol oxidase only adopts the slow catalytic cycle via the deoxy form, resulting in the apparent low reactivity toward the phenol compound. This proposition must be verified by the various approaches.
In summary, we can propose the chemically reasonable atomistic postulate of the reaction mechanism of tyrosinase toward the caddie Tyr98 residue, in which the coordination preference determined by the oxidation state of copper and the protonation state triggers the generation of reaction intermediates in order. However, the binding position of the caddie Tyr98 residue may be different from that of a genuine substrate. As proposed by other research groups [4,30,31,41], the hydroxyl group of a substrate is likely bound to the axial position of CuA in the oxy form (CuA-2). Although the proposition seems to be right for the small substrate, we believe that an important step for the tyrosinase reaction is a subsequent syn-to-anti rearrangement of copper ligands. As a result, substrate hydroxyl is bound to the equatorial position of CuA, and peroxide is optimally repositioned to attack the substrate. This rearrangement is enabled by the movement of CuA and the scission of a coordination bond between CuA and His63. The migration length of CuA for the hydroxylation of a genuine substrate may be shorter than that for the hydroxylation of the caddie Tyr98 residue. The hydroxylation reaction would accelerate with the decrease in migration length of CuA. In addition, in the binding position of a small substrate, Wat4 may be close to the peroxide and enable deprotonation from the hydroxyl group of the caddie Tyr98 residue via peroxide. Similarly, His54 or Wat4 may be close to the ortho-carbon of the substrate, enabling deprotonation at the second step. Identification of the actual bases for the small substrate is the next objective to be pursued.
To introduce an H63F mutation in tyrosinase, a QuikChange Site-Directed Mutagenesis Kit (Stratagene) was used. pET-tyrC [11], a plasmid for the expression of the His6-tagged tyrosinase, was amplified using sense primer (5′-CGTTCCTGCCCTGGTTCCGCAGATTCCTG-3′, where underline means the mutation site) and antisense primer (5′-CAGGAATCTGCGGAACCAGGGCAGGAACG-3′). The original plasmid was removed by DpnI digestion. The mutant plasmid was amplified in the Escherichia coli cells, and the introduction of the mutation was confirmed by DNA-sequencing analysis. To generate the plasmid for the coexpression of the H63F-mutated tyrosinase and caddie, the region containing the T7 promoter and mutated tyrC gene was amplified with the forward primer 5′-GCACGCATGCGAAATTAATACGACTCAC-3′ (the underline indicates the SphI site) and the reverse primer 5′-CTATGCATGCCAAAAAACCCCTCAAGAC-3′ (the underline indicates the SphI site) by using the mutated plasmid as a template. The amplified fragment was digested with SphI and inserted into the same site of pET-orf378 [11], a plasmid for the expression of the His6-tagged caddie protein. We chose a plasmid in which the direction of the tyrosinase and caddie genes is opposite.
Plasmid to overproduce the wild-type or Y98F-mutated complex has been already constructed [11,26]. The overproduction and purification of wild-type, Y98F-mutated, or H63F-mutated complex were performed by the method described previously [11].
The crystallization of copper-free tyrosinase in complex with the caddie protein (including mutated complexes) was conducted by the method described previously [12,26]. The typical formula of reservoir solution to obtain the crystals was 25% PEG3350, 0.2 M NaNO3, and 0.1 M Na-HEPES (pH 6.5). Crystallization was performed by the sitting-drop vapor-diffusion method, in which a drop mixing 2 μL of the protein solution, 2 μL of the reservoir solution, and 1 μL of the reservoir solution containing microseeds was kept in a well containing 1 mL of the reservoir solution. For the soaking experiments, crystals with similar sizes (0.3–0.4 mm in one dimension) were chosen. The crystals of a complex between met2-form tyrosinase and wild-type or Y98F-mutated caddie were obtained by soaking the copper-free crystals in a reservoir solution containing 1 mM CuSO4 for about 80 h at 25 °C [26]. The crystals of a complex between deoxy-tyrosinase and the wild-type caddie protein were obtained by soaking the met2-form crystal anaerobically in an N2-purged reservoir solution containing 0.1 mM CuSO4, 10 mM NH2OH, 5 mM glucose, 1 μM glucose oxidase, and 5 μM catalase for 2 h at 25 °C. The latter three reagents were added to keep a low dioxygen concentration. During soaking, the crystal-containing well was filled with the purged solution, and cover glass was put over the well. Other crystals were obtained by soaking the met2-form crystal aerobically in a reservoir solution containing 0.1 mM CuSO4 and 10 mM NH2OH for the indicated times at 25 °C. Exceptionally, to generate dicopper center in the H63F-mutated complex, the copper-free crystals were directly soaked in a reservoir solution containing 1 mM CuSO4 and 10 mM NH2OH for 24 h at 25 °C.
The diffraction intensities of the crystals were collected using synchrotron radiation from the stations BL26B2, BL38B1, or BL41XU at SPring-8, Japan. Some of the crystallographic studies at SPring-8 were performed with the approval of the institution (2013A1078). When using the high levels of X-ray exposure at BL41XU, the occupancies of CuA-1 and CuB-1 were calculated to be higher than the values obtained at BL26B2 and BL38B1, indicating the occurrence of copper reduction by hydrated electrons. Therefore, the crystal structures obtained from BL41XU were excluded from the discussion. The crystals were frozen by liquid nitrogen or by nitrogen gas stream and mounted on the goniometer. Diffraction of each crystal was measured using a CCD camera, and the intensities were integrated and scaled using HKL2000 [53] or using the combination of Mosflm and Scala programs in the CCP4 program suite [54]. The existence of the copper ion was confirmed in anomalous difference Fourier maps. The model was refined using conventional restrained refinement methods with the CNS program [55]. A subset of 5% of the reflections was used to monitor the free R factor (Rfree) [56]. Each refinement cycle included the refinement of positional parameters, individual isotropic B-factors, correction using the flat bulk solvent model, and addition of solvent molecules. After the CNS refinement was converged, the model was further refined by SHELXL-97 [43]. At this stage, anisotropic temperature factors were introduced for all atoms. The distance between the copper ion and its ligand atom was not restrained. At first, occupancies of the copper ions, bridging molecules (peroxide and water), and oxygen atom bound to the caddie Tyr98 were refined independently. With respect to the wild-type or the Y98F-mutated complex, several restraints were imposed on the occupancies. The model was revised using the electron density map visualized by the program Xfit in the XtalView software package [57]. Occupancies of CuA-1, CuA-2, CuA-3, peroxide, and Oζ2 atom added to Tyr98 at the final stage are shown in Table 2. The data collection and refinement statistics are shown in Table 1. The statistics for ST6 are worse than the values for the other five models, probably because of the low ratio of the number of reflections to that of the refined parameters. Other cryo-trapped structures obtained by the soaking at 4 °C are described in the Supporting Information.
|
10.1371/journal.pcbi.1006357 | Flexible resonance in prefrontal networks with strong feedback inhibition | Oscillations are ubiquitous features of brain dynamics that undergo task-related changes in synchrony, power, and frequency. The impact of those changes on target networks is poorly understood. In this work, we used a biophysically detailed model of prefrontal cortex (PFC) to explore the effects of varying the spike rate, synchrony, and waveform of strong oscillatory inputs on the behavior of cortical networks driven by them. Interacting populations of excitatory and inhibitory neurons with strong feedback inhibition are inhibition-based network oscillators that exhibit resonance (i.e., larger responses to preferred input frequencies). We quantified network responses in terms of mean firing rates and the population frequency of network oscillation; and characterized their behavior in terms of the natural response to asynchronous input and the resonant response to oscillatory inputs. We show that strong feedback inhibition causes the PFC to generate internal (natural) oscillations in the beta/gamma frequency range (>15 Hz) and to maximize principal cell spiking in response to external oscillations at slightly higher frequencies. Importantly, we found that the fastest oscillation frequency that can be relayed by the network maximizes local inhibition and is equal to a frequency even higher than that which maximizes the firing rate of excitatory cells; we call this phenomenon population frequency resonance. This form of resonance is shown to determine the optimal driving frequency for suppressing responses to asynchronous activity. Lastly, we demonstrate that the natural and resonant frequencies can be tuned by changes in neuronal excitability, the duration of feedback inhibition, and dynamic properties of the input. Our results predict that PFC networks are tuned for generating and selectively responding to beta- and gamma-rhythmic signals due to the natural and resonant properties of inhibition-based oscillators. They also suggest strategies for optimizing transcranial stimulation and using oscillatory networks in neuromorphic engineering.
| The prefrontal cortex (PFC) flexibly encodes task-relevant representations and outputs biases to mediate higher cognitive functions. The relevant neural ensembles undergo task-related changes in oscillatory dynamics at beta- and gamma frequencies. Using a computational model of the PFC network, we show that strong feedback inhibition causes the PFC to generate internal oscillations and to prefer external oscillations at similar frequencies. The precise frequencies that are generated and preferred can be flexibly tuned by varying the synchrony and strength of input network activity, the level of background excitation, and neuromodulation of intrinsic ion currents. We also show that the peak output frequency in response to external oscillations, which depends on the synchrony and strength of the input as well as the strong inhibitory feedback, is faster than the internally generated frequency, and that this difference enables exclusive response to oscillatory inputs. These properties enable changes in oscillatory dynamics to govern the selective processing and gating of task-relevant signals in service of cognitive control.
| Oscillatory neural activity is a common feature of brain dynamics. In vitro experiments have demonstrated that different brain regions can produce network oscillations at different frequencies [1, 2]. In vivo experiments have shown that field potential oscillations in prefrontal cortex (PFC) at beta- (15-35Hz) and gamma-(35-80Hz) frequencies undergo task-related modulations in their power [3] and synchrony [4] and that multiple frequencies can appear in the same region [5, 6]. Despite the wealth of experimental evidence suggesting changes in oscillation frequency and synchrony are functionally significant, little remains known about the mechanisms by which they affect processing in downstream networks (but see [7]). In this paper, we will explore the natural, resonant, and competitive dynamics of PFC networks and how the task-modulated properties of oscillatory signals affect those dynamics.
Neural systems at multiple spatial scales are known to exhibit larger responses to oscillatory inputs at preferred (resonant) frequencies. For instance, neurons can exhibit resonance in subthreshold voltage fluctuations [8, 9], and networks can exhibit resonance in the amplitude of suprathreshold instantaneous firing rates of self-inhibiting interneurons (INs) [10] and reciprocally connected populations of principal cells (PCs) and INs [11]. Given weak inputs, these systems often exhibit response amplitudes that scale linearly with the input, and they oscillate with the same frequency as the input. In the linear regime, analytical methods can be applied to fully characterize network responses [11]. However, the results of such analyses no longer hold when inputs are strong and responses become strongly nonlinear.
Neural models of fast network oscillations, like those observed in PFC, often involve populations of cells receiving strong feedback inhibition to synchronize the network and strong excitatory input to drive the oscillation [6]. Under a constant (possibly noisy) tonic input, self-inhibiting populations of INs and reciprocally connected PC and IN populations can generate (natural) gamma-frequency network oscillations, termed ING [1, 12] and PING [12, 13], respectively. Due to the strong input and oscillatory response to a tonic drive, the earlier work on network resonance does not extend to these inhibition-based oscillators.
In this article, we present a numerical study of the natural and resonant behavior of an inhibition-based PC/IN network oscillator. In contrast to the linear regime, we will show that the frequency of the network oscillation equals the input up to a maximal frequency, above which, it decreases; we call this phenomenon population frequency resonance. We will show that different input frequencies maximize inhibition and excitation when inputs are strong and that the population frequency peaks when inhibition is maximized. The importance of this phenomenon will be demonstrated by showing that the optimal driving frequency for suppressing responses to asynchronous input is that which maximizes population frequency and not that which maximizes the firing rate of excitatory PCs. Finally, we will show how network resonance depends on dynamic, task-modulated properties of the input as well as intrinsic properties of the resonant network. Our quantitative results identify mechanisms that are not model specific, as will be shown by analogous simulations in a Hodgkin-Huxley type model of PFC and a generic integrate-and-fire network model that exhibit qualitatively similar behavior.
The paper will begin with a characterization of network responses to asynchronous and oscillatory inputs. Responses will be characterized in terms of firing rate and population frequency, and then the latter will be shown to determine the maximal suppression of asynchronous activity. The dependence of response on experimentally-motivated input parameters will be described. Finally, the paper will end with a discussion of the functional relevance of these findings for flexible neural processing.
We explored the impact of modulating task-related signals on cortical processing using an experimentally-constrained, Hodgkin-Huxley type network model of prefrontal cortex (PFC). Principal cell (PC) activity in PFC can be interpreted as a bias signal that mediates higher cognitive functions. The model represents a deep output layer of reciprocally-connected PCs and fast spiking interneurons (INs) that provide strong feedback inhibition. The network was driven by collections of independent spike trains modeling upstream activity in populations of excitatory cells. The input spike trains were either asynchronous with constant rate or mediated by an oscillatory modulation of the rate. Rhythmic activity is considered task-relevant [14] while the asynchronous activity is task-irrelevant (see Discussion for further considerations). We studied how the network behavior varies with task-related changes in synchrony, frequency, and strength of periodic inputs, and how that behavior relates to the response driven by equal-strength, asynchronous activity.
The difference in natural and resonant frequencies in the PC/IN network has at least one functional consequence that we will introduce here. It will serve as motivation for our further exploration of the dependence of natural and resonant frequencies on other properties in the remaining sections below. Consider two parallel pathways driving separate output PC populations that are reciprocally connected to a shared pool of INs (Fig 3A). One pathway (the target) delivers a rhythmic signal to one PC population while the opposing pathway (the distractor) delivers an equal-strength asynchronous signal to the competing PC population.
Without competition, both PC populations would output periodic pulse packets of excitatory spikes, the target at the input frequency if f i n p ≤ f R p o p and the distractor at the natural frequency fN. The PC population frequency determines how frequently a PC population engages the IN population. When multiple outputs compete through shared INs, the output population with the highest frequency oscillation most frequently drives IN cells, tends to phase lock with them, and suppresses spiking in output populations oscillating more slowly. Any time the target population oscillates with a frequency faster than the natural frequency (i.e., fpop > fN), spiking in the distractor population is suppressed (Fig 3B). Importantly, peak suppression of the distractor population occurs when the target population frequency is maximal and not when the target PC activity is strongest. This implies that the optimal driving frequency to suppress responses to asynchronous distractors is the fpop-resonant frequency (Fig 3B and 3C). Such a rhythmically-driven output oscillating faster than the natural oscillation will always drive INs to continuously suppress responses to asynchronous distractors as long as the faster oscillation in the target remains. Internally-generated, nested oscillations with frequencies greater than fN would also suppress the response to asynchronous drive but only while present on the depolarizing phase of the slower driving oscillation. This distractor suppression occurs because the target population recruits interneuron-mediated lateral inhibition on every cycle of its oscillatory input with a period shorter than that required for the distractor population to reach threshold (Fig 3C, see membrane potential plots). Even if the lower-frequency distractor would otherwise have a higher firing rate than the target, its spike output is never fully realized when it receives another pulse of strong inhibition before reaching threshold. For this reason, the outcome of the competition is determined by the frequency of the population oscillation and not its amplitude. Dynamically, the suppression arises within a cycle as the target begins to oscillate more quickly than the distractor (S5 Fig). In contrast, there is no suppression of either pathway when the distractor input is strong enough so that the natural frequency that it induces equals the population frequency of the target (S6 Fig), despite the distractor PCs having lower firing rate, or when both PC populations receive asynchronous input (see T1 in S5 Fig).
Furthermore, the extent to which maximum fpop (i.e., f R p o p) exceeds fN determines the range of input frequencies that can activate targets that suppress competing responses to asynchronous distractors (Fig 3B). Since f R p o p > f N, there is always an input frequency that can suppress competing distractors. In this scenario, the expected firing rate difference does not determine who is suppressive as long as firing rates are sufficient for a PC population pulse to activate the inhibitory INs. This result provides further justification for considering fpop as an output measure (in addition to r ¯ P C) because that frequency can determine the outcome of competition. This example represents a novel, functionally-relevant reason for examining output fpop and demonstrates the importance of the separation between the natural fN and resonant frequencies (f R p o p = f R I N) in PC/IN networks with strong feedback inhibition.
For the remainder of the work presented here we will examine how the response properties of the PC/IN network (with one output PC population) depend on flexible parameters of the input and the slower effects of neuromodulation. It will be shown that the response properties of the PC/IN oscillator are not purely intrinsic and can be adaptively shaped by extrinsic influences. This represents a powerful means by which task-related modulations can influence cortical processing.
In this work, we characterized the prefrontal PC/IN network response to strong oscillatory inputs in terms of biologically-relevant input and output properties. The PC/IN network with strong feedback inhibition exhibited resonance in the spiking of PC and IN populations as well as the output population frequency of the network. We have shown that a separation of preferred frequency for output spiking (the frequency maximizing PC activity) and the maximal frequency that can be relayed by the network is enabled by the combination of (1) strong excitatory input that generates a response to all input frequencies, (2) strong feedback inhibition: the ability of fast spiking INs to synchronously silence the PC population, and (3) noisy spiking in a population of PCs for which an active subset are able to activate INs. The peak output frequency of the inhibition-based network oscillator was determined by the input frequency maximizing local inhibition from the INs and always exceeded the natural frequency induced by equal-strength asynchronous activity. This population frequency resonance was shown to determine the optimal driving frequency for suppressing responses to asynchronous input. Finally, we showed that the resonant properties of PC/IN networks can be flexibly tuned by task-modulated signal properties (synchrony and strength) to dynamically shape ongoing neural processing.
Resonance phenomena have been studied in neural systems at multiple scales. Peaks in the single neuron membrane potential response to subthreshold oscillatory inputs have been studied in terms of the interplay between intrinsic ion currents [9]; their ability to influence spiking has been demonstrated in single neurons [42]; and relationships between subthreshold resonance and the natural network frequency of electrically coupled excitatory cells have been shown [43]. Our PC model, in isolation, exhibits subthreshold resonance at delta frequencies (2Hz) (S1A Fig) that translates into an input strength-independent spiking resonance at the same frequency for suprathreshold inputs (S1B and S1C Fig). The addition of strong feedback inhibition suppresses the spiking response to delta-frequency inputs while a higher-frequency, input strength-dependent spiking resonance emerges in response to strong inputs (S1D Fig). We have explored this higher-frequency resonance in this work and shown how it depends on the strength of input (S1Ei and S1Eii Fig) and the time constant of feedback inhibition (S1Eiii Fig).
The mechanism that determines the precise value of f R P C in the inhibition-based PC/IN oscillator is not fully understood. Insight into the locking of an inhibition-based IN network oscillator to periodic drive with heterogeneous phases has been obtained using a timing map [18]; a similar analysis might provide insight in the present case of a PC/IN network locking preferentially to a particular periodic drive with heterogeneous spike times but is beyond the scope of this work. The work in [18] and our results suggest the value of f R P C is related to the number of PC spikes that are necessary to engage INs on a given cycle. Changing the network size while keeping the synaptic strengths constant did not affect f R P C (compare Fig 2Bi to S4 Fig). This is in contrast to work showing that resonant frequency in a network model of Wilson-Cowan oscillators depended strongly on network size [44]. Further work is needed to understand the differences between resonances in spiking versus activity-based networks.
[11] showed that strong feedback inhibition is required for firing rate resonance in integrate-and-fire networks of excitatory and inhibitory cells driven by sinusoidal inputs; they also showed that stronger inputs increase resonant frequencies. We have reproduced these qualitative findings in a more detailed network model and extended the results by showing how firing rate resonance relates to the natural and peak oscillation frequencies in the strong-input regime, and how they all depend on other properties of the oscillatory input (i.e., waveform and spike synchrony). Compared to the integrate-and-fire networks (S3 Fig), the prefrontal network exhibited lower natural and resonant frequencies. [45] showed that heterogeneity of PC intrinsic properties in a PC/IN network produces a range of resonant frequencies that supports combining, instead of selecting, inputs. In contrast, our PC population is homogeneous, and the network produces more selective responses favoring outputs with fpop-resonant inputs. See Methods for a comparison of output measures used in this and other studies of spiking resonance. [46] investigated the response of an inhibition-paced IN population to physiologically-relevant periodic pulse inputs, but their work did not include PC cells or examine resonance.
The model investigated in this work made the following simplifications: no NMDA synapses, usage of all-to-all connectivity between PCs and INs, no PC-to-PC or IN-to-IN connectivity, and the lack of noise driving INs. Preliminary simulations demonstrated that probabilistic connectivity, weak noise, and weak NMDA synapses did not disrupt the results. For strong noise to INs, additional IN-to-IN feedback inhibition is necessary to synchronize the IN population. Furthermore, IN-to-IN and PC-to-PC connectivity have been shown to modulate resonant frequencies [36]. We suspect the main requirement for our results to hold is that the network is in a regime that produces a natural oscillation in response to an asynchronous input to the PCs (i.e., external noise to INs must be weak enough, feedback inhibition strong enough, and PC-to-IN drives strong and fast enough for a fraction of spiking PC cells to control IN activity from cycle to cycle).
Another important limitation of the present work pertains to the role of modulatory intrinsic currents. We have focused on regimes where PCs are roughly regular spiking and INs are fast spiking. More work is needed to understand how the dynamics reported here would be affected by PCs that are intrinsically bursting (as observed in deep layers of cortex and thalamus) and INs exhibiting low-threshold spiking. Furthermore, our account of the effects of knocking out modulatory currents is limited to effects on overall activity levels across the PC population. In contrast, work by [47] shows that modulatory currents can impact the cycle-to-cycle probability of individual cells participating in the population rhythm.
The work reported here has introduced a distinction between time-averaged firing rate resonance in excitatory PCs and inhibitory INs that arises when inputs are strong. We have also introduced a new form of network resonance observed in strongly-driven inhibition-based PC/IN oscillators, called population frequency resonance that depends on firing rate resonance in INs; and demonstrated its importance for suppressing responses to asynchronous activity. These results have also made a significant contribution to understanding how PC/IN networks with strong feedback inhibition are affected by task-modulated changes in oscillatory inputs, in general, and why beta rhythms are so frequently associated with prefrontal activity.
The network model represents a cortical output layer with 20 excitatory principal cells (PCs) connected reciprocally to 5 inhibitory interneurons (INs). Hodgkin-Huxley (HH) type PC and IN models were taken from a computational representation of a deep layer PFC network consisting of two-compartment PCs (soma and dendrite) with ion channels producing INaF, IKDR, INaP, IKs, ICa, and IKCa currents (μA/cm2) and fast spiking INs with channels producing INaF and IKDR currents [48] (Fig 7A; see figure caption for channel definitions). IN cells had spike-generating INaF and IKDR currents with more hyperpolarized kinetics and faster sodium inactivation than PC cells, resulting in a more excitable interneuron with fast spiking behavior [48]. In the control case, PC and IN cell models were identical to those in the original published work [48] while network connectivity was adjusted to produce natural oscillations (not in [48]), as described below, and the number of cells in the network was decreased to enable exploration of larger regions of parameter space while remaining large enough to capture the dynamics of interest for this study; however, the same resonant frequencies were obtained in simulations using the original network size (S4 Fig). Knockout experiments were simulated by removing intrinsic currents one at a time from the PC cell model. All cells were modeled using a conductance-based framework with passive and active electrical properties of the soma and dendrite constrained by experimental considerations [49]. Membrane potential V (mV) was governed by:
C m d V d t = - I i n p ( t , V ) - ∑ I i n t - ∑ I s y n (1)
where t is time (ms), Cm = 1 μF/cm2 is the membrane capacitance, Iint denotes the intrinsic membrane currents (μA/cm2) listed above, Iinp(t, V) is an excitatory current (μA/cm2) reflecting inputs from external sources described below, and Isyn denotes synaptic currents (μA/cm2) driven by PC and IN cells in the network. We chose to explore the prefrontal model as part of a larger project on prefrontal oscillations. We confirmed the generality of our qualitative results using a leaky integrate-and-fire (LIF) model; see the caption of S3 Fig. for details on the LIF model. We explored single cell and minimal network versions of our HH type model to investigate potential relationships between single cell and network resonances; details on these simulations can be found in the caption of S1 Fig.
The output layer had either one or two populations of PC cells with each output population receiving either one or two input signals. Input frequency-dependent response profiles were characterized using a network with one input and one output (Fig 7A). Competition between the outputs of parallel pathways was investigated using a network with two homogeneous output populations receiving one input each while interacting through a shared population of inhibitory cells (Fig 7B).
PC cells provided excitation to all IN cells, mediated by α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) currents. IN cells in turn provided strong feedback inhibition mediated by γ-aminobutyric acid (GABAA) currents to all PC cells. This combination of fast excitation and strong feedback inhibition is known to generate robust network oscillations in response to tonic drive [12, 13]. AMPA currents were modelled by:
I A M P A = g A M P A s ( V - E A M P A ) (2)
where V is the postsynaptic membrane voltage, gAMPA is the maximal synaptic conductance, s is a synaptic gating variable, and EAMPA = 0 mV is the synaptic reversal potential. Synaptic gating was modeled using a first-order kinetics scheme:
d s d t = H ( V p r e ) 1 - s τ r - s τ d (3)
where Vpre is the presynaptic membrane voltage, τr = 0.4 ms and τd = 2 ms are time constants for neurotransmitter release and decay, respectively, and H(V) = 1 + tanh(V/4) is a sigmoidal approximation to the Heaviside step function. GABAA currents are modeled in the same way with EGABA = −75 mV and τd = 5 ms. Maximum synaptic conductances for PC cells were (in mS/cm2): GABAA (.1); for IN cells: AMPA (1).
Each PC cell received independent Poisson spike trains (Fig 8) with time-varying instantaneous rate λ(t) (sp/s) and time-averaged rate rinp = 〈λ〉; spikes were integrated in a synaptic gate sinp with exponential AMPAergic decay contributing to an excitatory synaptic current Iinp = ginp sinp(V − EAMPA) with maximal conductance ginp (mS/cm2). Input signals were modeled by collections of spike trains with the same instantaneous rate-modulation. A given input signal to a PC output population can be interpreted as conveying rate-coded information from a source population in a particular dynamical state.
Signals from sources in different dynamical states were generated by modulating instantaneous rates λ(t) according to the activity patterns exhibited by populations in those states. Signals from source populations in an asynchronous state were modeled by Poisson spike trains with constant rate λ(t) = rinp (Fig 8A) whereas signals from sources in an oscillatory state were modeled using periodically-modulated instantaneous rates (Fig 8B). Signals with sine wave modulation had λ(t) = rinp(1 + sin(2πfinpt))/2 parameterized by rinp (sp/s) and rate modulation frequency finp (Hz). Sinusoidal modulation causes spike synchrony (the interval over which spikes are spread within each cycle) to covary with frequency as the same number of spikes become spread over a larger period as frequency decreases. Thus, we also investigated oscillatory inputs with square wave modulation in order to differentiate the effects of synchrony and frequency while maintaining the ability to compare our results with other work. Square wave rate-modulation results in periodic trains of spikes with fixed synchrony (pulse packets) parameterized by rinp (sp/s), inter-pulse frequency finp (Hz), and pulse width δinp (ms). δinp reflects the synchrony of spikes in the source population with smaller values implying greater synchrony; decreasing δinp corresponds to decreasing the duty cycle of the square wave. For the square wave input, we chose to hold constant rinp so that across frequencies the only significant change is in the patterning of spikes and not the total number of spikes; this results in larger pulses being delivered to postsynaptic PCs at lower frequencies as would be expected if lower frequencies are produced by larger networks [50]. If the number of spikes per cycle was fixed, instead, as would be the case for a given input population with iFR fluctuating more rapidly and all cells spiking on every cycle, then the mean strength of the input would increase with frequency, and its effect on resonance would no longer be comparable to a sinusoidal input with increasing frequency. The consequence of holding the number of spikes per cycle fixed for a square wave input is discussed further below and related to the results for fixed-mean square waves in S8 Fig.
All principal cells in the output layer received additional asynchronous inputs representing uncorrelated background activity from 100 cells in other brain areas spiking at 1 sp/s. Notably feedforward inhibition was excluded from the present work so that asynchronous inputs were maximally effective at driving PC cells. The effects of adding feedforward inhibition and conditions under which each case holds are considered in the Discussion. Control values for input parameters were rinp = 1000 sp/s (corresponding to a source population with 1000 projection neurons spiking at 1 sp/s); δinp = 1 ms (high synchrony), 10 ms (medium synchrony), or 19 ms (low synchrony), and ginp = .0015 mS/cm2. High synchrony inputs are similar to strong, periodic spikes while medium and low synchrony inputs distribute spikes uniformly over intervals comparable to sine waves at 100 Hz and 53 Hz, respectively.
In simulations probing resonant properties, the input modulation frequency finp was varied from 1 Hz to 50 Hz (in 1 Hz steps) across simulations. In simulations exploring output gating among parallel pathways, input signals had the same mean strength (i.e., rinp); this ensures that any difference between the ability of inputs to drive their targets resulted from differences in the dynamical states of the source populations and not differences in their activity levels.
For each simulation, instantaneous output firing rates, iFR, were computed with
Gaussian kernel regression on population spike times using a kernel with 6 ms width for visualization and 2 ms for calculating the power spectrum. Mean population firing rates, r ¯ P C and r ¯ I N, were computed by averaging iFR over time for PC and IN populations, respectively; they index overall activity levels by the average firing rate of the average cell in the population (Fig 8Ci and 8Cii). The frequency of an output population oscillation, fpop, is the dominant frequency of the iFR oscillation and was identified as the spectral frequency with peak power in Welch’s spectrum of the iFR (Fig 8Ciii, S7 Fig). As defined, fpop usually reflects the rhythmicity of internal spiking; however, when nested oscillations are present at low frequencies, fpop may reflect either internal or external rhythm frequencies (see Fig 4D for a raster plot and PC iFR, Fig 4B for a fpop response profile, and S7 Fig. for an iFR power spectrum with nested oscillations). This ambiguity does not interfere with our study of resonance at higher frequencies where the signal has a single dominant frequency; however, a disambiguating measure of population frequency would be necessary to study regimes in which multiple frequencies are strongly present (e.g., strong, low-frequency, low-synchrony periodic inputs). The natural frequency fN of the output network was identified as the population frequency fpop produced in response to an asynchronous input.
Our measure of spiking activity in the strongly-driven network differs from measures used in work on resonance in weakly-driven networks [10, 11, 42]. In the weakly-driven (i.e., linear) regime, the iFR amplitude scales linearly with the input and can serve as a measure for detecting resonance. However, in the strongly-driven regime that we explore, iFR may scale nonlinearly with the input; in the case of high-synchrony inputs, iFR amplitude saturates below the resonant frequency (i.e., all cells spike once on every cycle), and it has a more complicated profile and scaling with input strength in other cases. [51] has explored spiking resonance in a strongly-driven single cell and defined a measure called spike frequency that is roughly equivalent to the time-averaged firing rate. We have chosen to use a similar measure, the time-averaged iFR, r ¯ *, to capture overall increases or decreases in the amount of spiking produced in the strongly-driven network.
Qualitatively, r ¯ P C profiles differ for the PFC PC/IN network with strong feedback inhibition depending on the waveform of the periodic input (Fig 9). Weak sinusoidal inputs produce band-pass responses like those observed in [10, 11] (Fig 9Ai; S1D and S1Ei Fig, blue curve). Increasing the strength of those inputs produces an all-pass regime in which inputs at all frequencies elicit a response, although a resonant peak remains (Fig 9Aii; S1D and S1Ei Fig, black curve). In contrast, a weak square wave with mean input held constant across frequencies produces a low-pass response due to the larger input pulses at low frequencies (Fig 9Bi). However, the curve still exhibits a peak that occurs at the same input frequency as for the sine wave given equal-strength input. Finally, a weak square wave with pulse amplitude held constant produces a high-pass response due to the increasing input strength that occurs with an increasing number of pulses (Fig 9Ci). Increasing the strength of square wave inputs also moves the network into an all-pass regime (Fig 9Bii and 9Cii), but only the fixed-mean square wave exhibits a resonant peak (Fig 9Bii). In this work, we focus on the sine and square wave cases where mean input strength is held fixed and resonance is well-defined in physiologically-relevant frequency ranges.
Across simulations varying input frequencies, statistics were plotted as the mean ± standard deviation calculated across 10 realizations. Input frequency-dependent plots of mean firing rates and population frequencies will be called response profiles. The time-averaged firing rate resonant frequencies, f R P C and f R I N, are the input frequencies at which global maxima occurred in the r ¯ P C and r ¯ I N firing rate profiles, respectively. Similarly, the resonant input frequency, f R p o p, maximizing output oscillation frequency was found from peaks in fpop population frequency profiles, excluding the peaks that are due to nested oscillations in response to strong, low-frequency, low-synchrony periodic drives.
All models were implemented in Matlab using the DynaSim toolbox [52] (http://dynasimtoolbox.org) and are publicly available online at: http://github.com/jsherfey/PFC_models. Numerical integration was performed using a 4th-order Runge-Kutta method with a fixed time step of 0.01 ms. Simulations were run for 2500ms and repeated 10 times. The network was allowed to settle to steady-state before external signals were delivered at 400 ms. Plots of instantaneous responses begin at signal onset. The first 500 ms of response was excluded from analysis, although including the transient did not alter our results significantly.
|
10.1371/journal.pntd.0003393 | How Effective is Integrated Vector Management Against Malaria and Lymphatic Filariasis Where the Diseases Are Transmitted by the Same Vector? | The opportunity to integrate vector management across multiple vector-borne diseases is particularly plausible for malaria and lymphatic filariasis (LF) control where both diseases are transmitted by the same vector. To date most examples of integrated control targeting these diseases have been unanticipated consequences of malaria vector control, rather than planned strategies that aim to maximize the efficacy and take the complex ecological and biological interactions between the two diseases into account.
We developed a general model of malaria and LF transmission and derived expressions for the basic reproductive number (R0) for each disease. Transmission of both diseases was most sensitive to vector mortality and biting rate. Simulating different levels of coverage of long lasting-insecticidal nets (LLINs) and larval control confirms the effectiveness of these interventions for the control of both diseases. When LF was maintained near the critical density of mosquitoes, minor levels of vector control (8% coverage of LLINs or treatment of 20% of larval sites) were sufficient to eliminate the disease. Malaria had a far greater R0 and required a 90% population coverage of LLINs in order to eliminate it. When the mosquito density was doubled, 36% and 58% coverage of LLINs and larval control, respectively, were required for LF elimination; and malaria elimination was possible with a combined coverage of 78% of LLINs and larval control.
Despite the low level of vector control required to eliminate LF, simulations suggest that prevalence of LF will decrease at a slower rate than malaria, even at high levels of coverage. If representative of field situations, integrated management should take into account not only how malaria control can facilitate filariasis elimination, but strike a balance between the high levels of coverage of (multiple) interventions required for malaria with the long duration predicted to be required for filariasis elimination.
| Integrated vector management aims to optimize efficacy and make better use of available funds, including targeting multiple diseases, using one or more interventions. However, we have relatively poor understanding of the programmatic demands that arise when controlling two diseases. For instance, does the intensity, duration of deployment, or type of intervention most suitable for each disease overlap or clash? We developed a mathematical model to explore these issues for the example of the vector-borne diseases malaria and lymphatic filariasis. Since the causative agents of these diseases are transmitted by the same mosquito species in certain areas, there is clear potential for an integrated approach using long-lasting insecticidal nets (LLINs) or larval source management. We found that the transmission potential of both malaria and LF is most sensitive to changes in mosquito survivorship and the duration of the feeding cycle, supporting the usefulness of LLINs. In areas where both diseases occur, malaria elimination was predicted to require high levels of both LLINs and larval source management, whereas either intervention at a low intensity was sufficient to eliminate LF, if maintained for a longer period. This highlights that integrated control programs should be flexible and dynamic in order to accommodate these demands.
| Vector control continues to play a major role in ameliorating the burden of vector-borne diseases, such as malaria, for instance through the use of long lasting insecticidal nets (LLINs) [1]. As progress is made toward meeting the goals of local elimination and global eradication that have been set for a number of vector-borne diseases, there is a need in many parts of the world to further reduce the intensity of transmission and make better use of existing funds over the long term [2], [3]. Integrated vector management (IVM) is seen as a way to make rational decisions about the choice of vector control tools, improve (cost-)effectiveness and sustainability of control and limit the use of insecticides based on an understanding of local ecological conditions [2], [4]. Within IVM there are two broad approaches: one which uses a combination of interventions against a single disease and one which uses one or more interventions against more than one vector-borne disease. Since there are few field studies that report the impact of vector control interventions on more than one vector-borne disease, we use mathematical modelling to explore how common vector control tools could impact two diseases transmitted by one vector species.
Malaria and lymphatic filariasis (LF), both transmitted by the same vectors in rural Africa, serve as an examplar of the potential benefits and complications of IVM. The potential for integrating control across both diseases stems from their broad geographic overlap, shared vectors across much of this range, and susceptibility to the same interventions [2], [5]. This is particularly relevant for the poorest countries where the burden due to these diseases remains the highest. In Africa, particularly in areas where Loa loa co-occurs, mass drug administration programmes to clear Wuchereria bancrofti microfilariae from the human population face the challenge that the most effective anti-helminthics are contraindicated and vector control may have to be relied on more heavily [6]. In areas with the highest malaria burden due to Plasmodium falciparum infections, and especially in rural areas where malaria is spread by the efficient and anthropophilic vectors Anopheles gambiae s.l. and An. funestus, that are also primary vectors of W. bancrofti [7], current control measures may not be sufficient to interrupt transmission, and increasing worries about insecticide resistance highlight the need for an efficient, sustainable, and well thought out approach to controlling multiple diseases [8].
To date, synergy between malaria and LF control programmes has been mostly in the form of almost accidental side-effects of malaria control on filariasis transmission. A notable example being the Solomon Islands malaria eradication initiative commenced in 1960, where both parasites were transmitted by An. farauti and An. koliensis [9], [10], [11]. Whilst indoor residual spraying for 10 years failed to eradicate malaria, it resulted in the disappearance of LF from the island. Bed nets have likewise been shown to affect transmission of LF. Use of untreated bed nets in one village in Papua New Guinea was associated with a reduction in the proportion of vectors harbouring infective larvae from 5.38% to 1.62% [12]. Despite the large number of studies that have investigated the impact of LLINs on malaria there have been few that investigated their effect on LF. In Kenya the introduction of treated nets lowered the number of indoor-resting An. gambiae s.l. and An. funestus and reduced the human blood index of Culex quinquefasciatius [13]. Similarly the proportion of infected An. punctulatus decreased from 1.8% to 0.4% after the distribution of LLINs in Papua New Guinea, and, importantly, no infective larvae were found [14].
The evidence suggests that interventions directed against malaria vectors may be more effective at controlling LF if their use is sustained at least as long as the average lifespan of adult worms in humans, with estimates ranging from 4–10 years [15], [16]. Central to this is that LF is a far less efficient disease to transmit than malaria [17] requiring a far greater critical density of mosquitoes per human for LF than for malaria. Whilst one study estimated that between 5 and 100 infective bites (depending on the level of immunity of humans) will result in a malaria infection [18], many thousands are required before a patent filarial infection is produced. In Yangon (formerly Rangoon), where Cx. quinquefasciatus is the primary vector, it was estimated that an average of 15, 500 infective bites resulted in one microfilaraemic case [19]. The feasibility of interrupting LF transmission where Anopheles spp. are the vectors could be enhanced further by the process of facilitation, which is a density-dependent parasite-vector interaction resulting in an increasing yield of infective larvae as the number of microfilariae ingested increases. It has been suggested that this introduces an additional unstable equilibrium point above that associated with worm mating probabilities [20]. Interactions between parasites and mosquitoes in areas co-endemic for malaria and LF, could, potentially, result in perverse effects of control programmes aimed at only one disease [20], [21]. An illustration of this was recently provided, where a higher prevalence of malaria infection in humans was predicted to occur in the absence of LF [22]. The integration of control measures aimed at multiple diseases will thus have to take the different sensitivities of the diseases to interventions, as well as the complexities in transmission dynamics, into account. To explore how multiple interventions could be used as an IVM strategy to control malaria and LF where both parasites are transmitted by the same vector species, we develop a combined mathematical model that takes the interaction between the vector and multiple parasites into account. The model shares characteristics with a recently published model [22], but diverges in a number of areas. The focus initially is on a general anopheline (e.g. gambiae, funestus, or punctulatus complex) without developing species-specific behavioural characteristics such as the degree of anthropophily or response to interventions. We derive expressions of the basic reproductive numbers, R0, of both diseases in the presence and absence of the other disease and calculate the local sensitivity of R0 to the transmission parameters it encapsulates in order to gain insight in which parameters contribute most strongly to transmission and therefore make attractive targets for interventions. Additionally, we perform simulations of the impact of LLINs and larval source management to gain insight in the relative efficacy of each for both diseases in the presence and absence of the other. LLINs are used on a massive scale for the control of malaria, particularly in sub-Saharan Africa [23], and are highly effective at controlling vectors entering houses, whilst larval source management can be used as a supplementary intervention in some settings where it will impact both the indoor and outdoor biting population [24]. This study is to our knowledge the first to consider how one or more interventions can impact multiple vector-borne diseases and was carried out to help guide the development of an IVM programme to assist the elimination of both diseases.
We model both malaria and LF as a system of ordinary differential equations, representing mean filarial worm and microfilariae burdens in humans based on parasite burden helminth models [25], and proportions of the human population that is susceptible, infected but not yet infective, infective, or immune to disease for malaria, based on extensions of the Ross Macdonald model [26], with no interaction of parasites in humans. We assume susceptible-exposed-infectious prevalence dynamics for both diseases in mosquitoes with the possibility of co-infection. For LF infection in mosquitoes, since this entails modelling the prevalence of infection rather than the mean larval burden of each mosquito, we have made the assumption of strong density-dependence in the parasite action on the vector (as in some models for schistosomes [27]). All state variables of the ordinary differential equations are shown in Table 1. A diagrammatic overview of the model is given in Fig. 1. Adult filarial worm and microfilariae burdens in humans are modelled in a similar way as other deterministic filariasis models [28], [29], although age-dependence in humans and potential effects of immunity on worm establishment (simplifications shared with [22] and [16], respectively) are ignored.
We use a susceptible, exposed, infective, resistant and susceptible (SEIRS) compartmental model for malaria, following Chitnis et al [30], ignoring human migration and assuming a constant per capita density independent death rate.
We use the recovered class, Rh, to model immunity to malaria in humans, where people harbour low levels of parasite in their blood that are often undetectable but still allow a lower probability of transmission to mosquitoes. Although in reality, humans frequently move between patent and sub-patent parasitaemia, and the duration of infection may be dependent on past exposure, we make the simplifying assumption here that the recovered stage has a fixed duration. Equations and further details are provided in the S1 Supporting Information.
The number of microfilariae ingested by mosquitoes depends on the density of microfilariae in the blood meal [31]. For our prevalence based model, we fitted an exponential curve to data on the probability of ingesting microfilariae for An. gambiae, An. arabiensis and An. melas [32], [33], [34] (Fig. 2). The probability of a mosquito ingesting microfilariae when biting a human with microfilariae is given by the following equation: (1)
Where G reflects the mean microfilariae burden among infected humans,(2)and P(F) is the prevalence of infection depending on the mean microfilariae burden. We approximate a negative binomial relation between microfilariae burden and prevalence of infection as in [28] with(3)
The force of infection on mosquitoes is:(4)where a is the number of bites per mosquito per day and m is the proportion of ingested microfilariae that passes the midgut barrier (see the supplementary material). The force of infection for malaria of human to mosquito, λm, is:(5)where b is the likelihood of a mosquito becoming infected when feeding on an infective human, and the likelihood when feeding on an immune human.
Interactions between parasites and vectors are multi-faceted and likely depend on the co-evolutionary history of the association. Filarial larvae can exert a number of costs on the mosquito, including damage inflicted while crossing the midgut, while developing in the thoracic musculature, or due to breakage of the labium [35], or a metabolic cost associated with an immune response to infection [36], with high mortality typically associated with high numbers of larvae. Relatively few studies have looked into this matter for the natural pairing of Anopheles spp. and W. bancrofti infection and a meta-analysis did not support the notion of density-dependent mortality for this pairing [37]. Although recently in An. farauti density-dependent mortality due to filarial infection was observed [38] and field data has also suggested this for An. gambiae s.l. and An. funestus [39]. The effects of infection with Plasmodium on Anopheles survival is likewise equivocal, with evidence suggesting it occurs mostly based on experimental systems consisting of mosquito-Plasmodium species combinations that do not occur in nature, such as An. stephensi infected with P. berghei [40], [41]. This induced mortality was most likely be due to the sporozoite stage [40], [42], but has also been linked to the oocyst burden [43]. We make the assumption that additional mortality due to harbouring sporozoites and filarial larvae occurs so additional malaria mortality acts only on infectious mosquitoes and additional LF mortality acts only on exposed mosquitoes. Since we use a prevalence- rather than intensity-based model for infection with LF, we assume a constant level of mortality. Putatively, susceptibility to co-infection could be likelier as infection with one parasite could weaken the mosquito's innate defences against subsequent infection, and survival of co-infected mosquitoes could be decreased further, but we do not model this effect here. Evidence for the first comes from field studies finding higher than expected proportions of mosquitoes carrying co-infections [44], [45], but the opposite has also been reported [46], and laboratory experiments suggest filarial infection may reduce development of Plasmodium in vectors [47] so we ignore this effect in our model. In terms of mortality, we assume the parasite-induced mortality acts additively. Further details on parameter values used for mosquito-filaria interactions are provided in S1 Supporting Information.
To investigate the robustness of our baseline analyses we performed simulations allowing for model parameter uncertainty, environmental variation in the form of seasonality, and density-dependent larval mosquito survivorship.
To explore the impact of vector control methods on P. falciparum and W. bancrofti prevalence we ran a series of 500 simulations. The parameter sets used were drawn using Latin hypercube sampling from uniform distributions of parameter values with ranges as specified in Table 2.
Here, mosquito immatures were assumed to be subject to density-dependent mortality, based on the formula of Dye [48], so that the daily emergence of female mosquitoes is given by:(6)where E represents the number of female eggs oviposited on average by a mosquito per day, R the number of offspring that would emerge were only density-independent mortality to operate, and f(t) and β modify the strength of density-dependent immature mortality.
The effect of seasonality on mosquito population dynamics was assumed to operate through the effect of rainfall patterns on the number and size of available larval development sites. We approximated this effect by making the strength of density-dependent mortality, f(t) vary following a sinusoidal pattern:(7)where f0 is the baseline value of f(t) and ε regulates the strength of seasonal variation.
The basic reproductive number, or ratio, R0, estimates the number of secondary infections that result from the infectious duration of a single case in a fully susceptible population and thus it provides a basis for judging whether a disease will thrive or be eliminated. When R0 is less than or equal to 1 the disease will be eliminated, whilst if it is greater than 1, the disease will survive. We derive expressions for the basic reproductive numbers of malaria and LF (where R0 represents the number of adult filarial worms arising from one adult filarial worm in the absence of density dependent regulation) in the absence and in the presence of the other pathogen, using a next-generation matrix approach [49], [50]. The derivations are provided in the supplementary material. The resulting expression for malaria in the absence of LF is:(8)
Where the first term represents the probability of a person in the latent stage progressing to the infective stage rather than leaving the compartment through dying, multiplied by the average amount of time spent in the infective stage, multiplied by the number of mosquito bites on that person per day that result in infection of a vector. The second term stands for the probability that a latent human passes to the immune stage (through the infective stage), the average amount of time spent in the immune or recovered stage, multiplied by the number of bites received per day that result in vector infection. The third term is then the probability of an infected vector progressing to the infective stage times the bites per day that result in infection in humans, multiplied by the average amount of time spent in the infective stage.
While for LF in the absence of malaria we obtain:(9)
Which, following a similar logic, consists of terms representing the fecundity of an adult worm and the mean lifespan of adult worms and microfilariae, multiplied by the number of mosquito bites on the population that result in establisment of infection in the vector, and terms for the probability that an infected mosquito progresses to the infective stage, the average duration of the infective stage, and the number of infective larvae that are delivered per day per human that reach maturity. Here and are the equilibrium values of the mosquito population and human population, respectively, in the absence of both diseases.
For malaria in the presence of filariasis and for filariasis in the presence of malaria the basic reproductive numbers are given by large expressions (see the supplement), but we note that these reduce to the expressions similar to the equations above when parasite-induced mortality due to the other parasite is ignored.
Potential positive or negative side effects associated with disease control methods that target a single parasite only (for instance, due to treatment of humans) can be investigated by simulating the artificial removal of either parasite from the co-endemic equilibrium state. The effects on the proportion of infectious mosquitoes of either disease (i.e. harbouring sporozoites or third-stage larvae) after setting all human and mosquito infections to zero for the other disease are shown (Fig. 3). When malaria is removed, LF transmission intensity is expected to initially increase, before decreasing, eventually to a slightly lower equilibrium. The initial rise in the proportion of infectious mosquitoes can be ascribed to the removal of Plasmodium-induced mosquito mortality. Over a much longer period this is balanced by the adjustment in human population size following the removal of malaria-induced human deaths. The implications or removing LF on malaria transmission, based on our model, are more straightforward as transmission will increase in intensity due to the absence of filarial-induced vector mortality.
To determine the key variables that drive transmission of LF and malaria in the presence of the other disease a local sensitivity analysis of R0 was performed by calculating the normalized forward sensitivity indices [51], equivalent to the concept of elasticity [52], [53], as:(10)where pi is any parameter of the model. Such an analysis can help identify which parameters influence R0 most strongly and therefore make appealing targets for disease management [52]. If high impact variables are shared between the two diseases, these should be the focus of integrated strategies. We follow the methodology as described in [54] and evaluate the sensitivity index of R0 for malaria at the malaria-free endemic equilibrium point for LF, and vice versa. The values for each of the parameters that constitute the expressions of R0 are given in Fig. 4. For both diseases, the most important parameter is the biting rate, a, followed by the base mosquito mortality, μm. Notable is that the parasite-induced mortality (μi for malaria, μe for filariasis) has only a weak impact, while parameters related to transmission of the other parasite, including parasite-induced mortality, have a relatively minimal impact. This suggests that although removal of one parasite could (temporarily) increase transmission of the other parasite, such an effect should be overshadowed if interventions simultaneously target a shared parameter, such as the biting rate, vector mortality or density.
We investigate the impact of vector control by calculating the basic reproductive number of malaria and LF over a range of coverage levels (indicated by φ), and by numerical simulations. Both LLINs and larval control are considered due to their targeting of different vector-related parameters: the mosquito biting rate and feeding-related mortality, and the emergence rate and resulting mosquito density. Larval control is here modelled in a simplified manner and assumed to have a linear effect on the daily emergence of mosquitoes, Λ, so that a 90% coverage of breeding sites is assumed to reduce emergence of adults by 90% (but note that a recent paper suggests that treating 50% of breeding sites was associated with a 90% reduction in adult mosquitoes [55]). The equilibrium number of susceptible mosquitoes then becomes Λ(1-φ)/μm. We note here that in certain settings such as urban areas, high coverage of larval breeding sites may be possible, but in many areas it may be difficult to achieve such high coverage.
LLINs are investigated using the formulae derived by Le Menach et al [56] for the feeding cycle duration and probability of survival under varying LLIN coverage levels, and substituting these equations in our expressions for R0 (further details provided in the supplement). We make the simplifying assumption that mosquitoes are fully anthropophilic, but this can be relaxed if species-specific differences in feeding behaviour are of interest.
The impact of LLINs and larval control over a range of 0–100% coverage on the basic reproductive number of malaria in the absence and presence of LF, and of LF in the presence and absence of malaria, is presented (Fig. 5). In both cases, the effect of an interaction with the other parasite is very small. For malaria this is particularly the case, because the range of coverage levels where filariasis persists is very narrow and above that level the expressions of R0 are equivalent. This is the case because our model, evaluated at the parameter values specified, predicts a very low R0 of filariasis even in the absence of interventions, and a very low level of vector control (6% LLIN coverage or 20% coverage with larval control) is sufficient to reduce this below one. The R0 for malaria is greater and requires coverage of 90% of bed nets, 100% of larval control by itself, or 70% coverage of both ITNs and larval control. At a higher mosquito density (160,000), 36% coverage of LLINs, 58% of larval control, or 26% of the interventions applied together, are sufficient to reduce the basic reproductive rate of LF to below one, while 78% of LLINs combined with larval control is now required for malaria, while 100% coverage of either on its own is required to eliminate malaria transmission.
Fig. 6 shows the effects of vector control measures when both diseases are present when simulated over time, in this case at a level of coverage, φ, of 80%, for mosquito population sizes of 80,000 and 160,000. The results are in line with those of Fig. 5, and show with that larval control or bed nets by themselves, malaria prevalence would reach a lower equilibrium, whereas prevalence is reduced to zero over time when LLINs are combined with larval control. For LF all interventions are sufficient to eliminate the disease, although this reduction takes place over a longer period. For instance, the time that interventions have to be in place to reduce the prevalence of patent infection to one half that of the initial, equilibrium level of prevalence was 218 days for malaria compared to nearly 8 years (2,884 days) for LF, when LLINs are combined with larval control, for the lower mosquito population size. To achieve a 32-fold reduction takes 10 years (3,552 days) for malaria, and 32 years (11,554 days) for LF, under those same conditions. The impact of a wider range of R0, simulated by varying the mosquito density between 80000 and 330000 on the prevalence of infection over time when LLINs and larval control are both employed at 80% coverage, is shown (Fig. 7). As the mosquito density increases far enough, malaria again reaches a new equilibrium, whereas for the parameter values used, LF will be eliminated at this level of vector control even at high mosquito densities. The robustness of this outcome to varying a key parameter regarding the efficiency of establishment of new adult filarial worms is provided (Fig. 8), while the impact of uncertainty in parameter values overall as well as environmental variability in the form of seasonally varying density-dependent immature mosquito mortality is explored in Fig. 9. For LF, the 95th percentile range decreased to zero over time for both LLINs used alone as well as in combination with larval control, with only the rate of decline being affected. There was more variability in the response of malaria, although the overall pattern corresponded to that of our baseline investigation (Fig. 6). When both LLINs and larval control were employed, 84% of simulations had reached a prevalence <1% at the end of the projected period. When only LLINs were used, only 5% of simulations had reached that level.
The principal findings of this analysis were that the basic reproductive number, R0, of both LF and malaria can be brought to below one by controlling their anopheline vectors. The value of R0 of LF was found to be lower than the R0 of malaria, and consequently controlling LF required a lower level of population coverage with LLINs and larval source management than did controlling malaria. However, the time over which the prevalence of infection is reduced is longer for LF than for malaria.
The difference between the basic reproductive numbers and consequently the critical densities of mosquitoes of malaria and LF are striking, though not unexpected. Care does have to be taken with the comparison, because while both indicate threshold criteria, there is a difference in interpretation where for malaria, R0 represents secondary cases of humans and for LF the number of adult worms arising from one worm [50]. A question that arises is whether we see in the field that the two diseases co-occur mostly in areas with high malaria transmission, as predicted here. It does raise the question whether the low value of R0 of filariasis in combination with the mosquito density required to maintain transmission that we found in our analysis is representative of field situations.
Threshold biting rates between 200–300 per month were reported in a theoretical study for Anopheles spp., and as low as 9 bites per month for LF transmitted by Culex spp [29]. Stolk et al. [16] reported a threshold in their model at approximately 400 bites per person per month. Our prediction was approximately 675 bites per month, and the critical density for filariasis predicted by our model of 67 mosquitoes per human is likewise higher than the value of 20 calculated by Webber [17], while in a later field study a density of 60 was found [57]. Our values appear slightly higher, perhaps as a result of the additional vector mortality due to infection included here.
The main difference in the values of R0 between malaria and LF clearly depends on the parameters that are not shared between equations (6) and (7). As the filarial worms are long-lived and fecund, and the extrinsic incubation periods comparable between the parasites, this most likely then hinges on either parameters associated with uptake and establishment of filarial larvae in the mosquito (kψ, i, m), or, more likely with the (in-)efficiency of establishment of (mated, female) adult worms from infective bites (ψ1). A central criticism of our work is that this inefficiency of LF transmission stems from only one field study in Myanmar (formerly Burma) and an estimate from one prior modelling study [28] and clearly this needs empirical confirming in other parts of the world, including sub-Saharan Africa. This parameter, ψ1, represents an amalgamation of a number of factors, such as the probability that an infective filarial larva egresses from the labium during a blood meal, the probability that the larva successfully enters the puncture wound created by the blood-feeding mosquito, and the probability that the larvae survives, develops and mates within the human. The second, the probability that an infective larvae exposed on the human skin in a drop of haemolymph survives and enters the blood stream has been shown to depend on the relative humidity and rate of evaporation [58], and may thus show considerable geographic or seasonal variation.
The values for this parameter used in other recent LF models are also variable, ranging from 2.4–3.4 times higher than our value [16], [22] to ca. 18 [59] and 700 [29] times greater. Note that a direct comparison of these values is complicated by the inclusion of mating functions and particularly immunity affecting worm establishment in some of these more complex models. The range for this value used in our global uncertainty analysis was therefore based on values used in additional models without immunity (Fig. 9), though a wider range was considered in a one-way sensitivity analysis (Fig. 8). Our conclusions are robust to parameter uncertainty overall, as well as the inclusion of seasonal mosquito population dynamics (Fig. 9). The implications of seasonality appear more pronounced for malaria than they do for lymphatic filariasis, likely as a result of the shorter duration of Plasmodium infection in humans and indirect effects on acquired immunity. Our predicted outcomes are robust to changes in mosquito abundance (Fig. 7) as well as the inefficiency of Wuchereria transmission if we increase this parameter value (ψ1) 3.4-fold, but if increased 18-fold (as in certain models where immunity is assumed to negatively affect worm establishment) then at higher mosquito densities, 80% coverage of both LLINs and larval source management are predicted to lead to a lower equilibrium prevalence of LF rather than elimination, as is also the case for malaria (Fig. 8). This highlights the need for further model fitting and, given the general paucity of information on the impact of LF on vectors in the wild, the need to collect more field data to help improve the models. However, we note that the intention of the current analysis was not to provide a predictive model, but rather a general model including parasite-vector interactions of malaria and LF, to gain insight into how to integrate vector control for these two diseases. The use of simplifying assumptions is not restricted to the filariasis side of the model. The current malaria model, based on extensions to the Ross-Macdonald model, includes a class of recovered individuals with a lower infectiousness, which is meant to portray both the confounding effects of superinfections resulting in a skewed distribution of infection duration and the effects of acquired immunity in reducing parasite densities. Other models of malaria have included more realistic and complex systems of ordinary differential equations [60] or have used stochastic individual based models [61], [62], [63], and as typical in modelling, depending on the balance between analysis and simulations desired, a higher or lower degree of realism can be adapted. In the case of parasite-vector interactions greater realism could come from modelling these processes (mortality and immunity) as functions of parasite density. Since we were most interested in the potential negative side-effects of controlling only one disease and how this could be mitigated by vector control, we included only mortality induced by parasites, but a more realistic model may have to additionally consider other behavioural modifications, such as an increased biting rate.
A further caveat to this model is that homogeneous mixing of vectors and hosts is assumed, but in reality fine spatial variation, for instance due to microclimatic differences or proximity to mosquito breeding sites, will exist and potentially allow for transmission hotspots where R0 will be elevated and transmission harder to interrupt [64], [65]. To gain insight into interactions between the two diseases at such a scale, modifications will be required to the model structure to take heterogeneous exposure to mosquito bites into account. In addition to the differences in R0 between LF and malaria, the difference in response times to vector control interventions was notable, with LF requiring longer durations of interventions to reach low levels of prevalence due to the longevity of the adult filarial worms. For malaria there is an initial fast drop-off as the proportion of humans in the infective compartment is rapidly reduced, followed by a slower decline as individuals remaining in the less infectious but longer-lasting recovered compartment continue to contribute to transmission. Note however, that we have focused solely on vector control and not included the use of antihelminthic or antimalarial drugs in this study. Worm burdens will fall more rapidly when diethylcarbamazine, ivermectin, or albendazole can be administered to the population and should therefore remain a priority, even in areas that have been subject to malaria-vector control for many years.
An additional consideration is that the modelling approach we used likely overestimates the required duration that interventions have to remain in place for LF, because in our model mortality of filarial worms follows an exponential function and transmission can resurge with a small fraction of infected cases. A model that includes stochasticity, a more suitable distribution for worm lifespan than an exponential one (e.g., a Gompertz function), as well as a mating function for filarial worms, would be more suitable for accurately determining the duration of programs needed for elimination. In our simulations, the R0 of LF required only very modest levels of vector control coverage to be brought below one. A consequence of this is that larval control, which by itself was not sufficient to control malaria in our simulations, was sufficiently effective to halt filariasis transmission. However, even these low levels of coverage had to be maintained over a long period. Larval control, for instance, had to be maintained for nearly 10 years to halve the initial prevalence of infection, or approximately 41 years to achieve a 32-fold reduction in prevalence. Faced with the rise in pyrethroid-resistance seen with An. gambiae s.l. and the relatively poor ability of nets to kill culicine mosquitoes [66], applying larval control should be considered as a supplementary measure for filariasis control in addition to LLINs. Such an approach could be effective across mosquito genera and ecological settings. Larval source management could also supplement LLINs in areas with very high malaria transmission where use of LLINs at high levels of coverage are insufficient to severely dampen transmission. This would be particularly attractive if larval control activities can be maintained over longer periods (30–40 years in these simulations) by communities than the average lifespan of LLINs.
Our combined malaria and LF model takes into account that both Plasmodium and Wuchereria parasites may have adverse effects on the survival of their vectors, agrees with previous investigations that interventions that target only one of the parasites may have negative, unintended consequences on transmission of the other parasite [20], [21], [22]. However, the outcome of our sensitivity analysis suggests that transmission of both malaria and LF are most strongly impacted by perturbations of the mosquito biting rate, and vector mortality. Perturbation of parameters related to the other disease has a relatively minor impact, suggesting that any putative negative consequences of disease control will be overshadowed by the implications of vector control measures. That LLINs, which affect the biting rate and vector mortality, have a stronger impact on the basic reproductive number than larval control, which (in our simplified model without a full model of mosquito population dynamics) only affects mosquito density in a linear fashion, is thus in agreement with the sensitivity indices. However, it should be appreciated that in our model the impact of larval control is likely to be underestimated [55]. This lends further support to the notion of integrating the rollout of LLINs into LF drug administration campaigns [2], [14]. Based on the sensitivity analysis, indoor residual spraying (IRS) would be expected to have an impact on LF transmission (see also [11]) comparable to LLINs, but whether it should be considered as an alternative or in addition to LLINs may depend more on issues related to the cost-effectiveness of performing spray rounds for a sufficient amount of time needed to eliminate LF (i.e. at least 4–8 years).
This analysis confirms that the massive roll-out of LLINs for malaria control will have additional impact on the transmission and control of LF. Elimination of LF via vector control only is plausible, but likely only feasible in the form of mosquito abatement sustained over many years. The synergies that come from attacking two diseases with the same interventions should be exploited to a greater extent in elimination programmes. This is particularly relevant in West Africa where drug treatment against LF cannot be administered in areas endemic for loaiasis. LLINs and, where applicable, larval source management should be used for the control of malaria and LF in areas where both diseases are transmitted by the same vector.
|
10.1371/journal.pcbi.1004626 | Computational Modelling of Metastasis Development in Renal Cell Carcinoma | The biology of the metastatic colonization process remains a poorly understood phenomenon. To improve our knowledge of its dynamics, we conducted a modelling study based on multi-modal data from an orthotopic murine experimental system of metastatic renal cell carcinoma. The standard theory of metastatic colonization usually assumes that secondary tumours, once established at a distant site, grow independently from each other and from the primary tumour. Using a mathematical model that translates this assumption into equations, we challenged this theory against our data that included: 1) dynamics of primary tumour cells in the kidney and metastatic cells in the lungs, retrieved by green fluorescent protein tracking, and 2) magnetic resonance images (MRI) informing on the number and size of macroscopic lesions. Critically, when calibrated on the growth of the primary tumour and total metastatic burden, the predicted theoretical size distributions were not in agreement with the MRI observations. Moreover, tumour expansion only based on proliferation was not able to explain the volume increase of the metastatic lesions. These findings strongly suggested rejection of the standard theory, demonstrating that the time development of the size distribution of metastases could not be explained by independent growth of metastatic foci. This led us to investigate the effect of spatial interactions between merging metastatic tumours on the dynamics of the global metastatic burden. We derived a mathematical model of spatial tumour growth, confronted it with experimental data of single metastatic tumour growth, and used it to provide insights on the dynamics of multiple tumours growing in close vicinity. Together, our results have implications for theories of the metastatic process and suggest that global dynamics of metastasis development is dependent on spatial interactions between metastatic lesions.
| We used mathematical modelling to formalize the standard theory of metastatic initiation, under which secondary tumours, after establishment in a distant organ, grow independently from each other and from the primary tumour. When calibrated on the experimental data of primary tumour and total metastatic burden in the lungs in an animal model of renal cell carcinoma, the initial model predicted a size distribution of metastatic foci that did not fit with observations obtained experimentally using magnetic resonance imaging (which provided size and number of macro-metastases). The model predicted an increase in the number of lesions, but of smaller size when compared to the data. This led us to revise the standard theory and to propose two hypotheses in order to explain the observations: 1) small metastatic foci merge into larger ones and/or 2) circulating tumour cells may join already established tumours. We then derived a spatial model of tumour growth in order to explore the quantitative implications of tumours merging on global tumour growth and estimated the numbers of required metastatic foci to obtain the observed metastatic volumes.
| Metastasis, the spread of cancer cells from a primary tumour to secondary location(s) in the body, is the ultimate cause of death for the majority of cancer patients [1,2]. Although studied for more than 180 years [3], increasing efforts in recent years contributed to a better understanding of this aspect of tumour development [2,4], with exciting new discoveries [5–8] that potentially have important clinical implications. The metastatic process can be coarsely divided into two major phases: 1) dissemination of detaching cells from the primary tumour to a secondary site and 2) colonization of this distant organ [1,9]. While the former has been relatively well elucidated, in particular due to recent advances about the epithelial-to-mesenchymal transition [10] and advances on our understanding of molecular and genetic determinants [11,12], the latter remains not fully understood, especially during the colonization phase [1,12]. This is due, in part, to experimental limitations that hinder our ability to observe colonization of organs by tumour cells and the development of tumour lesions.
In this context, mathematical models provide powerful tools to potentiate data analysis, infer hidden information, test biological hypotheses against the empirical data and simulate a range of conditions that may be confronted to the biological reality. In recent years, several models for tumour growth have been developed (see [13,14] for historical reviews), based on multiple modelling techniques from non-spatial ordinary differential equations models (see [15] for a benchmark of these against experimental in vivo data) to discrete agent-based models [16–18] and continuous partial differential equations based on tissue mechanics laws [19,20]. However, despite a large body of literature for modelling tumour growth, relatively little effort has been devoted to the development and validation of mathematical models describing the biology of the metastatic process (see [21,22] for an early and notable exception, [23,24] for more recent studies and [25] for a recent review). In 2000, Iwata and colleagues proposed a simple mathematical model for the growth of a population of metastatic colonies [26], which was recently shown able to fit experimental data describing the increase in total metastatic burden [27,28]. In this mathematical description, each metastasis grows independently from the others and from the primary tumour. We report herein a theoretical study to test this hypothesis using in vivo data derived from a metastatic renal carcinoma model in mice. We show that the standard theory of metastatic initiation in which distinct foci grow independently from each other (as assumed in [21]) predicted an unrealistically large number of metastases, while the tumours sizes were too small.
In a space-limited organ (such as the lungs), where two neighbouring metastatic foci are growing in close vicinity, they might enter in contact and interactions occur, ultimately leading to the merging of the metastatic foci. This phenomenon is not taken into account in a classical description of metastasis development, although it can lead to important differences in the number and sizes of the colonies. Moreover, mechanical interactions could occur during metastases merging, possibly impacting the global dynamics. Therefore, we next conducted a simulation study to quantify the effect of mechanical interactions between two neighbouring tumours. Based on mechanical laws for tissue growth, we derived a minimally parameterized model (2 parameters). This second, spatial model, based on a pressure-mediated growth law, once fitted to magnetic resonance imaging data of individual metastatic tumour growths, offered an adapted framework to perform simulations of spatially interacting tumours. These revealed significant impact of the interactions on the global growth and allowed to test if merging by passive motion could explain the data that are not in accordance with the classical model.
To our knowledge, this is the first time that data on size distribution of metastasis at this resolution (with such a small visibility threshold, of the order of 0.05 mm3) is reported and analysed in lights of a theoretical model.
As an initial step, we studied the growth rates of individual metastatic tumours. Then, we calibrated a more elaborated mathematical model of tumour growth and metastatic dissemination using quantitative data derived from green fluorescent protein (GFP)-tracking of primary and metastatic tumours (see Materials and methods, n = 31 mice). Finally, we used the model to investigate predictions of the standard theory with regard to number and sizes of metastatic lesions and compared them to Magnetic Resonance Imaging (MRI) data (see Materials and methods, n = 6 mice).
Using a combined approach between experimental data and mathematical models, we demonstrated that the standard theory of metastasis formation and growth, where metastases grow independently from the rest of the system, was biologically unlikely. To explain our findings, we proposed several hypotheses, including the possibility of metastatic foci merging by passive motion. To investigate whether this hypothesis would have quantitatively non-negligible impact on the kinetics of the total metastatic burden (thus requiring more intricate modelling for the model describing the size distribution at the scale of the organ), we introduced a parsimonious spatial model of tumour growth. After calibrating the model to the growth of single metastases, we found (in simulations) that spatial interactions resulted in a significant reduction of tumour growth. Our results indicate that spatial interactions should be considered in future efforts for the development of a general quantitative theory of metastatic colonization.
Based on the rationale that lung capillaries have a diameter of the order of one tumour cell (20 μm) and that metastatic cells have lost expressions of cell-cell adhesion proteins such as cadherins [2], we assumed in our simulations, that metastases originated from one cell. This might be arguable and metastasis could start from tumour cell clumps [31,32]. To resolve this further and assess the robustness of our results, we performed the entire data analysis (fit of the total metastatic burden and resulting prediction of the metastatic size distribution) for values of the initial number of cells of 1, 10, 100 and 500 (S3 Fig). Initial numbers of 10, 100 and 500 cells could be in agreement with the data at day 19. However, with V0 = 10 cells, the predicted number of macro-metastases at day 26 was 3-fold higher than in the data. For V0 = 100 cells and V0 = 500 cells, the predicted macro-burden was 2-fold smaller than the observed one. Moreover the largest metastasis at day 26 was still predicted much smaller in the model than in the data (3.11 mm3 for V0 = 10 cells, 3.58 mm3 for V0 = 100 cells, 3.8 mm3 for V0 = 500 cells, against 13.6 mm3 in the data). Furthermore, in animal experiments the vast majority of detaching tumour cell clumps has been shown to comprise less than 10 cells [31] with a range of 2–50 cancer cells [32], which makes the theories V0 = 100 cells and V0 = 500 cells unlikely. This suggests that, if the metastases started from a substantial amount of cells, the grouping of these cells probably occurred at the distant site, after extravasation from the blood circulation. Similarly, we did not consider any cell loss at the moment of initial sub-capsular injection. We could make theoretical assumptions of cell loss (of 10%, 20%, etc…), which would simply consist in multiplying V0 by the relevant factor. For instance, considering a 90% loss (i.e. that only 10% of the cells remain viable) would be equivalent to multiplying V0 by 10. As demonstrated in S3 Fig, it is necessary to assume an initial size of at least 100 V0 to recover plausible values for the number of metastases at time T = 26 days. Combining the two (cell loss of 10% and initial metastatic size of 10 cells) thus gives a hypothesis that we are not able to infirm given the data we dispose.
The spatial model for tumour growth that we introduced is based on a pressure-induced decrease of the growth rate. Contact inhibition between cells is a mechanism for maintaining tissue homeostasis [4]. The ability of cancer cells to ignore these inhibition signals is a hallmark of cancer. In a recent study, Stylianopoulos et al showed that the uncontrolled proliferation of tumour cells results in mechanical stresses in the surrounding micro-environment of transplanted and human tumours [48]. Furthermore, they also showed that such an exerted pressure impairs in vivo proliferation via two mechanisms: reduced cancer cell proliferation in direct response to increased pressure, as well as a pressure-induced collapse of blood vessels within the tumour, leading to nutrient deficiency for tumour cells [49]. Based on these considerations, it seems relevant to consider that tumour expansion depends on the pressure. In our spatial growth model, the tissues motion is mediated by pressure gradients. It means that cells within a tumour tissue proliferate and that the exerted pressure pushes the neighbouring tissues. This pressure is not solely due to mechanical constraints (solid stresses, interstitial fluid pressure,…) exerted by the neighbouring cells on each other, but represents a more phenomenological pressure, that reflects the basic assumption of our modelling strategy for the tumour tissue being constituted by a fluid mixture in a porous medium. The effect of the pressure on proliferation has also been studied using numerical simulations elsewhere. In [47], Montel et al discussed the fact that cells proliferate faster on the surface than in the bulk of a tumour spheroid. A classical reason is that nutrients do not penetrate deeply in the spheroid. However, Montel et al. suggested a mechanical effect due to the necessity for a cell to deform its environment in order to proliferate. In an in silico study on two-dimensional monolayers and three-dimensional spheroids, based on experimentally determined biophysical parameters, Drasdo and Höhme suggested that pressure conditions have a higher impact on doubling time than lack of nutrients [16]. Moreover, in [47], Montel et al. performed experiments where tumour cells were submitted to different pressure constraints and observed a decrease in proliferation when pressure was applied. In their study, simulation results that were compared to experimental ones showed an exponential decreasing of proliferation with pressure, consistently with the modelling adopted here. However, the bulk and surface division rate were not affected equally by stresses. In our model, we used a similar pressure-mediated proliferation law translating direct effects of mechanical stresses on proliferation as well as indirect effects of proliferation on the micro environment (collapsing of blood vessels leading to lack of nutrients).
Our proposed hypotheses should be further experimentally reinforced, by, for example, implanting orthotopically and injecting intravenously two groups of cells into mice, each group being tagged with a different colour, and by quantifying single or mixed-coloured tumour foci. Similar experimental protocols have been already performed in [7,32]. Furthermore, in vivo investigations by observing two (or more) growing tumours in close vicinity that would enter mechanical interactions and then assess with a Ki-67 staining if the proliferation is impaired in the contact area, would further reinforce our contentions.
The inability of the merging theory to explain all of the observed volumes may indicate that besides merging by passive motion due to proliferation, other mechanisms such as chemokine-mediated cells attraction occur [6,50]. Circulating tumour cells may be attracted by some established niches and explain the abnormally fast volume expansions that we observed. Indeed, such chemokine-mediated attractions are presumed to play an important role for the pre-metastatic and metastatic niches establishment, in mediating myeloid and tumour cells attraction [6,50,51]. Moreover, chemo-attractants may play a role in tissue tropism of metastatic cells [52]. Chemotactic gradients can attract metastatic cells that express the chemokine receptor to specific locations. In the future, additional phenomena such as aggregation and recruitment of cells during the metastatic process from the circulation should be integrated in the standard mathematical model. Another phenomenon that could possibly explain the observed volumes would be the presence of circulating tumour cell clusters that would give rise to metastases [32]. Indeed, Aceto et al. recently showed in a breast cancer animal model that metastases do not originate from single cells only but also from tumour cells clusters that have a higher metastatic potential than single cells. However, they did not show evidence of this phenomenon for kidney cancer and in their experiments, clusters were formed by at most 50 cells. As indicated above, this order of magnitude of the initial cell numbers that colonizes the lung is not able to describe the dynamics of metastasis formation in our model and experimental data.
Taken together, our results indicate that spatial interactions are an essential component for the dynamics of metastasis development in the lung and probably also in other organs. However, it is unlikely that they alone control metastasis expansion. Indeed, when trying to assess whether this concept alone explains the fast growth of various metastases from the beginning of organ colonisation (from the first cell at days 12–14 to 0.022-0.67 mm3 at day 19), unrealistic numbers were found for two of the tumours. Thus, other mechanisms are probably also involved such as recruitment of additional cells from the blood stream and micro-environmental cues such as nutrient depletion or responses to environmental stress. Our methodology and results illustrate, furthermore, how a combined approach using multimodal biological data on one hand, and multimodal modelling analysis on the other, provides powerful insights into tumour biology and, in particular, into the metastatic process.
Ethical approval for all animal studies was obtained from the Institutional Animal Care and Use Committee of the INSERM Institute in accordance with the National Advisory Committee for Laboratory Animal Research Guidelines licensed by the French Authority. Animal facility: Animalerie mutualisée de Bordeaux 1, authorisation number: B33-522, Date: February 8th, 2012. Investigator: Andreas Bikfalvi (authorisation number: R-45GRETA-F1-10).
|
10.1371/journal.ppat.1004485 | Autophagy Controls BCG-Induced Trained Immunity and the Response to Intravesical BCG Therapy for Bladder Cancer | The anti-tuberculosis-vaccine Bacillus Calmette-Guérin (BCG) is the most widely used vaccine in the world. In addition to its effects against tuberculosis, BCG vaccination also induces non-specific beneficial effects against certain forms of malignancy and against infections with unrelated pathogens. It has been recently proposed that the non-specific effects of BCG are mediated through epigenetic reprogramming of monocytes, a process called trained immunity. In the present study we demonstrate that autophagy contributes to trained immunity induced by BCG. Pharmacologic inhibition of autophagy blocked trained immunity induced in vitro by stimuli such as β–glucans or BCG. Single nucleotide polymorphisms (SNPs) in the autophagy genes ATG2B (rs3759601) and ATG5 (rs2245214) influenced both the in vitro and in vivo training effect of BCG upon restimulation with unrelated bacterial or fungal stimuli. Furthermore, pharmacologic or genetic inhibition of autophagy blocked epigenetic reprogramming of monocytes at the level of H3K4 trimethylation. Finally, we demonstrate that rs3759601 in ATG2B correlates with progression and recurrence of bladder cancer after BCG intravesical instillation therapy. These findings identify a key role of autophagy for the nonspecific protective effects of BCG.
| Next to its effects against tuberculosis, BCG vaccination also induces non-specific beneficial effects on immune cells to increase their ability to control unrelated pathogens. It has been recently proposed that the non-specific effects of BCG are mediated through epigenetic reprogramming of monocytes, a process called trained immunity. Little is known regarding the intracellular events controlling its induction. In this study we identified autophagy as a key player in trained immunity. Pharmacological inhibition of autophagy as well as polymorphisms in autophagy-related genes blocked BCG-induced trained immunity. Furthermore, BCG vaccine is also used to treat bladder cancer. Genetic polymorphisms in autophagy-related genes correlated with progression and recurrence of bladder cancer after treatment with BCG therapy. These findings open new possibilities for improvement of future BCG-based vaccines to be used against infections and malignancies.
| Immunological memory has long been viewed as being exclusively mediated by T and B cells. However, an increasing body of evidence indicates enhanced nonspecific protection against reinfections in plants [1] and insects [2] which lack adaptive immunity. Similarly, mammalian innate immune cells such as natural killer cells show features of immunological memory [3], [4]. Recently, we proposed the term trained immunity to describe the memory properties of innate immune cells [5]. Candida albicans or its major cell wall component β-glucan, as well as BCG, are prominent stimuli that can induce trained immunity through epigenetic reprogramming of monocytes [6], [7]. However, little is known regarding the intracellular events controlling the induction of trained immunity, impairing the ability to fully harness the therapeutic potential of this important immunological process. Therefore, we investigated the trained immunity-induced signaling pathways, discovering autophagy being one of the main players.
To identify new signaling pathways specifically activated upon training of monocytes with bacterial components, we compared the transcriptional profile of β-glucan-trained human primary monocytes isolated from healthy volunteers to the profile of monocytes stimulated with Escherichia coli-derived lipopolysaccharide (LPS), which stimulates inflammation but is unable to induce long-term training [5]. Transcriptomic assessment of these monocytes by microarrays and pathway analysis revealed specific clusters of genes significantly induced by β-glucan training with an intriguing signal found in the ubiquitin-related proteins and associated catabolic processes (Figure 1a). Since ubiquitination plays an important role in autophagy [8], a process that has previously been shown to improve intracellular processing of BCG [9], [10], we examined the role of autophagy in the induction of trained immunity.
Using an in vitro model of trained immunity [6], [7], adherent monocytes from healthy human volunteers were stimulated for 24 h with RPMI, BCG or β-glucan alone or in combination with the autophagy inhibitors 3-methyladenine (3MA) or wortmannin. After washing of cells and a resting period of 6 days in medium supplemented with 10% human serum, cytokine production was measured after a second stimulation with the unrelated stimuli LPS or Borrelia burgdorferi (B. burgdorferi) (Figure 1b). IL-6 and TNF-α production increased significantly in BCG- and β-glucan-trained cells compared to non-trained cells. When autophagy was blocked by 3MA or wortmannin, neither β–glucan nor BCG induced trained immunity (Figure 1c–f; Figure S1a–h). Notably, the putative cytotoxic effects of autophagy inhibitors used in this study were assessed by LDH measurements. None of the inhibitors used during the 24 h of primary cell stimulation enhanced LDH release compared to RPMI-treated cells (Figure S2 a–c), demonstrating that the molecules were not toxic to the cells.
To further explore the role of autophagy in the nonspecific protection of BCG in innate immune cells, we examined the effects of genetic polymorphisms in autophagy genes for the BCG-induced trained immunity in vitro and in vivo. The genotypes of nine SNPs in eight autophagy genes were correlated with the capacity of BCG to induce trained immunity in a group of 72 volunteers. The rs3759601 ATG2B SNP was found to be strongly associated with trained immunity; the ability to develop training characteristics following BCG treatment was observed in monocytes isolated from individuals carrying the GG (major) or CG genotype but not in those carrying the CC (minor) genotype (plus strand coding) (Figure 2a–f). A similar effect, though less clear, was apparent for the rs2245214 ATG5 SNP (Figure 2g–i). No significant association was found between the nonspecific protection of BCG and polymorphisms in ATG10, ATG16L1, EREG, IRGM, LAMP3 and WIPI (Figure S3).
To test the possibility that the association between SNPs and differences in cytokine production of BCG-trained monocytes was due to differential intrinsic capacity of the cells to produce cytokines, we stimulated monocytes bearing different ATG2B (Figure 2j) or ATG5 (Figure 2k) alleles with LPS or B. burgdorferi for 24 hours. We noted no differences in cytokine release, indicating that the capacity of cells to release proinflammatory cytokines upon stimulation was not responsible for the observed association between autophagy SNPs and BCG-induced trained immunity. Next to that, the effect of the rs3759601 SNP on the transcription of the ATG2B gene was assessed after training. We observed increased levels of ATG2B transcripts in BCG-trained cells of individuals carrying the GG genotype but not in those carrying the CC genotype (Figure 2l). Increased ATG2B levels could also be found in β-glucan trained individuals carrying the GG genotype (Figure S4a) but no difference in ATG2B levels could be found in the two groups after LPS stimulation (Figure S4b). The reduced expression of ATG2B in individuals carrying the CC genotype of the SNP upon training with BCG could indicate a role for autophagy in trained immunity since it has been shown that the ATG2 proteins are essential for the formation of autophagosomes [11].
To identify the effect of rs3759601 in ATG2B on autophagy, the amount of LC3+ vesicles in BCG stimulated monocytes of individuals carrying the major or minor variant of the SNP have been compared. A decrease in autophagosome formation of individuals carrying the CC genotype can be seen as demonstrated by a lower percentage of LC3+ monocytes (Figure 3a–b).
To corroborate the above data, we investigated BCG-induced training of monocytes in vivo by testing individuals carrying different ATG2B alleles. Monocytes were isolated from 16 healthy volunteers, before and 3 months after vaccination with BCG. Following stimulation with LPS (Figure S5a–b) or B. burgdorferi (Figure 4a–b), IL-1β and TNF-α production was significantly higher 3 months after vaccination in individuals who were bearing at least one G allele of the ATG2B SNP (n = 12), while monocytes isolated from individuals carrying the CC genotype (n = 4) showed no change in cytokine production after BCG vaccination.
In addition to the protective effects of BCG against secondary infections, non-specific therapy with intravesical BCG is also used as a therapeutic strategy for patients with non-muscle invasive bladder cancer (NMIBC; stages: Ta, T1, CIS) [12]. In a cohort of 192 NMIBC patients treated with at least 6 intravesical instillations of BCG we evaluated the association between the ATG2B SNP and prognosis in terms of recurrence and progression during the first five years after the primary NMIBC diagnosis. Analyses learned that those patients that carry one or two C alleles for ATG2B rs3759601 showed increased risk of recurrence (CG vs. GG: hazard ratio (HR) = 1.73 (95% confidence interval (CI): 0.99–3.03) and CC vs. GG: HR = 1.68 (95% CI: 0.78–3.27)) (Figure 4c) and progression (CG vs. GG: HR = 1.57 (95% CI: 0.79–3.12) and CC vs. GG: HR = 2.15 (95% CI: 1.00–4.66)) (Figure 4d). This finding of a correlation between the polymorphism in ATG2B to progression and recurrence of bladder cancer supports the hypothesis of a clinical relevance of the autophagy gene for the non-specific protective effects exerted by BCG. In addition, the responsiveness of circulating monocytes of bladder cancer patients has been investigated before and after BCG-therapy. Of high interest, individuals who received intravesical BCG therapy showed an increased cytokine response of their monocytes after stimulation with LPS in vitro (Figure 4e–g).
Epigenetic reprogramming of monocytes is a crucial immunological mechanism underlying nonspecific protection by BCG. Stable changes in histone trimethylation at the level of lysine 4 of histone 3 (H3K4), a post-translational modification associated with the regulation of immune-related genes [13], is one of the mechanisms responsible for enhanced cytokine production after re-stimulation of trained monocytes [5]–[7]. Therefore, we assessed whether trimethylation of H3K4 due to nonspecific training by BCG was influenced by the ATG2B polymorphism or inhibition of autophagy by 3MA. Consistent with our hypothesis, H3K4 trimethylation was significantly increased at the IL-6 and TNF-α promoters in BCG-trained monocytes from volunteers bearing the ATG2B G allele (Figure 5a–b). In contrast, volunteers homozygous for the ATG2B C allele did not show any increase in trimethylation at H3K4 at the cytokine promoters after BCG-training. Furthermore, inhibition of autophagy by 3MA blocked the H3K4 trimethylation at IL-6 and TNF-α promoters in BCG-trained monocytes (Figure 5c–d), supporting the hypothesis of a central role of autophagy in the epigenetic reprogramming of monocytes induced by BCG.
BCG is a live attenuated vaccine which is routinely administered at birth in low-income countries, protecting newborns against disseminated tuberculosis and tuberculosis meningitis [14]. However, in addition to its specific protection against childhood tuberculosis, epidemiological studies have demonstrated that BCG protects against infant mortality independent of its effect on tuberculosis, suggesting a nonspecific protection against unrelated infections [15]–[24]. Next to that, BCG treatment has long been used as a non-specific immunostimulatory therapy in urothelial cell carcinomas [25]. Recently, these non-specific protective mechanisms of BCG have been associated with epigenetic reprogramming of innate immune cells in a process called trained immunity [7]. In the present study we show that autophagy is a central event modulating trained immunity induced by BCG. Moreover, polymorphisms in autophagy genes such as ATG2B control trained immunity in both in vitro and in vivo models, as well as the non-specific therapeutic effects of BCG in patients with bladder cancer.
An important difference has to be noted between the effect of ATG2B polymorphism on BCG training against secondary infections and BCG used as a treatment against non-muscle invasive bladder cancer. BCG training of monocytes against unrelated secondary infections could only be modulated by an ATG2B polymorphism expressed on both alleles. Heterozygote individuals were still trainable with the vaccine. On the contrary, the prognosis in terms of recurrence and progression of non-muscle invasive bladder cancer decreased with only one affected allele. The different route of BCG administration, as well as several disease-related mechanisms could be the explanation of this event. To further unravel the different mechanisms behind this phenomenon, a pilot study has been performed to investigate whether BCG installation in the bladder could induce a state of trained immunity. The cytokine response of ex-vivo stimulated monocytes of BCG treated bladder cancer patients increased in response to LPS compared to the pre-treatment response.
In addition to the aspects discussed above, there are also a few limitations of the current study. Thus, although we demonstrate the role of autophagy for BCG-induced trained immunity, additional studies are needed to decipher the precise pathway linking autophagy to the epigenetic modifications observed during trained immunity. A second important aspect is the fact that the genetic study has been performed in a relatively small cohort of patients with bladder carcinoma, and it needs to be validated by independent studies. Finally, the role of autophagy gene SNPs for the effects of BCG on infections also needs to be evaluated. The role of BCG for protection against infection is currently investigated by a large Danish study in 4500 newborn children (http://calmette-studiet.dk/), and the effect of the autophagy polymorphisms on the effects of BCG is an important aspect to be assessed.
A key question regarding trained immunity refers to the signaling and molecular mechanisms responsible for its induction. As shown previously, exposure of monoctyes to BCG induces high levels of H3K4 trimethylation at the promoter level of inflammatory genes, which correlates with long-term increased production of proinflammatory cytokines, a hallmark of trained immunity [6], [7]. Next to that, the blockage of histone acetyltransferases inhibits the training of monoctyes [26] suggesting also an important role of acetylation in trained immunity which will be further studied in the future.
The discovery that autophagy modulates trained immunity may have important consequences. It provides understanding of an important immunological process, although future studies are warranted to identify the molecular mechanisms through which autophagy mediates the epigenetic changes responsible for trained immunity. Restriction of reactive oxygen species release from damaged mitochondria, or processing of microbial ligands such as peptidoglycans [9], may represent two potential candidate mechanisms. Furthermore, identification of autophagy as a driver of trained immunity opens new possibilities for improvement of future BCG-based vaccines to be used against infections and malignancies.
All human experiments were conducted according to the principles expressed in the Declaration of Helsinki. Before taking blood, informed written consent of each human subject was obtained. The study was approved by the review board of the department of Medicine of the Radboud University Nijmegen Medical Centre. The BCG in vivo study was approved by the Arnhem-Nijmegen Ethical Committee. For the NBCS, all participants gave written informed consent and the study was approved by the Institutional Review Board of the RUMC. All data analyzed were anonymized.
In vitro cytokine stimulation experiments were performed with PBMCs isolated from buffy coats obtained from healthy volunteers (Sanquin Bloodbank, Nijmegen, the Netherlands). To analyze the effect of gene polymorphisms on trained immunity, blood was drawn from a group of healthy volunteers (age 23–73). For the in vivo BCG model, subjects (aged 20–36) who were scheduled to receive a BCG vaccination at the public health service, due to travel or work in tuberculosis-endemic countries, were asked to participate in this trial. Blood was drawn before and 3 months after the BCG vaccination. Informed consent was obtained from all human subjects.
The bladder cancer patients included in this study were selected from a total of 1,602 patients with primary urinary bladder cancer (UBC) from the Nijmegen Bladder Cancer Study (NBCS). The NBCS served as the Dutch discovery population in the UBC genome-wide association study led by Radboud University Medical Centre (RUMC, Nijmegen, the Netherlands) and deCODE Genetics (Reykjavik, Iceland). The NBCS has been described in detail before [27]. Cases with a previous or simultaneous diagnosis of upper urinary tract cancer, based on information from the Netherlands Cancer Registry, were excluded. Detailed clinical data concerning diagnosis, stage, treatment, and disease course (tumor recurrence and progression) were collected retrospectively based on a medical file survey. In the analysis we included a total of 192 cases with non-muscle invasive bladder cancer (NMIBC; stage Ta/T1/CIS) that received at least 6 intravesical BCG instillations as initial treatment (median follow-up time from initial transurethral resection of bladder tumor until last urological check-up visit was 5.2 years (range: 0.4–20)). All patients were from Caucasian background.
C. albicans ATCC MYA-3573 (UC 820) yeast was heat-inactivated for 30 min at 95°C. B. burgdorferi, ATCC strain 35210, was cultured at 33°C in Barbour-Stoenner-Kelley (BSK)-H medium (Sigma-Aldrich) supplemented with 6% rabbit serum. Spirochetes were grown to late-logarithmic phase and examined for motility by dark-field microscopy. Bacteria were harvested by centrifugation of the culture at 7000× g for 15 min and washed twice with sterile PBS (pH 7.4).
The mononuclear cell fraction was obtained by density centrifugation of blood diluted 1∶1 in pyrogen-free saline over Ficoll-Paque (Pharmacia Biotech, Pittsburgh, Pennsylvania, USA). Cells were washed twice in saline and resuspended in culture medium (RPMI; Invitrogen, Carlsbad, California, USA) supplemented with 50 mg/L gentamicin, 2 mM L-glutamine and 1 mM pyruvate. PBMCs were counted in a Coulter counter (Coulter Electronics, Brea, California, USA) and their number was adjusted to 5×106 cells/ml. A total of 5×105 cells in a 100 µl volume was added to round-bottom 96-well plates (Greiner) with RPMI, E. coli LPS (10 ng/ml) or B. burgdorferi (1×106/ml). After 24 h, the supernatants were collected and stored at −20°C until being assayed.
For training experiments, PBMCs (5×105 for cytokine analysis; 10×106 for ChIP analysis) were incubated for 1 h at 37°C in 5% CO2. Adherent monocytes were selected by washing out nonadherent cells with warm PBS. Thereafter, cells were preincubated with RPMI, BCG vaccine (1 µg/ml BCG vaccine SSI from the Netherlands Vaccine Institute) or β-1,3-(D)-glucan (β-glucan) (10 ng/ml; kindly provided by Professor David Williams) for 24 h (4 h for Real-time PCR). After a resting period of 6 d in RPMI including 10% serum, cells were stimulated with E. coli LPS (10 ng/ml), C. albicans (1×106/ml), B. burgdorferi (1×106/ml), or RPMI for an additional 24 h. Supernatants were stored at −20°C until ELISA was performed. In the “inhibition” experiments, before training with BCG or β-glucan, the adherent monocytes were preincubated for 1 h with 10 mM 3-methyl adenine (3MA, Sigma).
Concentrations of human IL-1β, IL-6 and TNF-α were determined in duplicates using commercial ELISA kits (Sanquin, Amsterdam, or R&D Systems, Minneapolis), in accordance with the manufacturers' instructions.
RNA from stimulated monocytes was isolated using TRIzol reagent (Invitrogen) according to the manufacturer's instructions. Isolated RNA was reverse-transcribed into complementary DNA using iScript cDNA synthesis kit (Bio-Rad). Quantitative real-time PCR was performed using Power SYBR Green PCR Master Mix (Applied Biosystems) using a 7300 Real-time PCR system (Applied Biosystems). In each PCR a melting curve analysis was included to control for a specific PCR amplification. Primers used for the experiments (final concentration 10 µM) are shown below. Real-time quantitative PCR data were corrected for expression of the housekeeping gene B2M. Human ATG2B forward: ACCAGAGATAGCACCTTCTGAC and reverse: CCAATTAACCGTCCAATCTG; human B2M forward: ATGAGTATGCCTGCCGTGTG and reverse: CCAAATGCGGCATCTTCAAAC.
In vitro training experiment: Using NCBI SNP database we selected SNPs in autophagy genes previously associated to diseases or with a minor allele frequency of at least 5% (ATG10 (rs1864183), ATG10 (rs3734114), ATG16L1 (rs2241880), ATG2B (rs3759601) [allele frequency: G = 70%; C = 30%], ATG5 (rs2245214), EREG (rs78803121), IRGM (rs4958847), LAMP3 (rs482912), WIPI (rs883541)). Blood samples were obtained by venapuncture. Genomic DNA was isolated from EDTA blood using standard methods, and 5 ng of DNA was used for genotyping. Multiplex assays were designed using Mass ARRAY Designer Software (Sequenom) and genotypes were determined using Sequenom MALDI-TOF MS according to manufacturer's instructions (Sequenom Inc., San Diego, CA, USA) as described previously [28].
In vivo BCG-cohort: DNA was isolated using the Gentra Pure Gene Blood kit (Qiagen), in accordance with the manufacturer's protocol for whole blood. DNA was dissolved in a final volume of 100 µl buffer. Genotyping of single nucleotide polymorphisms (SNPs) was performed using a pre-designed TaqMan H SNP genotyping assay (Applied Biosystems) according to the manufacturer's protocol.
NBCS: All bladder cancer patients were genotyped using the Illumina Infinium HumanCNV370-duo Bead-Chips. Imputation was performed (IMPUTE version 2.1 software) using the 1000 Genomes low-coverage pilot haplotypes (released June 2010, 120 chromosomes) and the HapMap3 haplotypes (released February 2009, 1920 chromosomes) as a combined reference panel [27]. SNP rs3759601 was imputed with IMPUTE info_score 0.99. The SNP followed Hardy-Weinberg equilibrium.
Gene expression was performed as described previously [29] and assessed using Illumina Human HT-12 Expression BeadChip according to manufacturer's instructions. The Illumina LIMS platform, BeadStudio was employed to perform image analysis, bead-level processing, and quantile normalization of array data.
Adherent monocytes were cultured as described above (see Stimulation Experiments). ChIP was performed using antibodies against H3K4me3 (Diagenode). ChIPed DNA was processed further for qPCR analysis. The following primers were used in the reaction (5′-3′): TNF-α forward: CAGGCAGGTTCTCTTCCTCT, TNF-α reverse: GCTTTCAGTGCTCATGGTGT; IL-6 forward: TCGTGCATGACTTCAGCTTT, IL-6 reverse: GCGCTAAGAAGCAGAACCAC; myoglobin forward: AGCATGGTGCCACTGTGCT, myoglobin reverse: GGCTTAATCTCTGCCTCATGAT.
For immunofluorescence imaging, monocytes were seeded on coverslips pretreated with polylysine, fixed with 4% PFA for 15 min at room temperature followed by 10 min of fixation with ice-cold methanol at −20°C. After two washing steps with PBS, cells were permeabilized by 0.1% saponin (Sigma-Aldrich), blocked for 30 min in PBS plus 2% BSA, incubated for 1 h with a mouse mAb to LC3 (1∶50; Nanotools), washed twice in PBS plus 2% BSA and stained by a secondary Alexa Fluor 555 goat anti-mouse Ab (1∶500; Molecular Probes), followed by DNA staining with 10 µM TO-PRO-3 iodide (642/661; Invitrogen). After the washing steps, slides were mounted in Prolong Gold antifade media (Molecular Probes). Images were acquired using a laser-scanning spectral confocal microscope (TCS SP2; Leica Microsystems) and LCS Lite software (Leica microsystems). 2 fields/donor including at least 40 cells each were counted and compared for the amount of LC3.
Data are expressed as mean ± SEM unless mentioned otherwise. Differences between experimental groups were tested using the non-parametrical two-sided Mann-Whitney U test (no normal distribution of measured cytokines); differences between multiple time points within one group (before versus after treatment) were tested using the Wilcoxon matched pair test (unless stated otherwise) performed on GraphPad Prism 4.0 software (GraphPad). P values of ≤0.05 were considered statistically significant.
Kaplan-Meier survival and Cox proportional hazard regression analyses were performed to evaluate the association between rs3759601 and recurrence- and progression-free survival. Log-rank tests were calculated to compare survival curves between genotype categories. Imputed genotype probabilities were transformed to hard genotype calls based on a probability threshold of >0.90. Statistical analyses were performed using IBM SPSS Statistics for Windows 20 (IBM Corp., Armonk, NY, USA) and survival plots were drawn using R software v3.0.2 (package ‘survival’) (R Development Core Team, Vienna, Austria).
|
10.1371/journal.pcbi.1004453 | Quantitative Live Imaging of Human Embryonic Stem Cell Derived Neural Rosettes Reveals Structure-Function Dynamics Coupled to Cortical Development | Neural stem cells (NSCs) are progenitor cells for brain development, where cellular spatial composition (cytoarchitecture) and dynamics are hypothesized to be linked to critical NSC capabilities. However, understanding cytoarchitectural dynamics of this process has been limited by the difficulty to quantitatively image brain development in vivo. Here, we study NSC dynamics within Neural Rosettes—highly organized multicellular structures derived from human pluripotent stem cells. Neural rosettes contain NSCs with strong epithelial polarity and are expected to perform apical-basal interkinetic nuclear migration (INM)—a hallmark of cortical radial glial cell development. We developed a quantitative live imaging framework to characterize INM dynamics within rosettes. We first show that the tendency of cells to follow the INM orientation—a phenomenon we referred to as radial organization, is associated with rosette size, presumably via mechanical constraints of the confining structure. Second, early forming rosettes, which are abundant with founder NSCs and correspond to the early proliferative developing cortex, show fast motions and enhanced radial organization. In contrast, later derived rosettes, which are characterized by reduced NSC capacity and elevated numbers of differentiated neurons, and thus correspond to neurogenesis mode in the developing cortex, exhibit slower motions and decreased radial organization. Third, later derived rosettes are characterized by temporal instability in INM measures, in agreement with progressive loss in rosette integrity at later developmental stages. Finally, molecular perturbations of INM by inhibition of ACTIN or NON-MUSCLE MYOSIN-II (NMII) reduced INM measures. Our framework enables quantification of cytoarchitecture NSC dynamics and may have implications in functional molecular studies, drug screening, and iPS cell-based platforms for disease modeling.
| Brain development is a dynamic and complex process that requires highly orchestrated interaction between neural stem cells. Therefore, investigating these dynamics is fundamental for understanding brain development and disease. However, difficulties to record and quantify neural stem cells behavior inside the brain pose a major limitation. We were recently able to mimic brain development in the Petri dish by generating highly organized multicellular structures containing human neural stem cells termed Neural Rosettes. Here we present a newly developed method to record, quantify and analyze the dynamic movements of neural stem cells within rosettes as reflection of their behavior inside the developing brain. We first confirmed that neural stem cells move radially in rosettes similarly to authentic stem cells residing in the developing brain. We then defined novel measures to assess how well these neural stem cells organize into rosettes in culture and found that organization decreases as stem cells progress in culture. Finally, we demonstrated that disruption of rosette structures by specific drugs impairs organization dynamics of neural stem cells. Our findings offer a first insight into neural stem cell dynamics during brain development, and a potential methodology for functional studies and drug discovery.
| Neural stem cells (NSCs) are neural progenitor cells within the nervous system that are defined by their ability to self-replicate while retaining potential for generating neurons and glia [1–3]. During nervous system development, early emerging NSCs first undergo successive symmetric cell divisions that generate additional progenitor cells, resulting in expansion of the NSC pool. This phase is then followed by asymmetric cell divisions that generate a progenitor cell and a terminally differentiated cell such as a neuron or a glial cell, resulting in decreased NSC ratios [4] (for review see Ref. [5]). Together, these mechanisms are fundamental for the generation of distinct types of NSCs to account for cellular diversity of the nervous system.
Much progress has been made towards understanding the factors that regulate the self-replication or differentiation of NSCs both in vivo and in vitro [6,7]. One predominantly exciting but less understood aspect of stem cell biology is how the cytoarchitecture of the developing brain is linked to maintenance of NSC number and developmental potential. Particularly interesting is the nuclei dynamics within neuroepithelial cells and radial glial cells—the NSCs that build the cortex. These are elongated cells harboring distant apical and a basal processes that connect the two walls of the developing neural vesicles. Structurally, this layer of NSCs is pseudostratified; i.e., although several layers of nuclei are apparent between vesicle walls, the cytoplasm of each cell extends to contact both apical and basal surfaces of the wall, resulting in a bipolar cellular morphology that is much longer then the thickness of a single cell. Interkinetic nuclear migration (INM) is the process by which nuclei migrate between apical and basal ends of these pseudostratified neuroepithelial cells, in coupling with cell division at the apical surface. INM was first suggested by Sauer [8], further confirmed experimentally [9–11] and later suggested as a mechanism to ensure maintenance of sufficient NSC pools throughout embryonic neocortical development [12]. Thus, it is hypothesized that INM spatial composition and dynamics reflect critical abilities of NSCs during self-renewal and differentiation.
Time-lapse microscopy of mouse and human cortical slices excised from brain and grown in vitro has confirmed INM as a live dynamic process [13] and provided initial clues on mechanisms and functions. Specifically, a correlation between INM and cell cycle was established [14,15] and variations between apical vs. basal INM speed as well as motion patterns were shown in vivo [16]. INM dynamics is fascinating also because the pseudostratification of neuroepithelial cells entails high cellular traffic due to accommodation of many moving nuclei within a limited space. It was suggested that unsynchronized INM may serve as a mechanism to maximize probability of cell division at apical sites and consequent exposure to Notch signaling activation within these sites, which in turn promote maintenance of NSC fate following cell division [12].
Yet, very little is known regarding specifics of INM dynamics, reflecting the technical challenge to image and quantify multiple cells in vivo at high resolution and high content. Therefore, two components are needed in order to advance the field: (1) an in vitro model system that is physiologically relevant and reflects dynamics similar to those observed in vivo, (2) quantitative readouts for cell dynamics in such a system.
Here we used neural rosettes as an in vitro model to explore INM dynamics with relation to NSC maintenance. Neural rosettes are highly organized structures that appear in culture following differentiation of human pluripotent stem cells into cortical lineages. Neural rosettes contain NSCs resembling neuroepithelial and radial glial cells of the developing cortex that are radially organized to create a lumen, resembling the structure of the ventricular zone of the developing cortex. Initial characterization of neural rosettes revealed strong apicobasal cell polarity with the organization of apical ends surrounding a lumen [17]. Joining apical ends at rosette lumens also coincide with mitosis in these luminal regions [17,18], in accordance to INM observations in vivo [19]. We recently dissected the entire cortical differentiation process—from neuroepithelial cells towards distinct radial glial cell types. Specifically, we identified two rosette stages in culture corresponding to two developmentally distinct types of radial glial cells [20]. Early radial glial (E-RG) rosettes (day 14 in culture) contain highly proliferative NSCs exhibiting broad differentiation potential and minimal differentiation propensity in culture [20]. As cultures proceed in vitro, E-RG rosettes progress into a stage termed Mid radial glial (M-RG) rosettes (day 35 in culture), which is characterized by decreased NSC numbers and increased propensity to differentiate into neurons. Further culture of M-RG rosettes (beyond day 55) ultimately results in loss of rosette integrity, a further reduction in NSC numbers, and transition of fate potential from neuronal towards glial bias [20]. We further comprehensively dissected the regulatory networks that drive differentiation from pluripotent stem cells to E-RG and then to M-RG rosettes by employing extensive transcriptional and epigenetic characterization coupled with computational analysis [21]. Together, these findings propose a functional link between cortical development, NSC capacity maintenance and neural rosettes formation and disassembly. Thus, we hypothesized that INM dynamics within neural rosettes may predict NSC capacity of the developing cortex.
Here, we developed quantitative live imaging to characterize cellular dynamics within rosette structures, and investigated these kinetic properties during transition across developmentally distinct rosettes and following molecular perturbations. We devised three main readouts associated with INM to capture different aspects of cell dynamics that relate to migration orientation and speed. We further applied these measures to reveal inherent and stage-dependent observations that distinguish E-RG from M-RG rosettes. Finally, we were also able to detect and quantify reduction in cellular organized dynamic performance following specific inhibition of molecular motors involved in INM. Thus, our quantitative approach delineates a model that describes intrinsic dynamic features within rosettes and suggests for the first time a functional link between rosette dynamics and NSC competence. Hence, the presented kinetic quantitative readouts have the potential to serve for functional molecular studies drug screening, and may be implicated to gain novel insights into biology of NSCs in health and disease.
We used the HES5::eGFP Notch activation reporter human embryonic stem cell (hESC) line, expressing cytoplasmic GFP in Notch active cells [22]. HES5 is a major and direct downstream target of Notch activation pathway (for review, see Ref. [23]) and specifically marks NSCs in vivo. We recently showed that neural rosettes correspond to NSCs of the developing cortex based on their strong apico-basal epithelial polarity and the expression of cortex associated genes such as PAX6 together with the NSC marker HES5::eGFP [20] (Fig 1A). When newly formed E-RG rosettes appear in culture on day 14, most rosette cells express the cortical marker PAX6 and the NSC marker HES5::eGFP (Fig 1A, middle panel) (>80%; see Ref. [20]) in accordance with their high proliferative capacity and lack of differentiation in culture. In contrast, continued culture of E-RG rosettes results in their progression towards M-RG rosettes around day 35, and this is marked by significant loss in the NSC marker HES5::eGFP and the cortical marker PAX6 in rosette cells (Fig 1A, bottom panel) (<30%; see Ref. [20]). Importantly, PAX6 expression is now limited only to the regions adjacent to rosette lumens, reflecting the limited area where stem cells reside at that stage. Since the ability of neural progenitors to radially organize in rosettes is correlated with increased ratios of polarized epithelial NSCs [17,18,20], we hypothesized that this difference in NSC numbers among early and advanced rosettes would be phenotypically reflected in rosette dynamics, specifically that of INM.
Subjective live imaging observations suggested that both E-RG and M-RG rosettes exhibit INM characteristics, and this was even apparent in matching phase contrast images (S1 and S3 Movies; compare to non-rosettes, S2 Movie; phase contrast time lapses immediately follow GFP time lapses in each movie). To quantitatively validate the observed INM motility-patterns, we devised an automated objective framework to assess rosettes dynamics. The analysis was based on manual annotation of rosette contours and centers from the phase-contrast channel, where the cytoarchitecture outlines of the region performing INM become obvious to a human eye based on different texture-patterns in the image (Figs 1B, right panels and S1A; See Methods). Rosette outlines remained stable in the culture dish and did not change throughout the experiment (S1B Fig). Rosette areas were discretized to sub-cellular patches (S2A Fig; See Methods) and local cross-correlation was applied to estimate motion for each patch at each time point [24], similarly to particle image velocimetry (PIV) [25,26]. This approach was validated as highly correlative to manual single-cell tracking (S2B Fig). Motions were exceptionally fast ranging up to 120μm hr-1 (S2C Fig), and only motions of 15μm hr-1 or faster (≥ 2 pixels per time-lapse frame) were considered for further analysis.
We first estimated the average velocity orientation for each of the coordinates within each rosette over the entire time course (Fig 1C, left panels). Indeed, migration pattern followed the expected radial angle, as defined by orientation of each patch’s velocity relatively to the rosette center, i.e., the expected velocity direction assuming apico-basal radial motion (See schematics in Fig 1E). This pattern was also reflected by the normal distributions observed for motion angles grouped by their expected radial angle (S2D Fig). As expected, this pattern did not occur for cells in non-rosette areas that were adjacent to rosettes (Fig 1C, right panels).
More quantitatively, we calculated the distribution of the angular alignment γ between the observed velocities and their respective expected radial angles (See schematics in Fig 1E) to validate a dramatic bias of the angular alignment distribution toward structured motion (Fig 1D, compare E-RG Rosettes, left, to Non-rosettes, right). Strikingly, this dynamics was even obvious also when computing motions based on the phase contrast channel, further confirming that rosettes indeed perform radial migration (Fig 1B–1D, left, compare GFP columns to Phase columns). These observations indicate that motion within rosettes resemble in vivo INM [8,27] and suggest that radial migration within rosettes in vitro plays a functional role in the maintenance of NSCs.
Based on our initial observations we devised three objective measures to study cell dynamics in rosettes to enable functional quantification. Each measure was defined as a scalar readout per rosette that quantifies different aspects in its dynamics throughout time. The first measure, Radial Score (RS), was defined as the average angular alignment (γ) of all motions in each rosette over the entire time course (Fig 2A). RS quantifies the mean alignment between observed and expected radial angles. Thus, lower scores correspond to better alignment, reflecting a more organized radial migration (denoted radial organization henceforth). The second measure, Basal to Apical ratio (B/A ratio), was defined as the ratio between the number of basal (distal) motions to apical (luminal) motions within rosettes along the entire time course (Fig 2B). RS and B/A ratio were designed to quantify INM in vitro, which corresponds to the basal to apical migration observed for the developing neuroepithelium in vivo [16,19]. The third measure, Speed was defined as the average magnitude of velocity for all patches across time (Fig 2C), a measure that was quantified in vivo [16,19]. When calculated for each time frame over time, these three measures fluctuated around a mean value, validating that the progressive rosette-disassembly in culture is much slower than the four hour imaging course (S2E Fig), thus allowing us to focus on the mean measures as our readouts.
Based on our observations (Fig 1), we hypothesized that these quantitative measures could reveal inherent biophysical properties of neural rosettes and their associations. These measures may also enable to functionally distinguish between INM performance capacity of E-RG and M-RG rosettes, and further test the hypothesis that the INM of M-RG rosettes is compromised and reflects the beginning of rosette disassembly, in line with the increase in cells with non-epithelial character [17,18,20]. 25 E-RG rosettes and 14 M-RG rosettes were imaged and quantified to test our hypotheses as detailed next.
We hypothesized that larger rosettes are characterized by enhanced structured motion, as a response to increased mechanical constraints by the rosette cytoarchitecture. To test this, we examined the association between RS and rosette size. This property was first examined for E-RG rosettes, speculated to have a more structured dynamics compared to the more advanced M-RG rosettes. Indeed, RS of E-RG rosettes was found to be associated with rosette size (Fig 3A), indicating that larger rosettes exhibit enhanced radial migration. Similarly to E-RG rosettes, RS of M-RG rosettes was also found to be associated to rosette size (Fig 3A), suggesting that the association between rosette size and RS is an intrinsic property. This result also implies that the robustness of INM is augmented when larger numbers of nuclei move together towards apical or basal sites, as previously suggested for INM in pseudostratified neuroepithelial cells in vivo [12]. A similar model works also in other studies of collective cell migration showing enhanced group coordinated motility correlated with group size [28–32]. We therefore suggest that the confined structures of larger rosettes lead to more mechanical constraints that explain the increased radial organization of larger rosettes.
Next, we tested whether M-RG rosettes differ in their radial migration dynamics compared to E-RG rosettes. We found that RS of M-RG rosettes were larger (less organized) than the expected size-dependent values derived from the E-RG’s linear model (Fig 3A, most M-RG rosettes above the line and Fig 3B, quantitatively). We conclude that rosette size plays a prominent role with linear effect on radial migration and that RS can serve as a functional predictive measure for NSC capacity within rosettes.
Previous in vivo studies have shown differences in speed between nuclei migrating apically and basally [16,33]. We hypothesized that radial organization may also differ between apical and basal motion. To test this hypothesis we classified each motion vector as moving apically (inward) or basally (outward) with respect to rosette center, reflective of apical and basal nuclei migration during INM, and examined apical and basal motion independently. We partitioned the motion vectors of each rosette into basal and apical groups and calculated each group’s RS. We found that similarly to general RS, basal or apical RS were associated with rosette size (Fig 4A). However, basal motion tends to be more radially organized (i.e., lower RS) than apical motion (Fig 4B, most points above the y = x line), regardless of rosette stage (Fig 4B) or size (Fig 4A, basal RS tends to be lower than apical RS for all rosette sizes). This was also judged by the basal RS values for M-RG rosettes, which were higher (less organized) than the basal RS values predicted by the linear fit of E-RG rosettes (Fig 4C). These results led to the hypothesis that enhanced radial organization of basal motions contributes to the overall elevated radial organization observed for E-RG rosettes.
B/A ratio was hypothesized as a secondary mechanism for the elevated radial organization of E-RG rosettes, by enhancing the contribution of the more radially organized basal motions in a size-independent mechanism (S3 Fig, S1 Note).
The measures RS and B/A ratio were calculated based on the orientation of the velocity vectors. Next we considered the speed—the magnitude of these vectors—as a third measure for rosette dynamics (See Fig 2). We found a two fold increase in the fraction of cells moving at speed of 15 μm hr-1 or faster in E-RG rosettes compared to M-RG rosettes (Fig 5A). This strikingly fits our recent findings according to which there are twice as much NSCs (i.e., GFP+ cells) in E-RG rosettes compared to M-RG rosettes [20], further drawing a correlation between the actual number of NSCs measured within rosette cultures and the computed quantification of their INM motions. In contrast to RS, rosette speed was not correlated to rosette size (S4 Fig), suggesting that molecular motors driving nuclei motions are less affected by the rosette confining structure. When considering apical and basal motions independently, we found that apical motion was consistently faster than basal motion in a striking linear relation, and with higher speed for E-RG rosettes (Fig 5B). Higher apical (vs. basal) nuclear migration speeds were previously reported in time-lapse ex vivo cultured embryonic cortical slices [16] as well as in vivo in zebrafish retina and brain [33], providing further validation to our quantitative approach as an in vitro platform for investigating INM. Direct comparison further revealed an ordered relation for rosette speed as follows: apical speed of E-RG rosettes > basal speed of E-RG rosettes > apical speed of M-RG rosettes > basal speed of M-RG rosettes (Fig 5C). Inclusively, these data suggest that during early human cortical development, high NSC numbers are accompanied by elevated INM speed towards apical sites, similar to as shown for radial glial cells in cultured cortical slices ex vivo [16].
Our observations indicate that rosettes are generally characterized by basal motions that are slower but more radially organized, while apical motions are relatively faster yet less organized. Also, E-RG rosettes display elevated radial organization and higher speeds in general, for both basal and apical motions, compared to M-RG rosettes (Fig 5B and 5C). These two observations (fast & less organized versus fast & more organized) seem, at first, conflicting, but they are reconciled by the notion that E-RG rosettes exhibit high performance for both basal and apical motions to be faster (Fig 5C) as well as more organized (Fig 3B) than M-RG rosettes. This is while keeping in proportion the inherent hierarchy of basal motions that are generally slower and more organized, compared to apical motions, which are generally faster and less organized, in a rosette stage independent manner. Altogether, these findings also suggest that more than a single mechanism is involved in linking speed to radial organization of INM.
We noticed that all measures follow a spatial pattern. It is apparent that inner cells at the center of a rosette exhibit reduced organization, low B/A ratios and slower motion than cells located elsewhere (Fig 2). To understand the spatial dynamics distribution along the entire rosette area, we defined five circular rings with equal width (i.e., width is rosette size-dependent) for each rosette, starting from rosette center and outwards, and quantified measures for each ring. We found that luminal and peripheral rings exhibited reduced radial organization, B/A ratio and speed, while peak levels were recorded within intermediate rings (S5 Fig). This observation was more prominently expressed for E-RG rosettes. The poor performance of cells in distal rings of M-RG rosettes support a model where distal cells in M-RG rosettes “suffer” from mixed populations comprised of more differentiated cells that have detached the lumen (apical sites) towards periphery (Fig 1A, bottom panel; Ref. [20]) and thus should reduce organized motion.
The reduced dynamics and the broader cellular heterogeneity of M-RG rosettes suggest that these rosettes are mechanically compromised, reflecting the progressive disassembly of rosettes, which culminates around day 55 [20]. We therefore tested whether rosette disassembly is reflected functionally in INM measures also at the shorter range of time, i.e. throughout the time-lapse experiment. This was quantified by calculating each measure for each rosette at every time point independently and then recording its temporal variance (Methods). Importantly, no temporal trend was observed during the 4-hour imaging course of an experiment (S2E Fig). This implies that the variance encodes the fluctuations in a certain rosette measure over the imaging time course, a measure we term functional instability. Indeed, temporal variance of RS, B/A and speed was significantly higher for M-RG rosettes (Fig 6), suggesting functional instability as an indicator for the stage of progression in rosette disassembly. Importantly, these high temporal variances observed for M-RG rosettes measure their reduced ability in consistently performing INM, arguably due to their compromised structure. They do not measure the actual rosette disassembly process, which occurs in a longer time scale (from day 14 to day 35, and culminating towards day 55). Taken together, our results validate functional instability as a reliable readout for rosette organized dynamics.
Finally, to further strengthen the validity of our method and to shed some light on mechanisms of INM in vitro, we quantified the effects of pharmacological perturbation on INM. Different molecular motors are thought to mediate nuclei migration [12,34]. Such motors are believed to be a part of the cytoskeletal structural machinery such as actin, or motor proteins such as NMII, both shown to be involved in INM movements [35]. We treated rosettes with Blebbistatin or Cytochalasin-B, two agents known to alter INM by inhibiting NMII ATPase activity or depolymerizing actin, respectively. Quantifying INM dynamics in these rosettes following live imaging (S4–S6 Movies) showed decrease in INM measures (Fig 7A). This was further supported by demonstrating a loss of the ordered spatial composition of cell cycle components within rosettes. This was judged by immunostaining for mitosis (PHH3) and DNA synthesis (BrdU)–two distinct phases in the cell cycle that are spatially distributed to lumen and periphery, respectively (Fig 7B) (See also Ref. [20]). These findings further validate the ability of our quantitative approach to distinguish between rosettes under different molecular perturbations and further provide new evidence for possible roles of these molecular motors in driving INM in human radial glial cells.
Neural rosettes are highly organized multi-cellular structures that are formed by NSCs and are a cytoarchitectural hallmark during the transition of pluripotent stem cells into cortical fates in vitro. Here, we provided some initial insights and possible mechanisms for the complex dynamics of these structures, implicating intrinsic size-dependency of radial organization driven by mechanical constraints, inherently elevated radial organization for E-RG rosettes, and enrichment of basal motions as yet another potential contributor to enhanced and more stable INM dynamics of E-RG rosettes.
We propose three quantitative measures as means to quantify rosette dynamics: RS, B/A ratio and speed. These measures were used to assess differences in dynamics between early (E-RG) and late (M-RG) rosettes, revealing that INM of early rosettes is more efficient. This may well reflect the situation in the developing cortex: E-RG rosettes correspond to symmetrically dividing NSCs during early cortical development [20], which exhibit high self-replication rates and low levels of differentiation, resulting in increased number of cells undergoing INM within the ventricular zone. This implies that radial glial cells during early cortical development hold inherently elevated radial organization that may be required for accommodating the high traffic and orchestrating cell motion and cell division. In this regard, the radial expansion of the ventricular zone, which can be mirrored in vitro by the emergence of larger rosettes, may add greater mechanical constrains that ultimately contribute as well to enhance radial organization. At more advanced stages of cortical development, which are reflected by the M-RG stage in vitro [20], the production of neurons and intermediate progenitors—both non-polarized cells—is prevalent due to increase in asymmetric cell division of the corresponding radial glial cells. This occurs on the expanse of polarized NSCs adjacent to apical sites, which still perform INM. Thus, the accumulation of differentiated progeny increases non-NSC ratios, which in turn disrupt radial organization performance (Fig 8). To conclude, our analyses provide a first link between function and dynamics.
Literature survey demonstrated notable similarity between speeds measured across different labeling methods, in different species, in vivo and in vitro. Tsai et al (Ref. [16]) used cytoplasmic GFP for INM quantification of radial glial cells in mouse cortical slices ex vivo and shows higher speeds for apically directed motions (up to a 60μm/hour) compared to basally directed motions (up to 30 μm/hour). An in vivo study in zebra fish (Ref. [33]) shows comparable results using the more classic nuclei labeling, with speed of up to 20μm/hour and 3.4μm/hour for apical and basal motions, respectively. Similarly, although to a different extent, our in vitro findings show that human E-RG rosettes corresponding to early developing cortical radial glial cells exhibit faster speed for apical motions (38.8μm/hour) compared to speed of basal motions (35.3μm/hour), while the more advanced M-RG rosettes corresponding to mid neurogenesis moved apically at 30.3μm/hour and basally at 27.1μm/hour. This agreement is served as additional means to increase our confidence of using neural rosettes as physiologically relevant in vitro model system for cortical expansion as well as validating the capabilities of our analysis to quantify this process.
Our encouraging results suggest that this analytical framework may enable high-content quantification for diagnosis, molecular investigation and drug screening. Recent advancements in the stem cell field allow obtaining skin biopsies from patients and convert them into pluripotent stem cells (induced pluripotent stem cells (iPS cells)) [36]. Such iPS cells can be derived from patients with cortical diseases and then re-directed from the pluripotent stage into neural rosettes. We envisage that applying these quantitative measures on rosettes derived from patient iPS cells will have the potential to reflect damaged or genetic mutation-affected properties of NSCs. In addition, the effects of a specific drug / molecular perturbation could be predicted based on its alternation on these measures. Thus, this could be the first step toward developing platforms for understanding rosette dynamics in health and disease.
We used the human embryonic stem cell line WA-09 (H9) purchased from WiCell. Tel Aviv University Ethics Committee (IRB) approved the usage of existing human embryonic stem cell lines including H9. Material transfer agreement (MTS) was signed between Tel Aviv University (Vice President for R&D) and WiCell with regard to the transfer and usage of the human ES cell line WA09 in the Elkabetz lab, and following Agreement Letter between the Principal Investigator and WiCell. The use of human embryonic stem cells for therapeutic research is allowed in Israel. For more details see text on "The Use of Embryonic Stem Cells for Therapeutic Research", available at http://bioethics.academy.ac.il/english/report1/report1-e.html.
The human ES cell (hESC) line H9 (WA-09; Wicell—Wisconsin) and the H9-derived BAC transgenic HES5::eGFP line [22] were cultured on mitotically inactivated mouse embryonic fibroblasts (MEFs) (Globalstem). Undifferentiated hESCs were maintained as described previously [17] in medium containing DMEM/F12, 20% KSR, 1mM Glutamine, 1% Penicillin/Streptomycin, non-essential amino acids and beta-mercaptoethanol.
Experiments were performed as described in Refs. [20,21]. For neural induction and generation of neuroepithelial and radial glial cells, hESC colonies were removed from MEFs by Dispase (6U/ml, Worthington), dissociated with Accutase (Innovative Cell Technologies, Inc.), plated at sub confluent cell density (40-50K cells/cm2, although twice higher density or alternatively small hESC clusters work well and accelerate confluence) on Matrigel (1:20, BD) coated dishes, and supplemented with MEF-conditioned media and 10μM ROCK inhibitor (Y-27632, Tocris) with daily fresh FGF2 (10 ng/ml, R&D). Confluent cultures were subjected to dual SMAD inhibition neural differentiation using Noggin (R&D, 250 ng/ml) and SB-431542 (10 μM, Tocris), and further supplemented with LDN-193189 (100 nM, Stemgent) (denoted LNSB protocol). HES5::eGFP usually appears on day 8 or 9 of neural differentiation. To generate E-RG rosettes and subsequent progenitors, NE cells were scrapped from plates on day 10–12, pre-incubated with Ca+2/Mg+2 free HBSS followed by collagenase II (2.5 mg/ml), Collagenase IV (2.5 mg/ml) and DNAse (0.5 mg/ml) solution (all from Worthington) (37 degrees, 20 minutes). Cells were then dissociated and replated at high density (500,000 cells/cm2) on moist matrigel drops, and grown for additional days till rosettes appeared (E-RG stage). Neural induction and direct formation of E-RG stage rosettes could be also formed by co-culture of hESC clusters with MS5 stromal cells as previously described [17]. Briefly, early appearing rosettes on MS5 were harvested mechanically beginning on day 8–10 of differentiation, replated on culture dishes pre-coated with 15 μg/mL polyornithine (Sigma), 1 μg/mL Laminin (BD Biosciences) and 1 ug/ml Fibronectin (BD Biosciences) (Po/Lam/FN) till Day 14, to obtain E-RG rosettes. Under both protocols, early appearing NE cells were cultured from Day 9 with N2 medium (composed of DMEM/F12 and N2 supplement containing Insulin, Apo-transferin, Sodium Selenite, Putrecine and Progesterone), and further supplemented with low SHH (30ng/mL), FGF8 (100ng/mL) and BDNF (5ng/mL). Long-term culture of E-RG rosettes was performed by a weekly mechanical harvesting of rosettes and re-plating on Po/Lam/FN coated dishes with N2 medium, SHH and FGF8, till Day 28. These were replaced by FGF2 (20ng/mL) and EGF (20ng/mL) on Day 28 (all cytokines from R&D Systems). At day 35, E-RG rosettes reached the M-RG rosette stage. Cells were replated as clusters from one passage to another to reach the M-RG stage.
Cells were fixed in 4% paraformaldehyde, 0.15% picric acid, permeabilized and blocked with PBS, 1% FBS and 0.3% Triton solution, and stained with indicated primary antibodies followed by AlexaFluor secondary antibodies (Invitrogen). Cells were imaged in PBSx1 after staining. All cell imaging was carried out in 24 well glass bottom plates (In Vitro scientific). Fluorescence images were obtained using a Nikon Eclipse Ti-E microscope or a confocal LSM710 microscope (Carl Zeiss MicroImaging, Germany). The still or time-lapse images were captured using a 10x and a 20× objectives (NA = 0.3, 0.8 respectively, Plan-Apochromat).
Fluorescence emissions for eGFP, CY3, CY5 and DAPI channels were detected using filter sets supplied by the manufacturer.
For live imaging, cultured cells were maintained on the microscope stage in a temperature, CO2, and humidity-controlled environmental chamber. Time-lapse eGFP and phase matched contrast images were acquired using Nikon Eclipse Ti-E microscope every 5 minutes for over 4 hours (250 minutes). Physical pixel size was 0.64 x 0.64 μm. Images and movies were generated and analyzed using the NIS elements software (Nikon). All images were exported in TIF and then processed by our quantitative tools. 25 E-RG rosettes and 14 M-RG rosettes were live imaged in N = 2 independent experiments and quantified as detailed below.
E-RG rosette cells were treated with either Blebbistatin (5μM, Sigma) or Cytochalasin-B (0.5μg/ml, Sigma) and concomitantly recorded for 250 minutes as described above. 10 control E-RG rosettes, 8 E-RG rosettes treated with Blebbistatin and 14 E-RG rosettes treated with Cytochalasin-B were live imaged and analyzed.
Rosette centers and contours were manually annotated from the phase contrast channel, where the cytoarchitecture outlines of the region performing INM become obvious to a human eye based on different texture-patterns in the image. Rosette centers and peripheries were marked independently based on subjective identification of regions performing INM as reflected in the image-texture of the phase-contrast channel. Rosette contours were manually validated to remain stable throughout the time-lapse images (S1B Fig), in accordance with the different time scales of rosette imaging and rosette disassembly (S2E Fig). Independent annotations showed highly similar kinetic measures (S6 Fig). The expected radial angle was calculated in relation to the marked center for every location within the region-of-interest defined by the rosette contours. Rosette size was defined as the diameter of the circle that best fits the rosette-annotated contour. Local motion estimation was extracted by maximizing local cross correlation as described in Ref. [24]. Briefly, given two consecutive HES5::eGFP fluorescence images t, t+1 from the time-lapse sequence (i) Partition the current image (at time t) to a grid of sub-cellular sized local patches, of size 8.3μm x 8.3μm each (13 x 13 pixels); (ii) Find maximal cross-correlation to the next frame (t+1) to retrieve the local motion estimations for each patch. The search radius was defined based on maximal speed of 120μm / hour; (iii) Extract velocity angles and magnitude (speed) from the local motion estimation for each patch; (iv) Exclude motions below 15 μm hr-1 from all measures calculations. The quantized motion angles for each patch in the rosette were recorded for 250 minutes (50 frames). For each patch at every time point the following two angles were defined: Expected radial angle is the orientation of the vector between the rosette center and the patch at hand. The angular alignment γ of a given patch’s motion at a given time is the angle between the local velocity angle and its corresponding expected radial direction (Fig 2). A measure for radial organization was defined as the average angular alignment across all patches over time. This measure was termed Radial Score (RS), were high values reflected poor radial organization throughout a time-lapse experiment. Only patches that move at speed above 15 μm hour-1 (≥ 2 pixel per frame) were considered for calculating rosette radial organization, because small motions limit the discretization of the velocity angles which cause unreliable high angular deviations. The same minimal motion was considered for the rest of the analysis.
The velocity of each patch at every time was classified as apical or basal based on its direction in relation to the rosette center (Fig 2). Velocity angles pointing toward the rosettes periphery (+/- 90 degrees) were classified as basal, while toward the rosette center (+/- 90 degrees) as apical. The ratio of all basal to apical motions was termed basal-to-apical ratio (B/A ratio), where value of 1 reflects equal numbers of patches’ motions moving apically and basally, values > 1 corresponds to more motion toward the rosette periphery. RS for basal (correspondingly, apical) motions were calculated exactly as described above, only considering basal (apical) motions.
The measure Speed was defined as the average magnitude of velocity across all patches over time (Fig 2). Higher values reflect faster average motion throughout a time-lapse experiment. Basal or apical speed was calculated by considering the average of solely the basal or apical motions, respectively.
The variance for RS, B/A ratio and speed was calculated over time. Each of the measures was calculated independently for each frame in the time-lapse experiment (one scalar readout per frame), and the variance was recorded.
For spatial analysis, 5 different regions, at growing distances-intervals from the rosette center were defined for each rosette. These regions, termed circular rings, were each analyzed independently throughout the time-lapse experiment: all patches in each ring over time were used to calculate RS, B/A ratio and speed. Since rosette geometry was not a perfect circle, the last ring was not always complete, but confined by the rosette contour. Note that the width of a circular ring was rosette specific and changed as function of rosette size.
The nonparametric Wilcoxon rank sum test was used to assess statistical significance between E-RG and M-RG rosettes and for the perturbation experiments (Matlab function ranksum). Nonparametric Wilcoxon sign rank test was used to assess statistical significance between two measures calculated independently on the same rosette (e.g., apical vs. basal measures, Matlab function signrank). Pearson’s linear correlation was used to calculate associations and their corresponding p-values (Matlab function corr). Least square fit was applied to calculate the linear models that best fits the data (Matlab function polyfit). Box plots: solid black line inside the box is the median, bottom and top of the box are the 25% and 75% percentile, respectively.
|
10.1371/journal.pgen.1003588 | Alu Elements in ANRIL Non-Coding RNA at Chromosome 9p21 Modulate Atherogenic Cell Functions through Trans-Regulation of Gene Networks | The chromosome 9p21 (Chr9p21) locus of coronary artery disease has been identified in the first surge of genome-wide association and is the strongest genetic factor of atherosclerosis known today. Chr9p21 encodes the long non-coding RNA (ncRNA) antisense non-coding RNA in the INK4 locus (ANRIL). ANRIL expression is associated with the Chr9p21 genotype and correlated with atherosclerosis severity. Here, we report on the molecular mechanisms through which ANRIL regulates target-genes in trans, leading to increased cell proliferation, increased cell adhesion and decreased apoptosis, which are all essential mechanisms of atherogenesis. Importantly, trans-regulation was dependent on Alu motifs, which marked the promoters of ANRIL target genes and were mirrored in ANRIL RNA transcripts. ANRIL bound Polycomb group proteins that were highly enriched in the proximity of Alu motifs across the genome and were recruited to promoters of target genes upon ANRIL over-expression. The functional relevance of Alu motifs in ANRIL was confirmed by deletion and mutagenesis, reversing trans-regulation and atherogenic cell functions. ANRIL-regulated networks were confirmed in 2280 individuals with and without coronary artery disease and functionally validated in primary cells from patients carrying the Chr9p21 risk allele. Our study provides a molecular mechanism for pro-atherogenic effects of ANRIL at Chr9p21 and suggests a novel role for Alu elements in epigenetic gene regulation by long ncRNAs.
| Chromosome 9p21 is the strongest genetic factor for coronary artery disease and encodes the long non-coding RNA (ncRNA) ANRIL. Here, we show that increased ANRIL expression mediates atherosclerosis risk through trans-regulation of gene networks leading to pro-atherogenic cellular properties, such as increased proliferation and adhesion. ANRIL may act as a scaffold, guiding effector-proteins to chromatin. These functions depend on an Alu motif present in ANRIL RNA and mirrored several thousand-fold in the genome. Alu elements are a family of primate-specific short interspersed repeat elements (SINEs) and have been linked with genetic disease. Current models propose that either exonisation of Alu elements or changes of cis-regulation of adjacent genes are the underlying disease mechanisms. Our work extends the function of Alu transposons to regulatory components of long ncRNAs with a central role in epigenetic trans-regulation. Furthermore, it implies a pivotal role for Alu elements in genetically determined vascular disease and describes a plausible molecular mechanism for a pro-atherogenic function of ANRIL at chromosome 9p21.
| The chromosome 9p21 (Chr9p21) locus is the strongest genetic risk factor of atherosclerosis known today, yet, the responsible mechanisms still remain unclear. Chr9p21 lacks associations with common cardiovascular risk factors, such as lipids and hypertension, indicating that the locus exerts its effect through an alternative mechanism [1]–[4]. The risk region spans ∼50 kb of DNA sequence and does not encode protein-coding genes but the long non-coding RNA (ncRNA) antisense non-coding RNA in the INK4 locus (ANRIL; Figure 1A) [5], [6]. CDKN2BAS or CDKN2B-AS1 are used as synonyms for ANRIL. The closest neighbouring genes are the cyclin-dependent kinase inhibitors CDKN2A and CDKN2B, which are located ∼100 kb proximal of the Chr9p21 atherosclerosis risk region. While these genes are expressed in atherosclerotic lesions [7], the majority of studies in humans speak against a cis-regulation of CDKN2A and CDKN2B by Chr9p21 (reviewed by [8]). Studies in mice revealed no effect on atherosclerosis development [9], [10]. In contrast, a clear association of ANRIL with the Chr9p21 genotype has been established in several studies, even though the direction of effects is still a matter of dispute [5], [6], [8], [11]–[13]. Moreover, a correlation of ANRIL expression with atherosclerosis severity has been described [2], [8]. Based on these clinical and experimental data, ANRIL must be considered as a prime functional candidate for modifying atherosclerosis susceptibility at the Chr9p21 locus.
ANRIL belongs to the family of long ncRNAs, which are arbitrarily defined and distinguished from short ncRNA, such as microRNA, by their length of >200 bp [14]–[16]. Long ncRNAs have been implicated in diverse functions in gene regulation, such as chromosome dosage-compensation, imprinting, epigenetic regulation, cell cycle control, nuclear and cytoplasmic trafficking, transcription, translation, splicing and cell differentiation [15], [17]–[19]. These effects are mediated by RNA-RNA, RNA-DNA or RNA-protein interactions [17]–[19]. Previous mechanistic work on ANRIL in prostate tissue and cell lines has focused on its role in cis-suppression of CDKN2A and CDKN2B [3], [20]. Using RNAi against ANRIL, these studies showed impaired recruitment of chromobox homolog 7 (CBX7), a member of Polycomb repressive complex 1 (PRC1) [3], and of suppressor of zeste 12 (SUZ12), a member of PRC2 [20], to the Chr9p21 region. PRCs are multiprotein complexes, responsible for initiating and maintaining epigenetic chromatin modifications and thereby controlling gene expression [21]. Yap et al found that knock-down of ANRIL decreased trimethylation of lysine 27 residues in histone 3 (H3K27me3) and was associated with increased CDKN2A expression, while CDKN2B remained unchanged [3]. In contrast, Kotake et al showed that shRNA-mediated ANRIL knock-down disrupted SUZ12 binding to the Chr9p21 locus and led to increased CDKN2B expression whereas CDKN2A remained unaffected [20]. While results of ANRIL knock-down are conflicting with regard to expression of CDKN2A and CDKN2B, both studies demonstrated a significant reduction of cell proliferation [3], [20], a key mechanism in atherogenesis [22]. In these studies, however, potential effects of ANRIL knock-down on trans-regulation were not investigated.
Sato et al transiently over-expressed one specific ANRIL transcript in HeLa cells and found effects on expression levels of various genes in trans [23]. Even though the molecular mechanisms were not investigated in that work, this finding was of interest because trans-regulation of target genes has been proposed as a key mechanism for biological effects of other long non-coding RNA such as HOTAIR [24]–[26]. It is believed that these long ncRNAs mediate their effects through targeting epigenetic modifier proteins to specific sites in the genome [17], [19], [27]. Taken together, the previously available data suggested that ANRIL might influence gene expression by modulating chromatin modification and thereby affect cardiovascular risk.
The aim of the present study was to investigate the role of ANRIL in gene regulation and cellular functions related to atherogenesis on a mechanistic level. To this end, we performed genome-wide expression analyses in cell lines over-expressing distinct ANRIL transcripts that were associated with Chr9p21. We studied the molecular mechanisms of ANRIL-mediated gene regulation by investigating ANRIL binding to epigenetic effector proteins and their distribution across the genome. Using bioinformatics studies, we identified a regulatory motif characteristic for ANRIL-regulated genes. Finally, the functional relevance of the motif was confirmed by deletion and mutagenesis and results were validated in primary human cells from patients with and without the Chr9p21 atherosclerosis risk allele.
Using rapid amplification of cDNA ends (RACE) and subsequent PCR experiments, we identified four major groups of ANRIL transcripts in human peripheral blood mononuclear cells (PBMC) and the monocytic cell line MonoMac (Figure S1, Figure 1C). Consensus transcripts designated ANRIL1-4, comprising the most frequently occurring exon-combinations and most strongly expressed in MonoMac cells (Figure S1E), are shown in Figure 2A. Association of these transcripts with Chr9p21 was confirmed in PBMC (n = 2280) and whole blood (n = 960) of patients with and without coronary artery disease (CAD) in the Leipzig LIFE Heart Study [28] and in endarterectomy specimens (n = 193) (Figures 1D and 1E). The Chr9p21 risk allele was associated with increased ANRIL expression (Figure 1E) and different isoforms were positively correlated with each other (Figure S2). Using an assay detecting a common exon-exon boundary present in the majority of ANRIL isoforms (Ex1-5), we found a 26% overall increase of ANRIL expression per CAD-risk allele (P = 2.04×10−33) in PBMC of the Leipzig LIFE Heart Study. Strongest isoform specific effects were found for ANRIL2 (5% increase per risk allele; P = 0.002) and ANRIL4 (8% increase per risk allele; P = 3.02×10−6) (Figure 1E). To investigate the functional role of distinct ANRIL transcripts, we generated stably over-expressing cells lines (Figure 2B, Figure S3). ANRIL over-expression led to significant changes of gene expression in trans as determined by genome-wide mRNA expression analysis (Figure 2C). 219 transcripts were down- and 708 transcripts were up-regulated in ANRIL1-4 cell lines with average fold-changes smaller than 0.5 and greater than 2 compared to vector control, respectively (Table S1 and Table S2). These genes were distributed across the genome and there was no evidence for regulation of CDKN2A/B in the Chr9p21 region (Figure S4). Gene set enrichment analysis predicted an effect of ANRIL over-expression on movement/adhesion, growth/proliferation and cell death/apoptosis (Table 1), which are central mechanisms of atherogenesis [22]. We therefore aimed to experimentally validate these predictions in ANRIL over-expressing cell lines. Studies confirmed that ANRIL led to increased cell adhesion with strongest effects observed for cell lines over-expressing ANRIL4 (Figures 2D–2F). Over-expression of ANRIL further promoted cell growth and metabolic activity (Figures 2G–2I). Apoptosis, as determined by flow cytometric analysis of AnnexinV-positive cells (Figure 2J), caspase-3 activity (Figure 2K) and caspase-3 immunohistochemical staining (Figure 2L) was attenuated. Greatest biological effects were consistently found in cell lines over-expressing isoforms ANRIL2 and 4, the same isoforms, which also revealed the strongest associations with the Chr9p21 risk genotype (Figure 1E). Moreover, we demonstrated a dose-dependent effect on these mechanisms in independently established cell lines over-expressing these isoforms (Figure S5). Effects on cell adhesion, proliferation and apoptosis could be reversed by RNAi-mediated knock-down of ANRIL as shown in ANRIL2 and ANRIL4 cell lines (Figures 2M–2O, Figure S6), further supporting a pivotal role of ANRIL in these pro-atherogenic cellular functions.
To systematically identify ANRIL-associated epigenetic effector proteins, we next screened ANRIL binding to Polycomb group (PcG) proteins (AEBP2, BMI1, CBX7, EED, EZH2, JARID2, MEL18, PHF1, PHF19, RBAP46, RING1B, RYBP, SUZ12, YY1) and CoREST/REST (CoREST, REST, LSD1) using RNA immunoprecipitation (RIP) in nuclear extracts from ANRIL2 and 4 cells (Figures 3A–3B). ANRIL did not bind to CoREST/REST repressor proteins but bound to PRC1 proteins CBX7and RING1B and to PRC2 proteins EED, JARID2, RBAP46, and SUZ12. ANRIL also bound to PRC-associated proteins RYBP and YY1 which have been shown to induce gene expression [29], [30]. To investigate genome-wide distribution of Polycomb complexes, we chose CBX7 [3] and SUZ12 [20] as representative PcG proteins and performed chromatin immunoprecipitation followed by high-throughput sequencing (ChIP-seq) in ANRIL2 cells. Analysis of PcG distribution patterns in ANRIL-target gene promoters revealed reduced SUZ12 and CBX7 occupancy compared to not-regulated genes, following a wave-shaped binding pattern (Figure 3C and Figure S7). Reduced SUZ12 and CBX7 binding, as well as identical occupancy pattern, was replicated in publicly available data from BGO3 cells (Figure 3D). Over-expression of ANRIL increased SUZ12 and CBX7 binding to promoters of up-regulated genes (Figures 3E–3F), concordant with a recently described role of PRCs in regulation of active genes [31]. In further support, RNAi against CBX7 and SUZ12 largely reversed expression patterns not only of ANRIL-repressed, but also induced genes in ANRIL2 cells (Figure 3G). Whereas ANRIL binding to CBX7 and SUZ12 has previously been demonstrated in the context of cis-repression [3], [20], our experiments now extent the role of these proteins to ANRIL-mediated trans-regulation of gene networks pivotal in atherogenesis.
To address, which additional factors might be relevant for ANRIL-mediated trans-regulation, we performed bioinformatic analyses of promoter regions of ANRIL up- and down-regulated target genes. To this end, we used the MEME algorithm searching for differences in motif abundance and identified two partially overlapping DNA motifs (Figure 4A). Combined analysis of all trans-regulated genes validated the core motif CACGCCTGTAATCCCAGCACTTTGG (Figure 4A). The identified motif is an Alu-DEIN element [32], [33] with approximately 60.000 copies per human genome [34]. Since MEME does not provide information whether a motif is enriched or depleted, we quantitatively tested motif occurrence and found a significant depletion in up- and down-regulated genes compared to genes not regulated by ANRIL over-expression (4 vs. 6 occurrences per 5 kb promoter, respectively, P<10−15; Figure 4B). These data were consistent with decreased PcG occupancy in target-gene promoters (Figures 3C–3D). Notably, strong enrichment of PcG binding was found ∼150 bp downstream of the Alu motif compared to random DNA control (Figure 4C). This finding was replicated in the independent dataset for SUZ12 in BGO3 cells (Figure 4D) suggesting Alu element-dependent binding of PcG proteins as a general mechanism. Due to the repetitive nature of the investigated Alu motif, the significant spatial coherence between the motif and PcG occupancy was not detected when analyses were masked for repetitive elements (Figure 4E). Importantly, the same Alu motif which was found in the DNA promoter sequence of ANRIL-regulated genes was also present in ANRIL transcripts (Figure S8). Here, it was predicted to form a stem-loop structure in ANRIL RNA (Figure 4F) suggesting RNA-chromatin interactions as a potential effector mechanism [16]. Using RIP, we show that ANRIL co-immunoprecipitates with H3 and trimethylated lysine 27 of histone 3 (H3K27me3) (Figure 4G). These results speak further in favor of ANRIL-chromatin interaction at genomic sites where PRC-mediated epigenetic histone methylation takes place.
To validate the functional relevance of Alu sequences implemented in ANRIL transcripts, we generated stably over-expressing cell lines devoid of exons containing these sequences (Figure 5A, Figure S9). Over-expression of mutant isoforms ANRIL2a-c and ANRIL4a,b reversed expression changes of representative transcripts that were otherwise induced (TSC22D3; Figure 2C and Figure 5B) or suppressed (COL3A1; Figure 2C and Figure 5C) in ANRIL2 and 4 cells. Increased cell adhesion in ANRIL2 and 4 was abolished in cell lines lacking the Alu motif (Figure 5D). Consistent with this finding, ANRIL2a-c and ANRIL4a,b cell lines showed increased apoptosis and proliferated more slowly compared to ANRIL2 and 4 (Figures 5E–5F). To exclude that depletion of whole exons led to significant changes in ANRIL secondary structure and thus impaired ANRIL function, we next generated cell lines stably over-expressing ANRIL isoforms with single-base mutations of 25%, 33%, and 100% of nucleotides in the identified Alu motif (Figure 5G, Figure S9). Over-expression of these ANRIL isoforms confirmed the important role of the Alu motif by reversing expression changes of ANRIL trans-regulated genes (TSC22D3, Figure 2C and Figure 5H; COL3A1, Figure 2C and Figure 5I) compared to ANRIL2 cells. Moreover, ANRIL-mediated effects on adhesion, apoptosis and proliferation were attenuated (Figures 5J–5L), further supporting the functional relevance of ANRIL Alu motifs in pro-atherogenic cellular functions.
To investigate whether findings from cell culture studies could be translated into the human situation, we investigated genome-wide transcript expression in PBMC from 2280 subjects of the Leipzig LIFE Heart Study and associated gene expression with the Chr9p21 haplotype and ANRIL expression. While no transcripts were significantly associated with Chr9p21 on a genome-wide level, gene set enrichment analyses of genes correlated with ANRIL expression (n = 5066; P value≤0.01) and associated at a nominal significance level with Chr9p21 (n = 1698; P-value≤0.05) revealed comparable pathways to those identified in ANRIL over-expressing cell lines (Table 2). We next investigated adhesion and apoptosis in PBMC from patients, which were either homozygous for the protective or the CAD-risk allele at Chr9p21 (Figures 6A–6B). Consistent with our earlier cell culture studies, the risk allele, which was associated with increased expression of linear ANRIL transcripts (Figure 1), led to increased adhesion (P = 0.001; Figure 6A) and decreased apoptosis (P = 0.001; Figure 6B) compared to PBMCs from carriers of the protective allele.
Here we show that the same ANRIL isoforms, which were up-regulated in patients carrying the Chr9p21 atherosclerosis risk haplotype, modulate gene networks in trans leading to pro-atherogenic cellular properties. At the molecular level, we provide strong evidence for an Alu sequence (Figure 4A) as the key regulatory element responsible for ANRIL-mediated trans-regulation (Figure 6C). This Alu motif was not only expressed in ANRIL RNA transcripts but also marked promoters of target-genes and was associated with epigenetic effector protein binding. Depletion and mutagenesis of the motif reversed trans-regulation and normalized cellular functions. More generally, the proposed mechanism of ANRIL in atherogenesis highlights a novel role of Alu elements in epigenetic trans-regulation of gene networks, which might be relevant for other long ncRNAs as well.
To the best of our knowledge, this is the first study investigating the role of distinct ANRIL isoforms in several key mechanisms of atherogenesis using stable over-expression and knock-down approaches (Figure 1 and Figure 2). All investigated ANRIL isoforms had more or less pronounced effects on cellular functions (Figure 2, Table S3), but strongest effects were consistently found in cell lines over-expressing ANRIL isoforms which were up-regulated in patients carrying the Chr9p21 risk allele (Figure 1). These effects were also dose-dependent (Figure S5). Importantly, results from cell culture studies were validated in primary cells from patients with and without the Chr9p21 risk haplotype (Figure 6). So far, mechanistic work on ANRIL has focused on cell proliferation using RNAi approaches [3], [20], [35], [36]. In these studies down-regulation of ANRIL led to decreased proliferation in different cell culture models, which is well in line with our observation of increased proliferation in stable ANRIL over-expressing cell lines (Figure 2). Network analyses of trans-regulated genes in the present study indicated that in addition to proliferation, cell adhesion and apoptosis were also affected in ANRIL cell lines. These predictions were functionally validated revealing that ANRIL over-expression led to increased adhesion and decreased apoptosis. Again, effects were greatest for those isoforms showing the strongest associations with the Chr9p21 risk genotype and could be reversed by RNAi against ANRIL (Figure 2). Moreover, the direction of atherogenic cell functions (increased cell adhesion, increased proliferation and decreased apoptosis) nicely fits with evidence for a potential pro-atherogenic role of ANRIL from patient studies [6], [22]. Taken together, our data provide a plausible mechanism for pro-atherogenic functions of Chr9p21-associated ANRIL transcripts.
In previous studies, it has been shown that long ncRNA may affect gene expression of target genes [24]–[26], [37]–[39]. Among the best studied examples is the long ncRNA HOTAIR, which is transcribed from the HOXC cluster and was shown to repress genes in the HOXD cluster in trans [25], [26]. A potential role of ANRIL in trans-regulation has previously been postulated by Sato et al, investigating genome-wide mRNA expression in HeLa cells upon transient over-expression of a single ANRIL isoform [23]. In the current work, we used a model of stable over-expressing cell lines and found evidence for trans-regulation of distinct gene networks that overlapped between different ANRIL isoforms (Table 1). To translate these finding to patients with coronary artery disease, we performed pathway analysis of genes that were correlated with ANRIL expression and associated with the Chr9p21 risk genotype in 2280 participants of the Leipzig LIFE Heart Study. While no transcript was associated with Chr9p21 at a genome-wide level of significance, we found almost identical pathways in patients with high ANRIL expression and high genetic risk at Chr9p21 as in stable ANRIL over-expressing cell lines (Table 2). The lack of significant associations of gene expression with Chr9p21 has also been observed by Zeller et al [40] and might be explained by the rather subtle modulation of ANRIL by Chr9p21 with slighter effects on trans-regulation as opposed to stronger over-expression in the investigated cell culture models. Nevertheless, permanently elevated ANRIL levels might activate the observed gene networks leading to subtle changes of cellular functions (Figure 6) and thereby increasing cardiovascular risk over time.
On the mechanistic level, it has been postulated that long ncRNA may serve as a scaffold, guiding effector-proteins to chromatin [24]–[26], [37]–[39]. Indeed, previous work has demonstrated ANRIL binding to CBX7 [3] and SUZ12 [20], which are proteins contained in PRC1 and PRC2, respectively. Both previous papers focused on ANRIL-mediated cis-repression of CDKN2A and CDKN2B using RNAi but did not investigate potential trans-regulatory effects [3], [20]. In the current work, we found no effect of ANRIL over-expression on expression of CDKN2A and CDKN2B (Figure S4). Thus, the spatial coherence of ANRIL transcription with adjacent protein-coding genes might be relevant for cis-suppression.
This is the first study systematically investigating ANRIL binding to 17 different proteins contained in chromatin modifying complexes PRC1, PRC2 and CoREST/Rest (Figure 3). We demonstrated binding to predominantly inhibitory (CBX7, RING1B, EED, JARID2, RBAP46, SUZ12) [41] and potentially activating factors (RYBP, YY1) [29], [30]. CBX7 and SUZ12 were selected as representative proteins for PRC1 and 2, respectively, and were followed up in genome-wide ChIP-seq experiments (Figure 3 and Figure 4). Notably, ANRIL over-expression was accompanied with changes of CBX7 and SUZ12 distribution in ANRIL target-gene promoters (Figure 3). The role of PRCs in ANRIL-mediated gene regulation was further proven by RNAi against CBX7 and SUZ12 which reversed trans-regulation of down- and also of up-regulated target genes in ANRIL cells (Figure 3G). Induction of gene expression through PcG proteins might seem at odds with the current understanding of PRC-mediated gene silencing. However, binding of PcG proteins in active genes has been described in earlier studies [31]. Moreover, Morey et al demonstrated reduced gene repression in distinct PRC1 compexes containing RYBP, which was shown to associate with ANRIL (Figure 3) [42]. Alternatively, ANRIL-bound activating factors RYBP and YY1 may directly activate gene expression [29], [30] or ANRIL might bind other activating epigenetic effector proteins [43]. Taken together, we demonstrate a central role of PRC1 and PRC2 but not CoREST/Rest in ANRIL-mediated repression and induction of genes but additional work is clearly warranted to unravel the complex nature of these trans-regulatory mechanisms.
A key limitation in the current understanding of long ncRNA function, in particular with regard to their trans-regulatory effects, is the lack of knowledge about specific regulatory sequences or structural motifs. These sequences might be responsible for targeting long ncRNA to distinct regulatory sites in the genome. Here, we provide first evidence for an important role of Alu motifs in this process, which mark the promoters of ANRIL trans-regulated genes (Figure 4). Additionally, genome-wide ChIP-seq analysis revealed binding of ANRIL-associated PcG effector proteins in close proximity to the Alu motif. Importantly, the same Alu motif was included in ANRIL ncRNA transcripts and predicted to be located in a central stem-loop like structure (Figure 4F). Additional evidence for ANRIL binding to chromatin comes from its association with histone H3 and H3K27me3 (Figure 4G). In previous work, Alu-like, CG- and GA-rich sequences were proposed as potential RNA-DNA interaction sites associated with RNA:DNA:DNA triplex formation [44], [45]. Paradoxically, Alu occurrence was significantly depleted in ANRIL-regulated genes, suggesting that a certain spatial patterning or additional co-factors in promoter regions rather than motif abundance alone might be relevant (Figure 4, Figure 6C). Functional relevance of the motif was proven by depletion and mutagenesis in ANRIL RNA, which reversed trans-regulation and pro-atherogenic cellular properties (Figure 5). Thus, our data suggest that ANRIL may bind to chromatin through interaction via its Alu motif, thereby guiding PRC proteins to ANRIL-regulated genes (Figure 6C).
Despite long consideration as “genomic junk”, Alu elements are coming into the focus of intense research. Alu enrichment occurred over the course of evolution [46] and integration at genomic sites was associated with maturation and gain-of-function of ncRNAs [47]. Alu elements as well as distal ANRIL exons are not conserved in the orthologous chromosomal region on mouse chromosome 4, which might be an explanation for the lack of effect on atherosclerosis when deleting that region [48]. Recent work by Jeck et al has also demonstrated preferred inclusion of Alu motifs in non-coding RNA lariats which are commonly thought to represent inactive forms of ncRNAs [49]. Whether implementation of Alu motifs in ncRNA lariats leads to silencing of the effector sequences or not remains to be determined.
In summary, our work extends the function of Alu motifs to regulatory components of ncRNAs with a central role in ncRNA-mediated epigenetic trans-regulation. Furthermore, it implies a pivotal role for Alu elements in genetically determined vascular disease and describes a plausible molecular mechanism for pro-atherogenic ANRIL function at Chr9p21.
The Leipzig LIFE Heart Study has been approved by the Ethics Committee of the Medical Faculty of the University Leipzig (Reg. No 276-2005) and was described previously [28]. RNA from peripheral blood mononuclear cells (PBMC; n = 2280) and whole blood (n = 960) from the same patients of this study was isolated as described [6]. Human endarteryectomy specimens (n = 193) were collected in an independent cohort of patients undergoing vascular surgery and the utilization of human vascular tissues was approved by the Ethics Committee of the Medical Faculty Carl Gustav Carus of the Technical University Dresden (EK316122008) [7].
Genotyping of single nucleotide polymorphisms rs10757274, rs2383206, rs2383297, and rs10757278 in 2280 probands of the Leipzig LIFE Heart Study and in vascular tissue was performed as described [6], [7].
Initial screening of human cell lines MonoMac, THP1, U937, HEK293, HepG2, CaCo2, and Hutu80 for Chr9p21 gene expression (MTAP, CDKN2A, CDKN2B, ANRIL, qRT-PCR assays were described in [6]) revealed that MonoMac and HEK293 cells expressed all Chr9p21 transcripts (data not shown). This suggested that these genes might be functionally relevant in these cells whereas all other investigated cell lines were lacking expression of at least one of the Chr9p21 transcripts. HEK293 cells (DMSZ, ACC305) were cultured in DMEM (Life Technologies) containing 10% fetal calf serum (FCS, Biochrom), 1% penicillin/streptomycin (P/S, Life Technologies). MonoMac cells (DSMZ, ACC124) were cultured in RPMI 1640 (Biochrom) containing 10% FCS, 1% P/S, 1% MEM (Life Technologies), and 1% OPI (Sigma). Cryopreserved PBMC (n = 32) were thawed, cultured in RPMI 1640 (Biochrom) containing 10% FCS, 1% P/S and functional assays were performed within 48 hours.
ANRIL isoforms were cloned in the bicistronic pBI-CMV2 (Clontech) or pTRACER-SV40 (Life Technologies) vectors allowing parallel expression of a green fluorescent protein (GFP) and ANRIL transcripts. Using Lipofectamine 2000 (Life Technologies), HEK293 cells were either co-transfected with ANRIL-pBI-CMV2 vectors/empty vector control and neomycin-encoding vector or transfected with ANRIL-pTRACER vectors/empty vector control encoding a GFP-Zeocin resistance gene. Transfected cells were selected with geneticindisulfate (G418, Roth) or Zeocin (Life Technologies). After 2 weeks, stably transfected cells were selected by flow cytometry and over-expression of ANRIL was validated by quantitative RT-PCR. On average, 2–4 cell lines were generated per ANRIL isoform and vector control, respectively.
Cell adhesion assays were performed in 96-well plates coated with collagen (Roche), Matrigel (BD) or PBS. Cells were allowed to adhere for 40 min, quadruplicate measurements/cell line were performed. Numbers of adherent cells were determined using CellTiter-Blue/CellTiter-Glo (Promega) in relation to standard curves of the respective cell line. PBMC adhesion assays were performed accordingly. Quadruplicate measurements/subject were performed. Cellular proliferation was either determined by counting absolute cell numbers (trypan blue staining) or viability assays (CellTiter-Blue/CellTiter-Glo, Promega). Glucose in the cell culture supernatants was determined using standard chemistry (Roche). Apoptosis was determined using flow-cytometric detection of AnnexinV positive cells (GFP-Certified Apoptosis/Necrosis Detection Kit, ENZO), caspase-3 activity (Caspase-3/CPP32 Fluorometric Assay Kit, BioVision; CaspaseGlo 3/7, Promega) and caspase-3 staining.
siRNA knock-down of ANRIL (n272158, Life Technologies), CBX7 (s23926, Life Technologies), SUZ12 (s23967, Life Technologies) was performed using Lipofectamine 2000 (Life Technologies). Knock-down efficiency was determined by quantitative RT-PCR and Western blotting.
10 µg RNA from human PBMC and from MonoMac cells were reverse transcribed using the ExactSTART Eukaryotic mRNA 5′- & 3′-RACE Kit (Epicentre Biotechnologies) according to the manufacturer's instructions. RACE PCR reactions were prepared using Advantage 2 Polymerase Mix (Clontech) and subcloned (pCR2.1-TOPO Vector, Life Technologies). Primers used for RACE experiments are listed in Table S4. Amplification of full-length isoforms was performed using primers listed in Table S5 (schematic in Figure S1). Sequencing of PCR products was performed with an automated DNA sequencer (Applied Biosystems).
Quantitative RT-PCRs and analysis of data were performed as described [6]. Primers and probes for ANRIL isoforms (n = 5), TSC22D3, COL3A1, and U1 are given in Table S6.
RNA from cell lines ANRIL1-4 and vector control cell line, siRNA-treated ANRIL2 cells, and RNA from PBMC (n = 2280) was labeled and hybridized to Illumina HumanHT-12 v4 BeadChips. Arrays were scanned with an Illumina iScan microarray scanner. Bead level data preprocessing was done in Illumina GenomeStudio.
Antibodies against AEBP2 (11232-2-AP, Proteintech), BMI1 (05-637, Millipore), CBX7 (ab21873, Abcam), EED (03-196, Millipore), EZH2 (ACC-3147, Cell Signaling), JARID2 (NB100-2214, Novus Biologicals), MEL18 (ab5267, Abcam), PHF1 (sc-101107, Santa Cruz), PHF19 (11895-1-AP, Proteintech), RBAP46 (MA1-23277, Thermo Scientific), RING1B (NBP1-49966, Novus Biologicals), RYBP (NBP1-97742, Novus Biologicals), SUZ12 (ab12073, Abcam), YY1 (NBP1-46218, Novus Biologicals), LSD1 (39186, Active Motif), REST (07-579, Millipore), CoREST (07-455, Millipore), H3 (ab1791, Abcam), H3K27me3 (ab6002, Abcam), rabbit control IgG (kch-504-250, Diagenode), mouse control IgG (kch-819-015, Diagenode), and goat control IgG (sc-2346, Santa Cruz) were used.
The immunoprecipitation reaction followed with some modifications previously described protocols [50], [51]: 2×107 ANRIL2, ANRIL4 and vector control cells were treated with 0.1% formaldehyde for 10 min at 23°C. Nuclear extracts were preincubated with non-immune IgG and tRNA (100 µg/ml) for 1 h and primary antibodies and control IgG reacted overnight at 4°C. The immunoprecipitates were washed 3× with low-salt, 3× with high-salt and 1× Li-salt buffers. The retrieved RNA was quantitated, reverse transcribed using random hexamer primers, and analyzed by qRT-PCR with ANRIL-specific primers (Ex1-5 assay, Table S6). U1-specific primers were used as negative control (Table S6).
2–5×107 ANRIL2 and vector control cells were fixed with 1% formaldehyde for 10 min at 23°C. Cross-linking reaction was stopped by adding glycine to a final concentration of 0.125 M for 5 min followed by three washes with ice-cold PBS. Cells were harvested, suspended in 50 mM Hepes-KOH, pH 7.5, 100 mM NaCl, 1 mM EDTA, pH 8.0, 0.5 mM EGTA, pH 8.0 and incubated for 10 min at 4°C. Cells were washed with 10 mM Tris-HCl, pH 8.0, 200 mM NaCl, 1 mM EDTA, pH 8.0, 0.5 mM EGTA, pH 8.0 for 10 min, pelleted by centrifugation and the nuclei lysed in 10 mM Tris-HCl, pH 8.0, 200 mM NaCl, 1 mM EDTA, pH 8.0, 0.5 mM EGTA pH 8.0, 0.1% Na-Deoxycholate, 0.5% N-lauroylsarcosine for 10 min at 4°C. Samples were sonicated with Diagenode Bioruptor UCD-300TO (High level, 8×5 cycles 30 sec on/30 sec off, 4°C), centrifuged at 15,000× g for 10 min at 4°C and the supernatant was shock-frozen in liquid nitrogen and stored at −80°C. Nuclear extracts were preadsorbed with non-immune IgG for 1 h and treated with SUZ12, CBX7, and control IgG antibodies overnight at 4°C. Precipitates were processed with the HighCell#ChIP kit (Diagenode) and Illumina DNA sequencing libraries were generated using the ChIP-seq Sample Preparation Kit (IP-102-1001, Illumina). Purity and quantity was measured on an Agilent Technologies 2100 Bioanalyzer. Sequencing was performed with the Illumina Genome Analyzer II platform.
ChIP-seq data are deposited in the Sequence Read Archive (SRA) under accession number SRA052089.1. BGO3 (GSM602674) data are publicly available at (http://www.ncbi.nlm.nih.gov/geo).
|
10.1371/journal.pntd.0007620 | Informal genomic surveillance of regional distribution of Salmonella Typhi genotypes and antimicrobial resistance via returning travellers | Salmonella enterica serovar Typhi (S. Typhi) is the causative agent of typhoid fever, a systemic human infection with a burden exceeding 20 million cases each year that occurs disproportionately among children in low and middle income countries. Antimicrobial therapy is the mainstay for treatment, but resistance to multiple agents is common. Here we report genotypes and antimicrobial resistance (AMR) determinants detected from routine whole-genome sequencing (WGS) of 533 S. Typhi isolates referred to Public Health England between April 2014 and March 2017, 488 (92%) of which had accompanying patient travel information obtained via an enhanced surveillance questionnaire. The majority of cases involved S. Typhi 4.3.1 (H58) linked with travel to South Asia (59%). Travel to East and West Africa were associated with genotypes 4.3.1 and 3.3.1, respectively. Point mutations in the quinolone resistance determining region (QRDR), associated with reduced susceptibility to fluoroquinolones, were very common (85% of all cases) but the frequency varied significantly by region of travel: 95% in South Asia, 43% in East Africa, 27% in West Africa. QRDR triple mutants, resistant to ciprofloxacin, were restricted to 4.3.1 lineage II and associated with travel to India, accounting for 23% of cases reporting travel to the country. Overall 24% of isolates were MDR, however the frequency varied significantly by region and country of travel: 27% in West Africa, 52% in East Africa, 55% in Pakistan, 24% in Bangladesh, 3% in India. MDR determinants were plasmid-borne (IncHI1 PST2 plasmids) in S. Typhi 3.1.1 linked to West Africa, but in all other regions MDR was chromosomally integrated in 4.3.1 lineage I. We propose that routine WGS data from travel-associated cases in industrialised countries could serve as informal sentinel AMR genomic surveillance data for countries where WGS is not available or routinely performed.
| Our data demonstrate how routine WGS data produced by Public Health England can be further mined for informal passive surveillance of Salmonella Typhi circulating in different geographical regions where typhoid is endemic. We have shown the public health utility of a simplified approach to WGS reporting based on the GenoTyphi genotyping framework and nomenclature, which doesn’t require the generation of a phylogenetic tree or other phylogenetic analysis. These approaches yielded results consistent with previously reported antimicrobial resistance (AMR) patterns of S. Typhi, including prevalence of multi-drug resistant (MDR) and fluoroquinolone resistance in different regions in association with different pathogen variants. These data provide a rationale and framework for the extraction and reporting of geographically stratified genotype and AMR data from public health labs in non-endemic countries. Prospective analysis and reporting of such data could potentially detect shifts in regional S. Typhi populations, such as replacement or spread of different subclades and the emergence and dissemination of MDR, fluoroquinolone resistant and/or extensively drug resistant S. Typhi, providing valuable data to inform typhoid control measures in low and middle income countries that are still building their genomics capacity.
| Salmonella enterica serovar Typhi (S. Typhi) is the causative agent of typhoid fever[1], a systemic human infection responsible for an estimated 223,000 deaths each year resulting from more than 20 million infections[2,3]. The majority of the disease burden falls on children in low and middle income countries (LMICs)[2], however vaccination programmes are rare in endemic countries and antimicrobial therapy is considered crucial for the safe clearance of S. Typhi infections and the avoidance of clinical complications. Historically typhoid fever could be effectively treated using first-line drugs including chloramphenicol, ampillicin or co-trimoxazole. However the emergence of multi-drug resistant (MDR) S. Typhi, defined as displaying resistance to all three drugs, in the late 1980s and early 1990s resulted in changes to treatment guidelines, with fluoroquinolones becoming the recommended therapy[4–7]. Recent years have seen a rise in the proportion of S. Typhi disease isolates displaying reduced susceptibility to fluoroquinolones associated with point mutations in quinolone resistance determining region (QRDR) of gyrA and parC[2,8,9].
S. Typhi is genetically monomorphic, which has historically constrained molecular surveillance of S. Typhi before the era of high throughput whole genome sequencing (WGS)[5,10]. A phylogenetically informative genotyping scheme, GenoTyphi, was recently introduced to facilitate the interpretation of S. Typhi WGS data[11]. Application of the scheme to a global collection of isolates from >60 countries showed that the S. Typhi population is highly structured, comprising dozens of subclades associated with specific geographical regions[11,12]. This global genomic framework for S. Typhi revealed the majority of MDR S. Typhi infections worldwide are associated with a single phylogenetic lineage, designated as genotype 4.3.1 (H58 under the legacy scheme[5]), which has disseminated from South Asia since the 1990s, including into East Africa [11,12]. Contrastingly, in West Africa MDR S. Typhi is associated with a different genotype, 3.1.1[13,14]. The MDR phenotype in both S. Typhi 4.3.1 and 3.1.1 is encoded by a composite transposon carrying genes conferring resistance to five drug classes. These include chloramphenicol, penicillins and co-trimoxazole and the determinants are typically located on IncH1 plasmids (plasmid sequence type PST6 in 4.3.1, and PST2 in 3.1.1) [4,12–17]. Recently, migration of the MDR transposon to the S. Typhi chromosome via IS1 transposition has been reported, with four different integration sites identified in the S. Typhi 4.3.1 strain background (fbp, yidA, STY4438 and the intergenic region between cyaA and cyaY)[12,18–21].
Resistance to newer agents is also on the rise in S. Typhi. Reduced susceptibility to fluoroquinolones via QRDR mutations is commonly observed amongst S. Typhi 4.3.1[12], especially lineage II[16,22], but is rare in 3.1.1 and other genotypes[12,13]. S. Typhi 4.3.1 QRDR triple mutants (carrying two mutations in gyrA and one in parC) displaying complete resistance to ciprofloxacin are increasing in prevalence in South Asia[23]. WGS analysis showed that recent treatment failure with gatifloxacin in Nepal resulted from introduction of a 4.3.1 lineage II triple mutant from India, prompting a reconsideration of the reliance on fluoroquinolones in the region[8,22]. Subsequently an outbreak of S. Typhi 4.3.1 displaying resistance to ceftriaxone in addition to ciprofloxacin and all three first line drugs was reported in Pakistan[24,25]. The outbreak strain carried both the MDR composite transposon integrated in the chromosome at yidA, and an E. coli IncY plasmid harbouring the extended spectrum beta-lactamase (ESBL) gene blaCTX-M-15 (conferring resistance to ceftriaxone and other third generation cephalosporins) and the quinolone resistance gene qnrS (which combined with a gyrA-S83F mutation in the chromosome conferred resistance to ciprofloxacin)[24]. This strain has been designated XDR (extensively drug resistant, defined as resistance to chloramphenicol, penicillins, co-trimoxazole, ceftriaxone and ciprofloxacin) and severely limits treatment options, with azithromycin being the last remaining oral antibiotic (to which sporadic resistance has already been observed in South Asia)[7].
These dynamic trends highlight a need for prospective AMR surveillance in global S. Typhi populations, in order to inform empirical treatment options; genomic surveillance offers the added benefit of revealing resistance mechanisms and regional and international spread of emerging MDR and XDR strains[26]. Blood culture confirmation and isolate characterisation is a pre-requisite for such activities, but neither are routinely performed in laboratories in areas where typhoid fever is endemic. However, in many developed nations, S. Typhi infections are notifiable and the disease is primarily associated with travellers returning from high-risk regions[27], providing an opportunity for informal sentinel surveillance of those regions. In England, typhoid fever is notifiable and all S. Typhi isolates are sent to Public Health England (PHE) for confirmation and characterisation via WGS, and recent travel history is sought via enhanced surveillance questionnaires. Most cases identified in England (~80%) are associated with recent travel to typhoid endemic areas[27]. We recently reported that the geographic origin of travel-associated S. Typhi cases in London could be predicted by comparing WGS data to the global framework[11], and identified the first case of ESBL S. Typhi in England from a patient with travel to Pakistan[28], supporting the notion that WGS data on travel-associated cases has sentinel surveillance value. Here we analysed WGS data from S. Typhi isolates referred to PHE from three years of national surveillance between April 2014 and March 2017[29], and explored the distribution of lineages and AMR determinants in geographic regions frequented by UK travellers.
A total of 533 S. Typhi isolates from English cases received by PHE during the period of 1st April 2014 to 31st March 2017 were included in this study. Ethical approval for the detection of gastrointestinal bacterial pathogens from faecal specimens, or the identification, characteristation and typing of cultures of gastrointestinal pathogens, submitted to the Gastrointestinal Bacteria Reference Unit is not required as covered by Public Health England’s surveillance mandate. Patient travel information was available for 488/533 of the isolates and was obtained by PHE using an enhanced surveillance questionnaire (https://www.gov.uk/government/publications/typhoid-and-paratyphoid-enhanced-surveillance-questionnaire), this included questions pertaining to the destination of any foreign travel that occurred during the likely incubation period (28 days before onset of symptoms). No specific consent was required from the patients whose data were used in this analysis because Public Health England has authority to handle patient data for public health monitoring and infection control under section 251 of the UK National Health Service Act of 2006. For the remaining 45 isolates, no enhanced surveillance questionnaire was completed or the data collected was incomplete; these were spread across the four years (n = 15 for 2014, n = 24 for 2015, n = 4 for 2016 and n = 2 for 2017; mean 8.5%). Details of all 533 isolates are provided in S1 Table.
WGS was conducted as part of routine sequencing of all Salmonella isolates referred to the Gastrointestinal Bacteria Reference Unit at PHE, as previously described[29]. Briefly, DNA was fragmented and tagged for multiplexing with NexteraXT DNA Sample Preparation Kits (Illumina) and sequenced at PHE on a HiSeq 2500 yielding 100 bp paired end reads. FASTQ data is available from the NCBI Short Read Archive, BioProject accession PRJNA248792. Individual accession numbers for isolates analysed in this study are given in S1 Table.
Paired end Illumina reads were mapped to the CT18 reference genome (accession AL513382)[30], which is the standard reference for S. Typhi genomic studies, using RedDog (V1.beta10.3) available at (https://github.com/katholt/RedDog). Briefly, RedDog maps reads to the reference genome using Bowtie2 (v.2.2.9)[31], before using SAMtools (v1.3.1)[32] to identify high quality single nucleotide variant (SNV) calls as previously described[19]. A core SNV alignment was generated by concatenating alleles with high-quality consensus base calls (phred score >20), for all SNV sites that had such calls in >95% of genomes (this represents the 95% ‘soft’ core of the S. Typhi genomes); this alignment was filtered to exclude SNVs in phage regions and repetitive sequences (354 kb; ~7.4% of bases) in the CT18 reference chromosome as defined previously[10]; S2 Table) and recombinant regions identified by Gubbins[33]. The resulting alignment of 8053 SNVs was used as input to RAxML (v8.2.8) to infer a maximum likelihood (ML) phylogeny with a generalised time-reversible model and a Gamma distribution to model site-specific rate variation (GTR+ Γ substitution model; GTRGAMMA in RAxML), and 100 bootstrap pseudo-replicates to assess branch support.
Genotypes were inferred for all isolates by screening the Bowtie2 alignment (bam) files for SNVs used in the extended S. Typhi typing framework, GenoTyphi (code available at https://github.com/katholt/genotyphi)[11]. GenoTyphi uses 72 specific SNVs to assign isolates to one of four primary clusters; 16 clades; and 49 subclades[11], with the globally disseminated 4.3.1 (H58) subclade further delineated into lineages I and II (4.3.1.1 and 4.3.1.2).
S. Typhi genomes were screened for acquired AMR genes and plasmid replicons using PHE’s mapping-based Genefinder pipeline as previously described[34], which has been extensively validated in comparing phenotypic and WGS-derived AMR profiles on ~600 typhoidal Salmonella Typhi and Paratyphi [35]. Reference sequences in Genefinder were curated from those described in Comprehensive Antimicrobial Resistance Database[36] and PlasmidFinder[37]. Genes were called as present within a genome when detected with 100% coverage and >90% nucleotide identity to the reference gene. Genefinder was also used to detect point mutations in the QRDR of chromosomal genes gyrA (codons 83, 87) and parC (codons 79, 80, 84)[35,38]. Isolates were defined as being MDR if genes were detected by Genefinder in the Beta-Lactamases, Trimethoprim, Sulphonamides and Chloramphenicol classes.
ISmapper[39] (v2) was run with default parameters to screen all read sets for insertion sites of the transposase IS1 (accession X52534) relative to the CT18 reference chromosome sequence (accession AL513382). The binary data from ISmapper was processed in R using tidyverse (v1.2.1) (https://cran.r-project.org/package=tidyverse). For six isolates that carried MDR genes but no plasmid replicon genes, ISMapper was re-run with the minimum read depth threshold reduced to 1 read (from the default of 6 reads), in order to increase sensitivity to detect the associated IS1 site. This identified IS1 sites in five of the six isolates (SRR3049053, SRR5989319, SRR1967049, SRR5500465 and SRR5500455) but not in isolate SRR5500435.
The presence and subtypes of IncHI1 and IncN plasmids was further investigated using plasmid MLST (pMLST). Publicly available pMLST schemes for IncHI1[40] and IncN[41], available at https://pubmlst.org/plasmid/, were used to screen the relevant read sets using SRST2 (v0.1.8)[42]. The five isolates that were determined to have pST2 versions of the IncHI1 plasmid were then mapped to the reference sequence of plasmid pAKU1 (accession AM412236) with RedDog, using the same quality control thresholds as described for the chromosomal SNVs. The coverage level of the plasmid was addressed determining the ratio of average sequencing depth for the chromosome and the plasmid for all five isolates, with the ratio in the range 0.89–2.29 in all cases. A pST1 plasmid sequence (pUI1203_01 accession ERR340785) from S. Typhi strain UI1203 (genotype 3.2.1, isolated from Laos in 2001)[11] was included to provide an outgroup for tree rooting. The resulting SNV alignment was filtered to exclude non-backbone regions (66,458 bp) of the pAKU1 plasmid as previously defined [4] of the plasmid, and to exclude SNV sites present in <95% of plasmid sequences, and recombinant regions detected by Gubbins (v2.3.2), resulting in a final alignment of 17 SNVs for phylogenetic inference as above.
Genomes with less common AMR profiles were assembled using Unicycler v0.4.6 [43]. The assemblies were further interrogated for the presence of plasmid replicon genes using ABRicate (https://github.com/tseemann/abricate) and the PlasmidFinder database using a minimum coverage and minimum identity of 90. The location of AMR genes detected, IS1 sequences and IncHI1 plasmid backbone genes was visualised in Bandage v0.8.1 [44].
Trees annotated with MDR, IS1 insertion sites, plasmid replicons and QRDR point mutations were visualised using ggtree v1.8.1[45]. An interactive version of the S. Typhi ML phylogeny with associated metadata is found at microreact (https://microreact.org/project/r1njJ_qJ4).
The minimum inhibitory concentration (MIC) of ciprofloxacin was measured for 173 isolates (n = 117 from 2015 and n = 56 from 2016) as part of a previously reported study[35]. Briefly, MIC was determined by agar dilution using Mueller-Hinton agar, and S. Typhi were interpreted as displaying resistance (MIC ≥0.5 μg/mL) or susceptibility to ciprofloxacin (MIC <0.06 μg/mL), according to the EUCAST guidelines (v7.1)[35]. For the 173 phenotyped isolates, the total number of QRDR mutations per isolate was plotted against the reported MIC, in R using ggplot2 (v3.0.0)[46].
A total of 533 S. Typhi isolates were referred from English cases to PHE during the three-year period between April 2014 and March 2017 (n = 178, n = 176, n = 179 in each year running from April to March; listed in S1 Table). Of these, 449 cases (84.2%) reported recent foreign travel (within 28 days of onset of symptoms), spanning 26 countries. A further 39 cases (7.3%) reported no travel abroad. The proportion of cases reporting no recent travel, which may represent local transmission within the UK, displayed a non-significant increase from 5.6% in the first year to 8.9% in the third year (p = 0.5 using Chi-squared trend test). No information on travel was available for the remaining 45 cases (8.4%). The distribution of travel destinations for the cases with known travel history are given in Table 1 and Fig 1. The majority of cases (n = 387, 73%) reported travel to South Asia, particularly India (36%), Pakistan (29%) and Bangladesh (6%), reflecting travel and migration patterns between England and typhoid endemic areas. Annual case numbers from India and Bangladesh were constant (mean 64 and 11 per year, respectively; p = 0.2 in each case using Chi-squared trend test), but case numbers from Pakistan spiked significantly in the third year of the study, to n = 68 (38% of total cases) compared to n = 45 and n = 43 (25%) in the first two years (p = 0.01 using Chi-squared trend test). Small numbers of cases reported travel to South East Asia (n = 10, 2%), East Africa (n = 21, 4%), West Africa (n = 11, 2%) and the Middle East (n = 4, 1%; see Table 1). Five cases (1%) had reported travel to Europe or North America or South America only.
The 533 genomes were assigned to 31 unique S. Typhi genotypes using the GenoTyphi scheme (Fig 2). The majority (n = 391, 73.4%) belonged to subclade 4.3.1 (H58), which was identified in cases with travel to 14 different countries. Genotype 4.3.1 dominated amongst cases with reported travel to the South Asian countries associated with the greatest burden of travel-related cases (India, 79% 4.3.1; Pakistan, 92% 4.3.1; Bangladesh, 52% 4.3.1), and amongst cases with travel to East Africa (86% 4.3.1; see Table 1). Of the 4.3.1 isolates, 162 (41%) were further classified to lineage I (4.3.1.1), including 66% of the 4.3.1 isolates from cases with travel to Pakistan, 94% of those with travel to Bangladesh, and all 4.3.1 isolates from cases with travel to Mozambique, Tanzania and Zimbabwe (total n = 13). In contrast, lineage II (4.3.1.2) accounted for 156 (40%) of 4.3.1 genotypes, dominating in returning travellers to India (75% of 4.3.1 isolates) and accounting for the three 4.3.1 isolates associated with travel to Rwanda and Uganda (Fig 3). Notably, 4.3.1 was not detected in cases associated with travel to West Africa (0/11; see Table 1, Fig 3). Other common genotypes include clade 3.3 (6% of total isolates), found in cases with travel to India (9% of all isolates from this location), or Bangladesh (15%); clade 2.2 (3.1% of total isolates), found in cases with travel to India (2%) or Pakistan (4.5%); subclade 3.2.2 (2.6% of total isolates), found in 21% of cases with travel to Bangladesh; and subclade 3.1.1 (1.7% of total isolates), found in 7/11 (63%) of cases with travel to West Africa, including Nigeria (n = 4/6), Ghana (2/2) and Sierra Leone (1/1).
The overall frequency of MDR, defined as carriage of genes associated with resistance to ampicillin (blaTEM), chloramphenicol (catA1) and co-trimoxazole (sul1 or sul2 plus a dfrA gene), was 129 (24%). MDR S. Typhi was associated with reported travel to 10 countries from three main regions, South Asia (26%), East Africa (52%) and West Africa (27%), with high frequencies amongst isolates whose cases report travel to Zimbabwe (8/10, 80%), Nigeria (2/6, 33%), Tanzania (2/4, 50%), Ghana (1/2, 50%), Bangladesh (8/33, 24%) and Pakistan (85/156, 55%) (Table 1, Fig 3). MDR was also present at low frequency amongst cases with reported travel to India (3%), and singleton MDR isolates were detected from Mozambique, United Arab Emirates and Afghanistan. Notably the majority of MDR isolates were linked with travel to Pakistan (n = 85, 66% of all MDR isolates) and other South Asian countries (n = 100, 78% of all MDR isolates; see Table 1). No ESBL or carbapenemase genes were detected.
All MDR isolates belonged to either 4.3.1 (n = 125) or 3.1.1 (n = 4, see Fig 2b). The frequency of MDR was highest in 4.3.1.1 (96%), then 3.1.1 (3%), and one isolate in 4.3.1 (Figs 2b and 3). At the regional level, MDR was common amongst cases associated with travel to East Africa (52%), West Africa (27%) and South Asia (26%), but not with travel elsewhere (with the exception of one isolate from UAE; see Table 1). However, there were significant differences between countries within these regions, associated with differences in the dominant S. Typhi clades (Fig 3). In East Africa and South Asia, MDR was detected amongst cases associated with travel to the countries dominated by MDR-associated lineage 4.3.1 lineage I (Mozambique, Tanzania and Zimbabwe in East Africa; and Bangladesh and Pakistan in South Asia; see Table 1, Fig 3). In West Africa, MDR was detected in three isolates (n = 2 Nigeria, n = 1 Ghana), all belonging to the region’s dominant subclade 3.1.1 (a fourth MDR 3.1.1 infection had no reported travel).
QRDR mutations were identified in 455 isolates (85.4%) originating from 18 countries (Table 1, S1 Table). QRDR mutants were most frequent in India (97%), Bangladesh (94%), Nepal (66%), Pakistan (94%), Myanmar (100%), Uganda (100%), and Nigeria (50%). Singleton QRDR mutants were also identified in China, Malaysia, Afghanistan, United Arab Emirates, Egypt, Rwanda, Peru, Greece and Zimbabwe. QRDR triple mutants were identified only in isolates associated with travel to South Asia: 33% of isolates from Nepal, 23% of those from India and 4% of those from Pakistan (Table 1). Overall, the presence of QRDR mutations was very common in cases associated with travel to South Asia (95% of all isolates linked to this region), and significantly less common (p < 2.2e-16 using Chi-squared trend test) in East Africa (43%,) and West Africa (27%; see Table 1).
Ciprofloxacin MICs have been previously reported for 173 of the S. Typhi isolates (S1 Table)[35], and a comparison of QRDR mutations with these phenotypes is shown in Fig 4 to facilitate the interpretation of QRDR mutations. These data showed that all isolates carrying a single QRDR mutation had ciprofloxacin MIC of at least 0.064 μg/mL (exceeding the EUCAST threshold for susceptibility); all those with two QRDR mutations had MIC of 0.125 μg/mL; and all those with three mutations had MIC ≥1 μg/mL (above the threshold for resistance) (Fig 4). The most common mutations were gyrA-S83F (n = 271, 51%) and gyrA-S83Y (n = 83, 16%), which were found in 16 genotypes in cases associated with travel to eleven and six countries respectively (Table 2). Double mutants (gyrA-S38F or -S83Y plus a mutation in parC), associated with elevation of ciprofloxacin MIC to ≥0.25 μg/mL (Fig 4), were found in 16 isolates (3.0%) associated with travel to India or Bangladesh. A total of 61 isolates were identified as QRDR triple mutants, which display resistance to ciprofloxacin (MIC >1 μg/mL, see Fig 4) and have been associated with fluoroquinolone treatment failure[8]. Most of these (n = 59) carried gyrA-S83F, gyrA -D87N and parC-S80I mutations, while the remaining two had a unique profile of gyrA-S83F, gyrA-D87G and parC-S80I and of gyrA-S83FY, parC-Y74X and parC-P98X (Table 2).
All QRDR double and triple mutants belonged to genotype 4.3.1 and were associated with travel to South Asia (or no/unknown travel). Overall, QRDR mutations were significantly more common in 4.3.1 lineage II compared to lineage I (99% vs 91%, p = 0.006, two-sided test of difference in proportions). The 61 QRDR triple mutants were detected only in S. Typhi 4.3.1 lineage II isolates (Fig 2c) and were associated with either travel to India (n = 43, 22.5% of all isolates from this location, where lineage II is dominant); travel to neighbouring countries Pakistan (n = 6) or Nepal (n = 1); travel to multiple destinations including India (n = 5); no reported travel (n = 2); or no travel information available (n = 4) (Tables 1 and 2). Notably because all 4.3.1 MDR isolates belonged to lineage I, and all QRDR triple mutants belonged to lineage II (Fig 2), there were no isolates with both MDR and three QRDR mutations (Fig 2, S1 Table). There were however 119 cases with MDR plus 1–2 QRDR mutations (i.e. reduced susceptibility to fluoroquinolones, Fig 4). The vast majority of these were 4.3.1 lineage I (n = 115), most commonly associated with travel to Pakistan (n = 85, 54.4% of isolates from this country), Bangladesh (n = 8, 24%) and India (n = 6, 3%) but also occasional cases who reported travel to Zimbabwe (n = 2), Tanzania (n = 2) and United Arab Emirates (n = 1). Three 3.1.1 isolates were also MDR with one QRDR mutation (n = 2 travel to Nigeria, n = 1 with no recorded travel).
All MDR isolates belonged to 4.3.1 (n = 125) or 3.1.1 (n = 4) and carried the typical S. Typhi MDR composite transposon comprising Tn6029 (encoding blaTEM-1, sul2, strAB) inserted in Tn21 (carrying a class I integron encoding dfrA alleles in the gene cassette and sul1 at the end), which is in turn inserted within Tn9 (encoding catA1)[47] (see Fig 5a). All 125 MDR 4.3.1 isolates (associated with South Asia and East Africa) carried dfrA7 in the integron cassette and no plasmid replicons (Fig 5a). In most of these (n = 123, 98%), we detected chromosomal IS1 insertions at sites previously associated with IS1-mediated integration of the MDR composite transposon (S1 Fig) (cyaA or yidA sites[12,18]). A putative IS1 insertion was detected in the novel site STY3168 in a single, genome (SRR5500440). Further, of the six isolates with four IS1 sites detected, five had recent travel to Pakistan with last isolate having no reported travel (S1 Fig). Notably, most (93%) of the MDR 4.3.1 isolates also carried a QRDR mutation, the most common being gyrA-S83F (Table 2, S1 Fig). The four MDR 3.1.1 isolates (associated with West Africa) carried IncHI1 PST2 plasmids with dfrA15 in the integron cassette (Fig 5b). An IncHI1 PST2 plasmid was also identified in a single non-MDR 2.3.1 isolate associated with travel to Nigeria. The plasmid backbone was very closely related to that of the 3.1.1 West African plasmids but carried dfrA1 in the integron cassette and lacked the chloramphenicol and ampicillin resistance genes catA1 and blaTEM (Fig 5a and 5b).
Plasmid replicon and AMR gene screening identified additional plasmid replicons and AMR genes at low frequency. Two of the MDR 4.3.1.1 isolates from Zimbabwe carried IncN (subtype PST3) plasmids. In addition to the genes typical of the MDR composite transposon (with dfrA7 in the integron), these isolates also carried qnrS, dfrA14 and tet(A). It was not possible to resolve the precise locations of the acquired AMR genes due to the limitations of short read assembly. As these isolates lack QRDR mutations, the presence of the qnrS gene is predicted to confer reduced susceptibility to fluoroquinolones but not full resistance; indeed, one of the isolates (SRR4063811) was phenotyped and displayed reduced susceptibility to ciprofloxacin, with MIC 0.25 μg/mL. IncN plasmids (subtype PST5) were found in three non-MDR 4.3.1.2 isolates that carried dfrA15, sul1 and tet(A). These isolates (two from India, one with no reported travel) were also QRDR triple mutants and thus predicted to be fully resistant to fluoroquinolones. The combination of dfrA14, sul2, blaTEM-1, strA and strB was detected in three 4.3.1.1 isolates with no QRDR mutations (two with travel to Tanzania, one with no travel information). All three carried sequences with similarity (100% nucleotide identity and 50% coverage) to a FIBK plasmid carrying dfrA14, sul2 and blaTEM-1 that was previously sequenced from a 2008 Tanzanian S. Typhi, strain 129–0238 (GenBank accession LT904889)[12]. The same combination of AMR genes (dfrA14, sul2, blaTEM-1, strA and strB) plus tet(A) were also found in a 3.1.1 gyrA-S83Y isolate from Nigeria that harboured an IncY plasmid replicon.
Here we demonstrate the utility of using S. Typhi WGS data generated routinely at a public health laboratory in the UK to serve as informal surveillance for different geographical regions where typhoid fever is endemic. The PHE dataset encompasses a diverse collection of 533 S. Typhi isolates from multiple geographical regions collected over a three-year sampling period (Fig 1). As typhoid fever has not been endemic in England since the successful interventions of the major controlling measures, water sanitation and hygiene complemented with antimicrobial therapy[27,48], it is assumed that notified cases to PHE are associated with returned travellers and their contacts. The collection is biased towards isolates from South Asia and East Africa which reflects historical and contemporary ties between England and these regions, however these are also regions that experience a high burden of typhoid fever and could benefit from the AMR and WGS data obtained routinely by PHE. Notably, public health agencies in other countries receive S. Typhi isolates reflecting the distinct travel habits of their own populations (for example Institut Pasteur receives more S. Typhi from travellers visiting Francophone countries in Africa and former French colonies such as Vietnam), and the synthesis of these diverse collections could potentially provide more extensive sentinel surveillance coverage of typhoid endemic regions.
Typing the S. Typhi isolates using the GenoTyphi scheme enabled rapid classification of the 533 genomes into 31 distinct lineages (Fig 2). This simple tree-free approach showed clustering of subclades by geographical region of travel that was consistent with previous previously reported geographical patterns[11], providing support for the use of travel associated isolates as an indicator of local pathogen populations. Notable examples include the detection of MDR subclade 4.3.1 in East Africa and South Asia [8,15], and the complete absence of subclade 4.3.1 in the isolates from West Africa[14] (Fig 3). Genomes from cases reporting travel to West Africa were genotyped as 3.1.1, consistent with earlier studies where 3.1.1 was found to be the main S. Typhi lineage in the region[13,14]. Furthermore, for isolates with travel to more than one country, the genotype can help to discern the most likely origin of the infection; for example, for one case reporting travel to Nigeria and Turkey, the genome isolate (SRR558502) was identified as genotype 3.1.1, suggesting that the pathogen was most likely acquired in Nigeria. This further demonstrates the public health utility of WGS data on returning traveller isolates. Notably the GenoTyphi scheme provides a mechanism and nomenclature for such insights to be achieved simply and rapidly from individual genomes and by different laboratories working independently, without need for phylogenetic tree construction or other comparative analyses.
Data from this study revealed that reduced susceptibility to fluoroquinolones was common amongst S. Typhi associated with diverse geographic sites and genotypes. However, the ciprofloxacin resistant QRDR triple mutants were all from cases belonging to subclade 4.3.1.2 (Fig 2, Table 2), the majority of which had reported travel to South Asia, mainly India. None of these QRDR triple mutants were also MDR, which was associated with cases belonging to subclade 4.3.1.1. This is in line with previous observations, and others have speculated there may be a fitness cost to the carriage of QRDR triple mutants particularly in the MDR background [19]. These data align with previous findings[8,19], and support the hypothesis that high levels of fluoroquinolone exposure in India through healthcare and the environment are driving the emergence of resistant S. Typhi and other pathogens in the region[49] and contributing to treatment failure for typhoid fever[8]. Further, in addition to confirming that QRDR triple mutants are resistant to ciprofloxacin, the MIC data clearly showed that even a single QRDR mutation is associated with reduced ciprofloxacin susceptibility (MIC 0.06–0.25 μg/mL) (Fig 4), which has also been associated with clinical failure[35,50]. Single QRDR mutations are not associated with a fitness cost[51], and are present in MDR isolates, suggesting that this may further limit treatment options. While phenotypic MICs are still advised for clinical treatment, inferring susceptibility to antimicrobials from WGS data is appropriate for surveillance purposes [35].
Strikingly in this data set all 4.3.1 MDR isolates carried the composite transposon integrated in the chromosome, suggesting an enormous shift in the burden of MDR typhoid from plasmid-borne resistance. This is of grave concern as it means that there is likely to be very little fitness cost associated with carriage of the MDR transposon. In the late 1990s the increase in MDR typhoid in Asia prompted a switch to fluoroquinolones for treatment, which in Nepal and other regions was followed by almost complete loss of the MDR plasmid from the S. Typhi population, suggesting that the fitness cost of the plasmid leads to plasmid loss in the absence of selection from the first-line drugs. However, the integration of the MDR transposon into the S. Typhi chromosome likely alleviates any fitness cost, making it more likely that MDR will be maintained even in the absence of selection for the specific resistances encoded.
A key limitation of our MDR element analysis is that the routine genomic surveillance data generated at PHE is restricted to short-read Illumina sequences, which prohibited the full resolution of complex MDR loci and their precise locations in the S. Typhi genome. Nevertheless, using short-read mapping approaches, we were able to confirm that in nearly all MDR isolates the MDR loci could be localised to plasmids or chromosomal integration sites that have been previously fully resolved in S. Typhi using long reads [52,18]. Notably, 98% of plasmid-free MDR isolates showed evidence of chromosomal integration of the MDR element at the known sites yidA or upstream of cyaA in the 4.3.1.1 lineage. The other MDR isolates may harbour the MDR transposon in novel genetic contexts, which could potentially be resolved using long-read sequencing. PHE has demonstrated the utility of this approach previously, including in S. Typhi [28], but does not currently conduct long-read sequencing as a routine part of Salmonella genomic surveillance.
Of particular note was the increase in number of IS1 insertions in from cases in subclade 4.3.1.1 with reported travel to Pakistan (S1 Fig). We may hypothesise that the S. Typhi from this region may be more likely to acquire novel mechanisms in response to local selective pressures. Indeed, the recent acquisition of an IncY plasmid harbouring blaCTX-M-15 and qnrS genes has resulted in the emergence of an XDR lineage of S. Typhi from Pakistan[24] and an IncI1 plasmid encoding blaCTX-M-15 in an S. Typhi from Bangladesh[52] provides evidence for this hypothesis, highlighting the importance of ongoing surveillance of these regions that experience a high burden of typhoid fever.
Currently there are three critical AMR threats posed by S. Typhi, namely the dissemination of mobile AMR genes mediating MDR profiles, the evolution of point mutations in gyrA and parC, two core housekeeping genes, that confer differing levels of fluoroquinolone resistance (Figs 2 and 4), and the recent emergence of XDR S. Typhi. The WGS data presented here provide insight into changing AMR dynamics within S. Typhi. Importantly, the concordance of genomic and phenotypic AMR data for ciprofloxacin resistance in this study (Fig 4) which have been extensively characterised for S. Typhi for multiple drugs previously[35], demonstrates the utility of WGS for robustly characterising AMR profiles. Here, two of the AMR threats where characterised within the S. Typhi collection, reflecting geographical differences in AMR profiles. While no XDR S. Typhi had been detected in the PHE collection between April 2014 and March 2017, the first XDR S. Typhi isolate in a returned traveller with recent travel to Pakistan shortly after XDR S. Typhi were reported from Pakistan has been identified[24,28]. This highlights the value in using these data as informal surveillance of S. Typhi as we hypothesise that resistance to second-line drugs such as azithromycin will arise under continued selective pressure as has occurred with previous drugs.
|
10.1371/journal.pntd.0001539 | Clear Genetic Distinctiveness between Human- and Pig-Derived Trichuris Based on Analyses of Mitochondrial Datasets | The whipworm, Trichuris trichiura, causes trichuriasis in ∼600 million people worldwide, mainly in developing countries. Whipworms also infect other animal hosts, including pigs (T. suis), dogs (T. vulpis) and non-human primates, and cause disease in these hosts, which is similar to trichuriasis of humans. Although Trichuris species are considered to be host specific, there has been considerable controversy, over the years, as to whether T. trichiura and T. suis are the same or distinct species. Here, we characterised the entire mitochondrial genomes of human-derived Trichuris and pig-derived Trichuris, compared them and then tested the hypothesis that the parasites from these two host species are genetically distinct in a phylogenetic analysis of the sequence data. Taken together, the findings support the proposal that T. trichiura and T. suis are separate species, consistent with previous data for nuclear ribosomal DNA. Using molecular analytical tools, employing genetic markers defined herein, future work should conduct large-scale studies to establish whether T. trichiura is found in pigs and T. suis in humans in endemic regions.
| Trichuriasis is a neglected tropical disease (NTD) caused by parasitic nematodes of the genus Trichuris (Nematoda), causing significant human and animal health problems as well as considerable socio-economic consequences world-wide. Although Trichuris species are considered to be relatively host specific, there has been significant controversy as to whether Trichuris infecting humans (recognized as T. trichiura) is a distinct species from that found in pigs (recognized as T. suis), or not. In the present study, we sequenced, annotated and compared the complete mitochondrial genomes of Trichuris from these two hosts and undertook a phylogenetic analysis of the mitochondrial datasets. This analysis showed clear genetic distinctiveness and strong statistical support for the hypothesis that T. trichiura and T. suis are separate species, consistent with previous studies using nuclear ribosomal DNA sequence data. Future studies could explore, using mitochondrial genetic markers defined in the present study, cross-transmission of Trichuris between pigs and humans in endemic regions, and the population genetics of T. trichiura and T. suis.
| Soil-transmitted helminths ( = geohelminths), including whipworm, are responsible for neglected tropical diseases (NTDs) of humans in developing countries [1]–[3]. Trichuris trichiura infects ∼600 million people worldwide. This parasite is transmitted directly via a direct, faecal-oral route. The thick-shelled (infective) eggs are ingested and then hatch, following gastric passage, in the small intestine. First-stage larvae (L1s) are released and migrate to the large intestine (caecum and colon), where they develop, following multiple moults, into adults (∼30–50 mm in length). The worms burrow their thin, thread-like anterior end into the mucosal lining of the large intestinal wall, feed on tissue fluids, mature and produce eggs. In the large intestines, large numbers of worms cause disease ( = trichuriasis), which is usually associated with entero-typhlocolitis and clinical signs, such as dysentery, bloody diarrhoea and/or rectal prolapse, in people with a high intensity of infection. Children (∼5–15 years of age) often harbour the largest numbers of worms [2]. Whipworms also infect other animal hosts, including non-human primates, pigs and dogs, and can cause clinical disease similar to trichuriasis of humans [4]–[6].
Based on current knowledge, Trichuris species are considered to specifically infect a particular host species or a group of related hosts. Trichuris species are usually identified based on host origin and the morphological features of the adult worm (spicule and pericloacal papillae) [7], [8]. However, it is not always possible to unequivocally identify and differentiate Trichuris species based on the morphology of adult worms alone. Importantly, T. trichuria cannot be unequivocally differentiated morphologically from T. suis or Trichuris from some other animals, such as non-human primates [7]. Over the years, there has been considerable discussion as to whether T. trichuira and T. suis are the same or distinct species [9]–[13], and whether humans can become infected with T. suis, and pigs with T. trichiura in endemic countries in which both host species live in close association. Although the authors of a recent molecular study suggested that T. suis is a separate species from T. trichiura [14], only a small number of specimens from one country (Spain) was used in this study, and amplicons (from the first and second internal transcribed spacers, ITS-1 and ITS-2, of nuclear ribosomal DNA) were subjected to cloning prior to sequencing, which has significant potential to lead to artefacts [15], [16]. Therefore, the findings from this study [14] need to be interpreted with some caution at this stage. Moreover, internal transcribed spacers (ITS) of nuclear ribosomal DNA might not be suited as specific markers for enoplid nematodes, because of sequence polymorphism (heterogeneity) that occurs within species (or individuals) [14], [17].
Given this heterogeneity in nuclear rDNA, barcoding from whole mitochondrial (mt) genomes (haploid) has major advantages, particularly when concatenated protein sequences derived from all coding genes are used as markers in comparative, phylogenetic-based analyses [18]–[25]. Therefore, in the present study, we (i) characterised the mt genomes of human-derived Trichuris and pig-derived Trichuris, (ii) compared these genomes and (iii) then tested the hypothesis that human-Trichuris and pig-Trichuris are genetically distinct in a phylogenetic analysis of sequence data sets representing both genomes and those from selected nematodes for comparative purposes.
This study was approved by the Animal Ethics Committee of the Lanzhou Veterinary Research Institute, Chinese Academy of Agricultural Sciences. For the collection of Trichuris from humans, the subjects provided informed, written consent. All pigs, from which Trichuris specimens were collected, were handled in accordance with good animal practices required by the Animal Ethics Procedures and Guidelines of the People's Republic of China.
Adult specimens of Trichuris were collected from the caecum of a human patient during surgery in Zhanjiang People's Hospital in Zhanjiang, Guangdong Province, China. Adult specimens of Trichuris were also collected from the caecum from a pig slaughtered in an abattoir in Zhanjiang in the same province. Adult worms from each host were washed separately in physiological saline, identified morphologically [9], [26], fixed in 70% (v/v) ethanol and stored at −20°C until use. Total genomic DNA was isolated separately from two individual worms (coded Ttr2 and TsCS1 for human-Trichuris and pig-Trichuris, respectively) using an established method [27]. The region spanning ITS-1, the 5.8S gene and ITS-2 was amplified from each of these individuals by PCR using previously reported primers [14] and sequenced directly. The ITS-1 sequence of the human-Trichuris sample had 99.3% similarity with that of T. trichiura from human in Thailand (GenBank accession no. GQ352554). The ITS-1 and ITS-2 sequences of the pig-Trichuris sample had 98.6% and 98.5% similarity with that of T. suis from pigs in Spain (GenBank accession nos. AJ781762 and AJ249966, respectively) [14].
To obtain some mt gene sequence data for primer design, we amplified regions (400–500 bp) of the cox1 and nad1 genes by using (relatively) conserved primers JB3/JB4.5 and JB11/JB12, respectively [28], and of nad4 and rrnL genes using primers designed in this study (Table 1) by PCR. The amplicons were sequenced from both directions, using BigDye terminator v3.1, ABI PRISM 3730. We then designed primers (see Table 2) to regions within cox1, nad1, nad4 and rrnL and amplified from total genomic DNA (from an individual worm) the entire mt genome in three (for human-Trichuris) or four (for pig-Trichuris) overlapping fragments (of ∼2–4 kb each) between nad1 and nad4, nad4 and rrnL, and rrnL and cox1, cox1 and nad1. The cycling conditions used were 92°C for 2 min (initial denaturation), then 92°C for 10 s (denaturation), 50°C for 30 s (annealing), and 60–68°C for 10 min (extension) for 10 cycles, followed by 92°C for 10 s, 50°C for 30 s, and 60–68°C for 10 min for 20 cycles, with a cycle elongation of 10 s for each cycle and a final extension at 60–68°C for 7 min. Each amplicon, which represented a single band in a 0.8% (w/v) agarose gel, following electrophoresis and ethidium-bromide staining, was column-purified and then sequenced using a primer walking strategy [29].
Sequences were assembled manually and aligned against the complete mt genome sequences of other nematodes (available publicly) using the computer program Clustal X 1.83 [30] to infer gene boundaries. The open-reading frames (ORFs) and codon usages of protein-coding genes were predicted using the program MacVector v.4.1.4 (Kodak), and subsequently compared with that of Trichinella spiralis [31]. Translation initiation and translation termination codons were identified based on comparison with those reported previously [31]. Codon usages were examined based on the relationships between the nucleotide composition of codon families and amino acid occurrence, for which codons are partitioned into AT rich codons, GC-rich codons and unbiased codons. The secondary structures of 22 tRNA genes were predicted using tRNAscan-SE [32] and/or manual adjustment [33].
Amino acid sequences inferred from the 12 protein-coding genes (i.e. not atp-8) common among all of the nematodes included here were concatenated into a single alignment, and then aligned with those of 9 other enoplid nematodes (Trichinella spiralis, GenBank accession number NC_002681; Xiphinema americanum, NC_005928; Hexamermis agrotis, NC_008828; Agamermis sp., NC_008231; Romanomermis culicivorax, NC_008640; Romanomermis iyengari, NC_008693; Romanomermis nielseni, NC_008692; Strelkovimermis spiculatus, NC_008047; Thaumamermis cosgrovei, NC_008046), using the chromadorean nematode, Brugia malayi (NC_004298) as the outgroup. Any regions of ambiguous alignment were excluded using Gblocks (http://molevol.cmima.csic.es/castresana/Gblocks_server.html; Talavera and Castresana 2007) using stringent selection criteria (do not allow many contiguous nonconserved positions). Phylogenetic analyses were conducted using three methods: Bayesian inference (BI) analysis was conducted with four independent Markov chain runs for 1,000,000 metropolis-coupled MCMC generations, sampling a tree every 100th generation in MrBayes 3.1.1 [34]; the first 2,500 trees represented burn-in, and the remaining trees were used to calculate Bayesian posterior probabilities (pp). Maximum likelihood (ML) analyses were performed using PhyML 3.0 [35], and the (mtREV for amino acid sequences and GTR for rrnL nucleotide sequences) models were determined based on the Akaike information criterion (AIC). Bootstrap support was calculated using 100 bootstrap replicates. Maximum parsimony (MP) analysis was conducted using PAUP 4.0 Beta 10 [36], with indels treated as missing character states; 1,000 random additional searches were performed using TBR. Bootstrap support was calculated using 1,000 bootstrap replicates, and 100 random taxon additions in PAUP. Phylograms were drawn using the program Tree View v.1.65 [37].
Two primers, rrnLF (5′-TAAATGGCCGTCGTAACGTGACTGT-3′) and rrnLR (5′- AAAGAGAATCCATTCTATCTCGCAACG-3′), were employed for PCR amplification and subsequent sequencing of a portion (471 bp for human-Trichuris and 482 bp for pig-Trichuris) of the large subunit of mt ribosomal RNA (rrnL) from multiple individuals of human- and pig-derived Trichuris (Table 3). The rrnL sequence from T. spiralis (accession number NC_002681) [31] was used as the outgroup in phylogenetic analyses, because this morphologically distinct species is related to Trichuris [38]. All rrnL sequences were aligned using Clustal X, and the alignment was modified manually, based on the predicted secondary structure of the rrnL for Trichuris [31], and then subjected to phylogenetic analysis using the same methods as described above.
The complete mt genome sequences were 14,046 nt (human-Trichuris) and 14,436 nt (pig-Trichuris) in length, respectively (GenBank accession numbers GU385218 and GU070737). Each mt genome contains 13 protein-coding genes (cox1-3, nad1-6, nad4L, cytb, atp6 and atp8), 22 transfer RNA genes and two ribosomal RNA genes (rrnS and rrnL) (Table 4). The genes nad6 and cox3 overlapped by 25 bp (human-Trichuris); nad5 overlaps by 7 and 1 bp with tRNA-His, tRNA-Ser(AGN) by 8 and 3 bp with rrnS, tRNA-Asp by 13 and 4 bp with atp8 for human-Trichuris and pig-Trichuris, respectively. The atp8 gene is encoded (Fig. 1), as is typical for adenophorean nematodes [31]. The protein-coding genes are transcribed in different directions, as described for T. spiralis and X. americanum [31], [39]. Except for four protein-coding genes (nad2, nad5, nad4 and nad4L) and six tRNA genes (tRNA-Arg, tRNA-His, tRNA-Met, tRNA-Phe, tRNA-Pro and tRNA-Thr) encoded on the L-strand, all other genes were encoded on the H-strand. The AT-rich regions are located between nad1 and tRNA-Lys, and nad3 and tRNA-Ser(UCN), which differs from those of secernentean nematodes [19], [33]. The nucleotide composition of the entire mt genome is biased toward A and T, with T being the most favoured nucleotide and G being least favoured, which is consistent with mt genomes of some other nematodes for which mt genomic data are available [18], [19], [21], [31], [33]. The overall A+T content is 68.1% for human-Trichuris (33.6% A, 34.5% T; 15.0% G and 16.9% C) and 71.5% for pig-Trichuris (35.6% A, 35.9% T; 13.4% G and 15.1% C).
Protein-coding genes were annotated by aligning sequences and identifying translation initiation and termination codons by comparison with inference sequences for other nematodes. For both human-Trichuris and pig-Trichuris, the lengths of protein-coding genes were in the following order: nad5 (1548–1557 bp) >cox1>nad4>cytb>nad1>nad2>atp6>cox3>cox2>nad6>nad3>nad4L>atp8 (165–171 bp) (Table 4). The longest gene is nad5, and the lengths of the nad1 and nad3 genes are the same for human-Trichuris and pig-Trichuris (Table 5). The inferred nucleotide and amino acid sequences of each of the 13 mt proteins of human-Trichuris were compared with pig-Trichuris. For individual genes, the nucleotide and amino acid sequence differences between human-Trichuris and pig-Trichuris vary from 25.4 to 37.4% and 13.6 to 62.5%, respectively (Table 5).
A total of 3559 and 3562 amino acids are encoded in the mt genome of human-Trichuris and pig-Trichuris, respectively. As the mt genomes of nematodes can contain non-standard initiation codons [18], the identification of initiation codons can sometimes be challenging. For human-Trichuris, five genes (cox1, cox2, cox3, cytb and nad4) start with ATG and eight genes (nad1, nad2, nad3, nad5, nad6, nad4L, atp6 and atp8) use ATA. All genes have complete termination codons, 11 genes (cox1, cox2, cox3, nad1, nad2, nad3, nad4, nad4L, nad5, nad6 and atp6) use TAA and two genes (atp8 and cytb) use TAG as a termination codon, respectively. For pig-Trichuris, six genes (nad2, nad3, nad4 nad4L nad5 and atp6) start with ATA, four genes (cox1, cox2, cox3 and cytb) with ATG, and one gene (atp8) with TTG. All protein-coding genes have complete termination codon; seven genes (cox2, cox3, nad3, nad4, nad6, atp6 and atp8) stop with TAA and six genes (cox1, nad1, nad2, nad5, nad4L and cytb) use TAG. No abbreviated stop codons, such as TA or T, were detected. Such codons are known to occur in the mt genomes of other nematodes, such as Strongyloides stercoralis (cytb, nad4, nad1 and atp6) and T. spiralis (cytb and nad4) and Caenorhabditis elegans (nad1 and nad3) [19], [31], [40]. The atp8 gene was inferred by comparison with the homologous gene of T. spiralis. The ORF of this gene was located between genes tRNA-Asp and nad3 and inferred by the presence of a methionine for human-Trichuris and a leucine for pig-Trichuris.
Twenty-two tRNA genes were predicted from the mt genomes of human-Trichuris and pig-Trichuris and varied from 50 to 67 nt in length. Most of the tRNA genes are smaller than the corresponding genes in the mt genomes of other nematodes due to a reduced TΨC stem-loop region (TV-replacement loop) or DHU stem-loop region [39]. Most of the tRNA gene sequences can be folded into conventional secondary four-arm cloverleaf structures. In these tRNA, there is a strict conservation of the sizes of the amino acid acceptor stem (11–15 bp) and the anticodon loop (7 bp). Their D-loops consist of 5–9 bp. The two tRNA-Ser each contain the TΨC arm and loop, but lack the DHU arm and loop.
The two ribosomal RNA genes (rrnL and rrnS) of human-Trichuris and pig-Trichuris were inferred based on comparisons with sequences from T. spiralis; rrnL is located between tRNA-Val and atp6, and rrnS is located between tRNA-Ser(AGN) and tRNA-Val. The length of rrnL is 1011 bp for both human-Trichuris and pig-Trichuris. The lengths of the rrnS genes are 698 bp for human-Trichuris and 712 bp for pig-Trichuris. The A+T contents of rrnL for human-Trichuris and pig-Trichuris are 72.5% and 76.4%, respectively. The A+T contents of rrnS for human-Trichuris and pig-Trichuris are 69.9% and 75.4%, respectively.
Two AT-rich non-coding regions (NCRs) were inferred in the mt genomes of both human-Trichuris and pig-Trichuris. For these genomes, the long NCR (designated NCR-L; 162 bp and 144 bp in length, respectively) is located between the nad1 and tRNA-Lys (Fig. 1), has an A+T content of 71–72%. This overall A+T content is lower than those reported for nematodes (77.9–93.1%) studied to date [18], [19], [21], [33]. In this NCR, there are also 26 nt (human-Trichuris) and 17 nt (pig-Trichuris) AT dinucleotide repeats. Similar repeats have been detected in this region in C. elegans and A. suum [40]. For both human-Trichuris and pig-Trichuris, the short NCR (NCR-S; 93 bp and 117 bp in length) is located between genes nad3 and tRNA-Ser (UCN) (Fig. 1), with an A+T content of 65.6% and 84.6%, respectively. This region contains dinucleotide [AT]26 repeats and might form a hairpin loop structure (cf. AAAAAAAATTTTTTTTTT). Although nothing is yet known about the replication process in the mt DNA of parasitic nematodes, the high A+T content and the predicted structure of the AT-rich NCRs suggest an involvement in the initiation of replication [41].
A substantial level of nucleotide difference (32.9%) was detected in the complete mt genome between an individual of human-Trichuris and pig-Trichuris from China. The sequence variation detected in the 13 protein-coding genes (25.4–47.4%) and in NCRs (36.6%) was consistent with previous findings of variation in the nucleotide sequences of the nuclear ITS rDNA from human and pig [14]. However, for many nematodes [42], [43], there is usually greater within-species variation in mt protein-coding genes than in the ITS. For example, the magnitude of the nucleotide sequence variation in the 12 common mt protein genes (3–7%) [20] was greater than the 15 (1.8%) variable positions in the ITS (over 852 bp) detected among multiple individuals of the human hookworm, N. americanus [44].
Comparison between human- and pig-derived Trichuris from China also revealed variation at 1299 amino acid positions in the 13 predicted mt protein sequences. This level of amino acid variation (36.4%) is very high, given that mt proteins are usually conserved within a species due to structural and functional constraints [45]. In addition, previous studies of other nematodes have detected little to no within-species variation in protein sequences. For example, no within-species variation was detected in a COX1 region of 131 amino acids for N. americanus and for related hookworms, including A. caninum Ancylostoma and A. duodenale [24], [33]. Similarly, amino acid substitutions were recorded at only two of 196 (1%) positions (based on a comparison of conceptually translated sequences originating from GenBank accession nos. AF303135-AF303159) in partial COX1 among 151 N. americanus samples from four locations in China [46]. In the present study, the greatest numbers of amino acid differences between human-Trichuris and pig-Trichuris were in the NAD4 (n = 235; 40.9%), NAD5 (n = 220; 35.9%) and NAD2 (n = 120; 34.1%) sequences; these percentages were significantly higher than that (4.9–10%) between the two hookworms A. caninum and A. duodenale [24], [33]. The nature, extent and significance of the amino acid sequence variation between Trichuris from the human and pig hosts and from different geographical origins needs to be evaluated further, because there is virtually no published data on the magnitude of within-species variation in mt protein sequences for members of the genus Trichuris.
Genetic variation between human- and pig-Trichuris was also detected here in the two mt ribosomal RNA gene subunits (rrnL and rrnS). These subunits are usually more conserved in sequence than the protein genes [45], which is also supported by the present data. Comparison of the complete mt genomic data set between the two Trichuris individuals (Ttr2 and TsCS1) displayed less sequence variation in rrnS and rrnL (24.6% and 25.1%) compared with most protein genes (25.4–47.4%) and the non-coding regions (36.6%) (Table 5). A region (∼430 bp) in the conserved rrnL gene was used to examine the magnitude of genetic variation in Trichuris between the two different host species. A comparison of the partial rrnL sequences among 16 Trichuris individuals revealed 89 (20.7%) variable positions between human-Trichuris and pig-Trichuris, which is comparable with previous findings of a significant genetic difference (17%) in nuclear ITS between the two operational taxonomic units (OTUs) in Spain [14]. Taken together, the molecular evidence presented here supports the hypothesis that the gene pools of human-Trichuris and pig-Trichuris have been isolated for a substantial period of time and that they represent distinct species. In spite of the genetic distinctiveness recorded here between them, host affiliation is not strict [47]. Cross-infection of Trichuris between humans and pigs (both directions) has been described, but infection in the heterologous host is usually abbreviated [47].
In spite of the compelling evidence of genetic distinctiveness between Trichuris specimens from human and pig hosts, interpretation from this study needs to be somewhat guarded until detailed population genetic investigations have been conducted. Future studies could (i) explore, in detail, nucleotide variation in ribosomal and mt DNAs within and among Trichuris populations from humans and pigs from a range of different countries employing, for example, mutation scanning-coupled sequencing [27], (ii) establish, using accurate molecular tools, whether there is a particular affiliation between Trichuris and host in endemic regions and whether cross-host species infection is common or not, and (iii) attempt to establish an experimental infection of Trichuris of human origin in pigs, in order to be able to investigate the genetic and reproductive relationships between human-Trichuris and pig-Trichuris. Moreover, given the advent of high throughput genomic sequencing technologies, and the recent success in sequencing the nuclear genomes of the parasitic nematodes, B. malayi [48] and Ascaris suum [49], it is conceivable that the genomes of human-Trichuris and pig-Trichuris will be characterized in the near future. The transcriptome, and inferred proteome, characterised recently [50] will assist in future efforts to decode these genomes. Such work will pave the way for future fundamental molecular explorations and the design of new methods for the treatment and control of one of the world's socio-economically important nematodes [3]. This focus is important, given the impact of Trichuris and other soil-transmitted helminths (STHs), which affect billions of people and animals world-wide. Although Trichuris species are seriously neglected, genomics and related approaches provide new opportunities for the discovery of novel intervention strategies, with major implications for improving animal and human health and well being globally. In addition, the implications of genomic studies could also be highly relevant in relation to finding new treatments for immune-pathological diseases of humans [50]. Interestingly, various studies [51]–[55] have indicated that iatrogenic infections of human patients suffering from immunological disorders (such as inflammatory bowel disease, IBD) with nematodes, such as pig-Trichuris eggs can significantly suppress clinical symptoms. Although the mechanisms by which Trichuris modulates the human immune system are still unclear [52], [56], [57], studies have proposed that a modified CD4+ T helper 2 (Th2)-immune response and the production of anti-inflammatory cytokines, such the interleukins (IL-) IL-4 and IL-10, contribute to the inhibition of effector mechanisms [56], [58], [59]. Therefore, detailed investigations of pig-Trichuris at the molecular level could provide enormous scope for studying immuno-molecular mechanisms that take place between the parasite and humans affected by autoimmune or other immune diseases. The mt genetic markers defined in the present study should be useful to verify the specific identity of Trichuris employed in such studies.
|
10.1371/journal.pmed.1002124 | Scheduled Intermittent Screening with Rapid Diagnostic Tests and Treatment with Dihydroartemisinin-Piperaquine versus Intermittent Preventive Therapy with Sulfadoxine-Pyrimethamine for Malaria in Pregnancy in Malawi: An Open-Label Randomized Controlled Trial | In Africa, most plasmodium infections during pregnancy remain asymptomatic, yet are associated with maternal anemia and low birthweight. WHO recommends intermittent preventive therapy in pregnancy with sulfadoxine-pyrimethamine (IPTp-SP). However, sulfadoxine-pyrimethamine (SP) efficacy is threatened by high-level parasite resistance. We conducted a trial to evaluate the efficacy and safety of scheduled intermittent screening with malaria rapid diagnostic tests (RDTs) and treatment of RDT-positive women with dihydroartemisinin-piperaquine (DP) as an alternative strategy to IPTp-SP.
This was an open-label, two-arm individually randomized superiority trial among HIV-seronegative women at three sites in Malawi with high SP resistance. The intervention consisted of three or four scheduled visits in the second and third trimester, 4 to 6 wk apart. Women in the IPTp-SP arm received SP at each visit. Women in the intermittent screening and treatment in pregnancy with DP (ISTp-DP) arm were screened for malaria at every visit and treated with DP if RDT-positive. The primary outcomes were adverse live birth outcome (composite of small for gestational age, low birthweight [<2,500 g], or preterm birth [<37 wk]) in paucigravidae (first or second pregnancy) and maternal or placental plasmodium infection at delivery in multigravidae (third pregnancy or higher). Analysis was by intention to treat.
Between 21 July 2011 and 18 March 2013, 1,873 women were recruited (1,155 paucigravidae and 718 multigravidae). The prevalence of adverse live birth outcome was similar in the ISTp-DP (29.9%) and IPTp-SP (28.8%) arms (risk difference = 1.08% [95% CI −3.25% to 5.41%]; all women: relative risk [RR] = 1.04 [95% CI 0.90–1.20], p = 0.625; paucigravidae: RR = 1.10 [95% CI 0.92–1.31], p = 0.282; multigravidae: RR = 0.92 [95% CI 0.71–1.20], p = 0.543). The prevalence of malaria at delivery was higher in the ISTp-DP arm (48.7% versus 40.8%; risk difference = 7.85%, [95% CI 3.07%–12.63%]; all women: RR = 1.19 [95% CI 1.07–1.33], p = 0.007; paucigravidae: RR = 1.16 [95% CI 1.04–1.31], p = 0.011; multigravidae: RR = 1.29 [95% CI 1.02–1.63], p = 0.037). Fetal loss was more common with ISTp-DP (2.6% versus 1.3%; RR = 2.06 [95% CI 1.01–4.21], p = 0.046) and highest among non-DP-recipients (3.1%) in the ISTp-DP arm. Limitations included the open-label design.
Scheduled screening for malaria parasites with the current generation of RDTs three to four times during pregnancy as part of focused antenatal care was not superior to IPTp-SP in this area with high malaria transmission and high SP resistance and was associated with higher fetal loss and more malaria at delivery.
Pan African Clinical Trials Registry PACTR201103000280319; ISRCTN Registry ISRCTN69800930
| Malaria infection during the course of pregnancy can have devastating consequences on the mother and unborn child.
Intermittent preventive treatment in pregnancy (IPTp) with the antimalarial sulfadoxine-pyrimethamine (SP) is one of the main interventions to protect pregnant women during pregnancy in malaria endemic areas in sub-Saharan Africa.
The effectiveness of SP, however, is threatened by increasing resistance of the malaria parasite to this drug in east and southern Africa.
We conducted this study to evaluate if an alternative strategy consisting of screening pregnant women for malaria with rapid diagnostic tests at regular intervals during pregnancy and then treating the test-positive women with dihydroartemisinin-piperaquine (DP) would reduce the risk of malaria infection and the adverse consequences to the mother and newborn. This strategy is called intermittent screening and treatment in pregnancy (ISTp).
Our team conducted a two-arm, open-label trial to compare the effect of the new ISTp with DP (ISTp-DP) strategy against the existing IPTp with SP (IPTp-SP) strategy (the control arm) in 1,873 pregnant women in southern Malawi, where almost all of the malaria parasites were highly resistant to SP.
We found that the rate of malaria infection was high in both groups and that the new ISTp-DP strategy was not any better than the existing IPTp-SP strategy in terms of reducing malaria infection or improving pregnancy outcomes; in fact, women in the ISTp-DP arm had more malaria than women in the IPTp-SP arm.
ISTp-DP with the current generation of rapid diagnostic tests is not a viable alternative strategy to replace IPTp-SP in malaria endemic areas in sub-Saharan Africa, despite the high levels of resistance to SP.
IPTp with SP should still be used as one of the interventions against malaria in pregnancy in sub-Saharan Africa.
Further studies to explore alternative drugs that can replace SP for IPTp will be required in these areas of high SP resistance.
| Malaria during pregnancy is a major preventable cause of poor birth outcomes in sub-Saharan Africa [1]. In sub-Saharan Africa, the World Health Organization (WHO) currently recommends intermittent preventive treatment in pregnancy (IPTp) with sulfadoxine-pyrimethamine (SP) (IPTp-SP) for HIV-seronegative women. The effectiveness of IPTp-SP to clear peripheral parasitemia decreases in areas where parasites are resistant to SP; this resistance results from a series of mutations in the parasite genes that encode the targets of pyrimethamine (dhfr) and sulfadoxine (dhps). For example, in settings where >90% of parasites harbor high-level SP resistance encoded by five mutations in dhfr and dhps, up to 40% of asymptomatic parasitemic women who receive SP for IPTp are parasitemic again by day 42, reflecting the failure of SP to clear existing plasmodium infections and prevent new infections [2]. Nevertheless, even in these high resistance settings, SP retains some beneficial effect on birthweight [2,3]. However, an additional mutation at codon 581 in dhps is emerging in parasites in East Africa that renders IPTp-SP unable to inhibit parasite growth and may significantly compromise IPTp-SP when present [4–6]. Consequently, alternative approaches are required to prevent malaria during pregnancy.
Most of the proposed alternative drugs to replace SP are too poorly tolerated for IPTp use, including amodiaquine alone or combined with SP [7], mefloquine monotherapy [8,9], and the fixed-dose combination of chloroquine-azithromycin [10]. A proposed alternative strategy to IPTp consists of scheduled antenatal testing with rapid diagnostic tests (RDTs) and the treatment of RDT-positive women with artemisinin-based combination therapy (ACT), referred to as intermittent screening and treatment in pregnancy (ISTp) [11]. In West African settings, where parasite resistance to SP is low, ISTp with artemether-lumefantrine (AL) (ISTp-AL) was not inferior to IPTp-SP in reducing low birthweight and was well-accepted by providers and patients [12–14]. Nevertheless, in these studies, women in the ISTp-AL arm had lower mean birthweights and more clinical malaria during pregnancy.
We hypothesized that, owing to widespread parasite SP resistance, ISTp with the ACT dihydroartemisinin-piperaquine (DP) would be superior to IPTp-SP for the prevention of the adverse sequelae of malaria in pregnancy. However, a recent trial in an area with high levels of malaria transmission and parasite resistance to SP in western Kenya showed that ISTp with DP (ISTp-DP) was not superior to IPTp-SP and was associated with increased incidence of clinical malaria and malaria infection [15]. These findings need to be confirmed urgently in other areas that have high levels of parasite resistance to SP. Here we report the results of a similar trial comparing IPTp-SP against ISTp-DP in Malawi.
Ethical approval was obtained from the Liverpool School of Tropical Medicine (LSTM) and the Malawian National Health Science Research Committee. Written informed consent was obtained from all participants prior to randomization.
This was a three-site, open-label, two-arm individually randomized superiority trial using a stratified design with one strata for primi- and secundigravidae (paucigravidae) and one for multigravidae (third pregnancy or higher). The study was conducted at the Mpemba and Madziabango Health Centers and the Chikwawa District Hospital in southern Malawi. The area has moderate to intense year-round malaria transmission and high levels of SP resistance, as evidenced by near fixation of parasites harboring mutations at codons 51, 59, and 108 of dhfr and 437 and 540 of dhps [16,17].
Women of all gravidae attending their first antenatal visit were eligible if they were HIV-seronegative, were resident in the study catchment area and willing to deliver at the study clinics/hospital, had a hemoglobin > 70 g/l, had a pregnancy between 16 and 28 wk gestation, and had not yet received IPTp-SP. Exclusion criteria included multiple gestation and other high-risk pregnancies according to national guidelines, previous enrollment in the same study, and history of allergy to any of the study drugs.
Randomization sequences were computer-generated by the study statistician at LSTM, one for each gravidity strata and study site, using variable block randomization and an allocation ratio of 1:1. In each clinic, eligible women were allocated to the IPTp-SP or ISTp-DP arm by the coordinating study staff in order of their study identification number by drawing sequentially numbered opaque envelopes containing the allocation arm from a box corresponding to each gravidity stratum. Following allocation, women and care providers were aware of the arm allocation. All laboratory staff were blinded to the treatment assignment. The study statistician remained blinded until after database lock and approval of the statistical analysis plan by the data and safety monitoring board.
At enrollment, demographic, socioeconomic, and educational information was collected, a medical and obstetric history taken, and the gestational age ascertained by ultrasound. A 5-ml venous blood sample was taken for malaria microscopy, PCR, immunology, and testing for syphilis, HIV serostatus, and hemoglobin concentration (Hemocue). All women received a long-lasting insecticide-treated net.
Participants were randomized to receive either IPTp-SP or ISTp-DP, at enrollment and all subsequent scheduled antenatal visits. The IPTp-SP arm received three tablets of SP (500 mg/25 mg sulfadoxine/pyrimethamine tablets). If they had fever or history of fever, they were tested for malaria by RDT. RDT-positive women were treated with AL and then received their first course of SP during the first scheduled follow-up visit. Women in the ISTp-DP arm were screened for malaria using the histidine-rich protein 2 (HRP2)/plasmodium lactate dehydrogenase (pLDH) combination RDT (First Response Malaria pLDH/HRP2 Combo Test, Premier Medical Corporation). All RDT-positive women in the ISTp-DP arm received a standard 3-d course of DP (Eurartesim, Sigma Tau; 40 mg/320 mg dihydroartemisinin/piperaquine tablets) at a dose of 2.5, 3, 3.5, and 4 tablets for women weighing <50, 50–59, 60–69, and ≥70 kg, respectively. All SP and DP doses were provided with a slice of dry bread as directly observed therapy. All doses in both arms were supervised. In case of vomiting within 30–60 min, the full dose was repeated. If the repeat dose was vomited, the women received AL. Sigma Tau provided the Eurartesim free of charge.
The follow-up schedule consisted of three or four scheduled antenatal visits every 4 to 6 wk: four if enrolled at 16–24 wk gestation or three if enrolled at ≥25 wk gestation. At each such visit, a clinical and obstetric examination was conducted, and a blood sample taken for RDT (ISTp-DP arm), malaria microscopy, and PCR. Hemoglobin was assessed during the last scheduled visit. Women were encouraged to make unscheduled visits if they felt ill or were concerned about their pregnancy. In the IPTp-SP arm, women with uncomplicated clinical malaria (fever/history of fever and RDT-positive) during or in between scheduled visits received AL. Women with uncomplicated malaria in the ISTp-DP arm received DP, or AL if they had received DP within the previous 4 wk.
At delivery, a maternal venous sample was taken for the same malaria metrics, and a placental and cord-blood sample for histology, RDT, microscopy, and PCR. Children were weighed and the gestational age assessed using the modified Ballard score [18]. The presence of congenital abnormalities and jaundice was assessed at delivery, at day 7, and at the final visit at 6–8 wk, coinciding with their childhood vaccination visit. In between scheduled visits, infants were followed passively.
RDT results were used to determine care. RDT positivity was defined as either pLDH or HRP2 antigen positivity. See S1 Text for details of microscopy and real-time PCR used for detection and identification of parasites, as well as baseline parasite genotyping.
The primary outcome among paucigravidae was “adverse live birth outcome,” defined as the composite of having a singleton baby born small for gestational age (SGA) [19] or with low birthweight (<2,500 g), or preterm (<37 wk) (S1 Text). The primary outcome among multigravidae was a composite of any evidence for plasmodium infection at delivery detected in peripheral maternal blood (microscopy, RDT, or PCR) or placenta (incision smear, impression smear, PCR, or active or past infection detected by histology) (S1 Text). The rationale for using a different primary outcome for multigravidae was based on systematic reviews showing that preventing plasmodium infection by IPTp-SP or long-lasting insecticide-treated nets is associated with improved birth outcomes primarily among women in their first and second pregnancies [20,21]. Plasmodium infection status at delivery was used as the primary outcome in multigravidae because plasmodium infection is associated with an increased risk of malaria [22–25] and anemia [26–28] in infancy, particularly in those born to multigravidae [24].
Key secondary efficacy outcomes included the individual components of the composite primary outcomes, fetal loss (spontaneous abortion at <28 wk gestation, stillbirth), any adverse birth outcome (adverse live birth outcome or fetal loss), maternal hemoglobin concentrations and anemia, clinical malaria (documented fever/history of fever plus positive malaria RDT), plasmodium infection, mean birthweight, mean gestational age at delivery, congenital plasmodium infection (cord blood positive at birth by microscopy, RDT, or PCR, or clinical malaria within 7 d of birth with parasitological confirmed diagnosis by microscopy or RDT), neonatal and infant (by 6–8 wk) clinical malaria, all-cause severe anemia and all-cause illness detected at scheduled or unscheduled postnatal visits, and perinatal and infant mortality by 6–8 wk.
The primary safety outcomes included maternal death, severe cutaneous skin reaction in the mothers within 30 d of drug intake, other serious adverse events (SAEs) in the mother or infant, congenital malformations, and neonatal jaundice.
See S1 Text for details about sample size calculations. Log binomial regression was used for binary endpoints to obtain relative risk (RR) values and corresponding 95% confidence intervals. The identity-link function was used to obtain risk differences. Linear regression was used for continuous variables, and results expressed as mean difference (95% CI). The unadjusted analysis, stratified by gravidity (pauci- and multigravidae), was considered the primary analysis. Secondary, covariate-adjusted analyses for the primary endpoints were conducted using seven prespecified covariates (in addition to gravidity and site) and simple imputation for missing covariates (<1%). These same covariates were included in subgroup analyses. Poisson regression with time of follow-up as an offset was used for count variables to obtain incidence rate ratios (95% CIs). A two-sided p-value < 0.05 was used to define statistical significance. The intention to treat (ITT) analytical population was defined as all eligible women who were randomized and contributed to the outcome. The per protocol population included women who attended every scheduled visit, who took all the daily study doses on each occasion, and who contributed to the endpoint. For the safety analysis, women in the ISTp-DP arm were considered overall and split by recipients and non-recipients of DP (i.e., those who were RDT-negative throughout). All analyses were prespecified, unless otherwise indicated, in a statistical analysis plan (see S2 Text) approved by the data and safety monitoring board. Analysis was done in SAS version 9.3 and Stata version 14.
Between 21 July 2011 and 18 March 2013, 3,214 women were screened for inclusion; 1,873 women were randomized (paucigravidae, n = 1,155; multigravidae, n = 718). Recruitment was stopped when the full sample size for paucigravidae had been reached (S1 Text). Of the randomized women, 1,743 (93.1%) were seen at delivery (Fig 1). Overall, 6,504 of 6,942 (93.7%) scheduled antenatal follow-up visits were attended (S1 Table), and 1,742 women (94.5%) attended all scheduled visits. Ultimately, 1,676 (89.5%) contributed to the primary endpoint, with proportions of participants equally distributed between the study arms (ISTp-DP arm, 89.3%; IPTp-SP arm, 89.6%) (S2 Table). The baseline characteristics were well balanced between the study arms, within each gravidity strata, and overall (Table 1). At baseline, about half of the women were infected with malaria parasites, and this proportion was slightly lower (not significant) in those not contributing to the primary analyses (S3 Table). Overall, 99.5% and 2.7% of the parasites harbored the dhps K540E and A581G mutation, respectively (S4 Table). In both arms, the median (interquartile range) follow-up time was 4.0 (3.2–4.7) mo, and the median (range) number of scheduled visits was 4 (1–4) (S1 Table). In the ISTp-DP arm, 48.8%, 38.0%, 12.4%, and 0.9% received 0, 1, 2, and 3 courses of DP, respectively (S1 Table). Overall, 3,048 and 604 courses of SP and DP were administered in the respective study arms, and in the IPTp-SP arm, 251 courses of AL were administered for clinical malaria (S1 Table).
Among paucigravidae, the prevalence of adverse live birth outcome was similar in the ISTp-DP (33.7%) and IPTp-SP (30.6%) arms (RR = 1.10 [95% CI 0.92–1.31], p = 0.282; Fig 2). The prevalence was also similar between arms among multigravidae.
Among multigravidae, the risk of malaria at delivery was higher in the ISTp-DP (34.9%) than in the IPTp-SP (27.2%) arm (RR = 1.29 [95% CI 1.02–1.63], p = 0.037). This increased risk was also evident among paucigravidae and all gravidae. In absolute terms, the risk of malaria was increased in multigravidae by 7.8% (95% CI 0.6%–14.9%) and amongst all gravidae by 7.9% (95% CI 3.1%–12.6%) (Fig 2).
Similar results for both primary outcomes were obtained from prespecified covariate-adjusted analyses, with and without prespecified imputation for missing covariates (S5 Table), with per protocol population analysis (S6 Table), and in a sensitivity analysis that restricted analysis to birthweight obtained within 24 h of delivery (S7 Table). Results were also consistent across subgroups (S1 and S2 Figs), although the increased risk of malaria at delivery appeared lowest in primigravidae (S2 Fig).
Following enrollment, 45.8% of women had ≥1 episode of plasmodium infection prior to delivery (PCR, microscopy, or RDT), and 11.4% had ≥1 episode of clinical malaria. These proportions were similar in both arms (Fig 3). At delivery, 22.2% of women had peripheral malaria detected by PCR, RDT, or microscopy. This value was higher in the ISTp-DP arm (RR = 1.34 [95% CI 1.12–1.61], p = 0.002; S3 Fig), particularly for subpatent infections (PCR-positive, RDT- or microscopy-negative; S3 Fig). The overall prevalence of placental malaria detected by histology, PCR, RDT, or microscopy was 38.0%, and this value was higher in the ISTp-DP arm (RR = 1.16 [95% CI 1.03–1.32], p = 0.018; S4 Fig), reflecting differences in acute rather than chronic or past histological infections (S4 Fig). Congenital malaria was common (12.0%) in both groups (Fig 4).
At delivery, relative to the IPTp-SP arm, paucigravidae in the ISTp-DP arm had higher mean hemoglobin concentrations (S8 Table) and a lower prevalence of anemia (hemoglobin < 110 g/l) (Fig 3). The individual components of the primary endpoint adverse live birth outcome are provided in Fig 4. Low birthweight was more common in the ISTp-DP arm (RR = 1.29 [95% CI 0.97–1.71], p = 0.079).
Overall, DP was well tolerated (S9 Table). There was no difference between arms in the number of maternal SAEs or deaths (S10 Table). There were no severe cutaneous reactions. Fetal loss was highest in the ISTp-DP arm (2.6% versus 1.3%; Fig 4). Further stratified analysis within the ISTp-DP arm showed fetal loss was highest among women who had never received DP (i.e., who remained RDT-negative throughout) (3.1% versus 2.2% in DP recipients; S10 Table). Perinatal and infant (by 6–8 wk) mortality were not statistically different between the arms (perinatal mortality: RR = 1.62 [95% CI 0.87–2.99], p = 0.127; infant mortality: RR = 1.42 [95% CI 0.63–3.17], p = 0.398) (Fig 4), but overall, the composite of fetal loss or infant death by the end of follow-up (6–8 wk) occurred more often in the ISTp-DP arm (4.0% versus 2.3%, RR = 1.76 [95% CI 1.04–2.98], p = 0.036; Fig 4; S10 Table). One case of neonatal jaundice was detected in the ISTp-DP arm (mother was a non DP-recipient), and none in the IPTp-SP arm. The frequency of congenital malformations was 1.2% in the ISTp-DP arm (0.9% in infants whose mother was a DP-recipient) and 1.0% in the IPTp-SP arm (RR = 1.11 [95% CI 0.45–2.71], p = 0.824).
Despite the high levels of parasite resistance to SP, ISTp-DP was not superior to the standard IPTp-SP regimen in this trial: ISTp-DP was not associated with improvements in the composite outcome of small for gestational age, low birthweight, and preterm birth (primary outcome for paucigravidae) and was associated with more malaria at delivery (primary outcome for multigravidae) and more fetal loss. Although the relative increase in malaria risk was modest, this affected an additional eight out every 100 pregnancies. These results suggest that ISTp-DP may not be a suitable alternative strategy to replace IPTp-SP in settings similar to ours and may even predispose to unfavorable pregnancy outcomes in these settings.
The results may not be representative of areas where >10% of parasites harbor the “sextuple mutant” haplotype carrying the dhps A581G mutation [29]; however, our efficacy findings are similar to those reported recently from areas in western Kenya [15] with similarly high transmission (malaria prevalence detected by PCR at enrollment 33% versus 44% in this study) and SP resistance (5.8% dhps A581G mutation), and are also consistent with two previous non-inferiority trials conducted in West Africa, despite marked geographic differences in prevailing SP resistance, which is low in West Africa [11,12]. In both West African studies, ISTp was non-inferior to IPTp-SP in the reduction in low birthweight among paucigravidae; mean birthweights were higher in the IPTp-SP recipients than in those receiving ISTp-AL (p = 0.04) [12], but there was no significant difference compared to those receiving ISTp with amodiaquine-artesunate (p = 0.06) [11]. Additionally, the incidence of clinical malaria was higher in the ISTp-AL arm compared to the IPTp-SP arm. This was not observed in our trial, but the trial in western Kenya also observed higher incidence of clinical malaria as well as of plasmodium infection during pregnancy [15].
Because DP has very high anti-parasitic efficacy in Africa [30], the lack of superiority of ISTp-DP may result either from the ineffectiveness of ISTp as a strategy in high malaria transmission areas or from the continued effectiveness of IPTp-SP despite prevalent SP resistance. To this latter point, in our study area in Malawi, 99.5% of parasites harbor the “quintuple mutant” haplotype, but only 2.6%–4% carry the additional dhps A581G mutation [29]. Therefore, it is likely that IPTp-SP continued to provide some benefits, as has been observed in settings with similar parasite populations [2,3]. Another factor likely contributing to continued effectiveness of IPTp-SP was our use of the frequent dosing regimen [31] now recommended by WHO, which may mitigate the shortening of posttreatment prophylaxis that results from SP resistance [2]. It would also be of interest to further explore whether SP, which also has broad antimicrobial activity, may have conferred additional protection from other pathogens [32].
It is unlikely that suboptimal dosing or subtherapeutic levels of DP contributed to the non-superior performance of ISTp-DP: each dose was supervised, and there is no evidence that pregnancy alters the pharmacokinetics of DP to a degree that requires dose adjustment [33]. The same DP regimen was shown to be highly effective (PCR-corrected success rate by day 63: 99%) in a concurrent treatment trial conducted by the same team in this area using the same batch [34].
ISTp-DP may also have been ineffective owing to a failure to detect low-level parasitemias, although the biological impact of such infections during pregnancy is unclear [35]. RDTs detected about 45% of the PCR-positive infections in paucigravidae and about 30% in multigravidae, thereby allowing the majority of infections to persist in the placenta. Conceptually, ISTp is intended to prevent both existing infections from progressing and new infections from occurring for up to 6 wk after each DP course. Because only the RDT-positive women receive treatment, many do not benefit from the posttreatment prophylaxis. Furthermore, the infrequency of screening (approximately monthly) in a high transmission setting may have allowed new infections to develop and persist between scheduled visits. These factors combined may explain the higher prevalence of plasmodium infections at the time of delivery in the ISTp-DP arm.
The ineffectiveness of ISTp to prevent malaria in high transmission settings may also explain the higher rate of fetal loss in the ISTp-DP arm (2.6% versus 1.3%), consistent with the results from previous meta-analyses that showed a 1.5-fold higher risk of fetal loss among women randomized to control arms in trials of insecticide-treated nets [36]. The excess risk of fetal loss was not due to an adverse effect of DP, as the risk was highest among women who had never received DP (3.1% versus 2.2%). An alternative explanation could be that the broad antimicrobial effect of SP reduced the risk of fetal loss relative to ISTp [32]. Lastly, the effect could also be a chance finding, as the trial in Kenya did not observe an excess risk of fetal loss in the ISTp arm [15].
Overall, DP was well tolerated, which is consistent with the results of a recent four-arm treatment trial comparing the four fixed-dose ACTs in the case management of malaria in pregnancy [34]. This is important as almost all RDT-positive women in our trial were asymptomatic, and tolerance can be a major factor determining adherence.
ISTp is a labor-intensive strategy, but a separate qualitative substudy using in-depth interviews and focus group discussions showed it was highly acceptable to both patients and clinic staff. Although ISTp requires more frequent blood sampling, women appreciated its importance and the fact that they could be shown the RDT test results, corroborating findings from similar acceptability studies in Ghana [14,37]. The venous sampling at the first antenatal visit was deemed more convenient by women than repeated finger pricks, as it allowed health workers to tests for malaria, anemia, syphilis, and HIV testing with a single blood draw.
Limitations of our trial include the open-label design used. Another limitation is that we were not able to include a third arm with IPTp with DP as there was insufficient safety information for repeat courses of DP available at the time this trial was designed. Approximately 9% of the randomized women did not contribute to the primary outcome of adverse live birth outcome and 12% did not contribute to the primary outcome of plasmodium infection at delivery. However, this loss to follow-up was well balanced between the study arms, with little differences in baseline characteristics between those who contributed to the primary endpoint versus those who did not; thus, this loss to follow-up is unlikely to have biased the findings. The proportion of multigravidae reporting using a bednet the night prior to enrollment was slightly lower in the ISTp-DP arm; however, this did not explain the observed difference in the risk of plasmodium infection at delivery, as all women received an insecticide-treated net on enrollment, and bednet use thereafter was near universal in both arms (99% in each arm).
ISTp-DP was not superior to the existing IPTp-SP regimen in this area with high SP resistance in southern Malawi. These results should be equally relevant to other high endemic areas in east and southern Africa with similar or lower levels of parasite SP resistance. The identification of alternative drugs to replace SP remains a pressing research priority for the control of malaria in pregnancy before levels of SP resistance render IPTp-SP fully ineffective.
|
10.1371/journal.pntd.0000528 | Fluorescent Imaging of Antigen Released by a Skin-Invading Helminth Reveals Differential Uptake and Activation Profiles by Antigen Presenting Cells | Infection of the mammalian host by the parasitic helminth Schistosoma mansoni is accompanied by the release of excretory/secretory molecules (ES) from cercariae which aid penetration of the skin. These ES molecules are potent stimulants of innate immune cells leading to activation of acquired immunity. At present however, it is not known which cells take up parasite antigen, nor its intracellular fate. Here, we develop a technique to label live infectious cercariae which permits the imaging of released antigens into macrophages (MΦ) and dendritic cells (DCs) both in vitro and in vivo. The amine reactive tracer CFDA-SE was used to efficiently label the acetabular gland contents of cercariae which are released upon skin penetration. These ES products, termed ‘0-3hRP’, were phagocytosed by MHC-II+ cells in a Ca+ and actin-dependent manner. Imaging of a labelled cercaria as it penetrates the host skin over 2 hours reveals the progressive release of ES material. Recovery of cells from the skin shows that CFDA-SE labelled ES was initially (3 hrs) taken up by Gr1+MHC-II− neutrophils, followed (24 hrs) by skin-derived F4/80+MHC-IIlo MΦ and CD11c+ MHC-IIhi DC. Subsequently (48 hrs), MΦ and DC positive for CFDA-SE were detected in the skin-draining lymph nodes reflecting the time taken for antigen-laden cells to reach sites of immune priming. Comparison of in vitro-derived MΦ and DC revealed that MΦ were slower to process 0-3hRP, released higher quantities of IL-10, and expressed a greater quantity of arginase-1 transcript. Combined, our observations on differential uptake of cercarial ES by MΦ and DC suggest the development of a dynamic but ultimately balanced response that can be potentially pushed towards immune priming (via DC) or immune regulation (via MΦ).
| Schistosomiasis is caused by the parasitic worm Schistosoma with over 200 million people infected across 76 countries. The parasitic larvae (called cercariae) infect mammalian hosts via the skin, but the exact mechanisms by which dermal cells interact with molecules released by invading larvae are unclear. A better understanding of the infection process and stimulation of the early immune response would thus enable a targeted approach towards the development of drugs and vaccines. Here, we have used the fluorescent tracer CFDA-SE to label infectious cercariae and, together with confocal microscopy, have for the first time tracked in real time the parasite infecting via the epidermis and depositing excretory/secretory material in its wake. Phagocytic macrophages and dendritic cells in the skin internalised excretory/secretory molecules released by the larvae, and both cell types were subsequently located in the draining lymph nodes where priming of the acquired immune response occurs. In vitro studies determined that macrophages were slower to process released parasite material than dendritic cells; they also secreted lower levels of pro-inflammatory cytokines but greater quantities of regulatory IL-10. The relative abundance of macrophages versus dendritic cells in the skin infection site and their differential rates of antigen processing may be crucial in determining the success of adaptive immune priming in response to infection.
| Trematode parasites (e.g. Schistosoma sp, Fasicola sp, and Trichobilharzia sp) are important parasites of mammalian hosts in the developing, as well as the developed world, and cumulatively are a major health burden to humans and domestic animals. Infective schistosome larvae gain entry to the host as free-swimming cercariae which penetrate the host via a percutaneous route. The precise mechanism by which Schistosoma larvae penetrate the skin to facilitate their onward migration is a matter of debate [1]–[3]. Infection of mouse skin by S. mansoni cercariae occurs rapidly but many of the larvae are still in the skin by 40 hours [4],[5]. Excretory/secretory (ES) molecules released by invading larvae aid penetration of the skin but also lead to the stimulation, and down-regulation, of the dermal inflammatory response [6]. Indeed, the extended contact between ES molecules released by invading larvae and innate immune cells in the skin, particularly following exposure to protective radiation-attenuated (RA) larvae [7], indicates that the innate response may be critical in limiting the success of initial infection. Therefore, the innate immune system in the skin could provide a target for manipulation in the pursuit of anti-schistosome vaccines and/or drugs but the cellular target(s) and mechanisms by which larval ES molecules act on the innate immune response are poorly understood.
The skin is populated with a range of innate immune cells [8], and pro- and anti-inflammatory innate responses occur quickly following cercarial penetration [9]. An initial neutrophil-rich cutaneous response resolves shortly after the majority of larvae have left the skin [4],[10]. The cutaneous response also involves macrophages (MΦ) [7], dendritic cells (DC) [7] and Langerhans' cells (LC) [11] which form cellular foci around the sites of parasite entry [12],[13]. Activation of cells with antigen presenting function in the skin also directs their emigration to CD4+ rich areas of the skin draining lymph node (sdLN), where DCs and LCs have been observed to accumulate following exposure to RA schistosomes [11].
The ES products released in the first 3 hours after transformation of S. mansoni cercariae into schistosomula (termed 0-3hRP) stimulate cytokine production by MΦ in a MyD88-dependent fashion implying the involvement of one or more Toll-like receptors (TLR) [14]. Moreover, 0-3hRP stimulates DC that in turn drive strong Th2 responses both in vitro and in vivo [15], likely resulting from its capacity to limit the maturation and hence stimulatory capacity of the DC population [16]. Several studies have characterised the composition of ES material released by cercariae and was found to comprise a number of proteases [17]–[19], and molecules with potential immunomodulatory function (e.g. Sm16 [20]). Carbohydrates may also play an important role in the stimulation of innate immune cells and are abundant on the surface of cercariae [21], and wide range of O- and N-linked oligosaccharides are present in 0-3hRP [22]. Such glycans are known stimulators of C-type lectins such as DC-SIGN [23] and TLR-4 [24], and consequently may also be involved in the innate immune response to schistosome larvae.
To better understand how cercariae penetrate the skin and stimulate dermal inflammatory and regulatory factors, an efficient method of tracking the invading parasite and the fate of their ES is required. Histological studies [4],[7],[25] have localised larvae in relation to the dermal inflammatory reactions but they do not reveal whether the constituent cells have taken up parasite material, or whether they have become activated. Fluorescent amine reactive tracers such as carboxyfluorescein diacetate succinimidyl ester (CFDA-SE) provide a novel approach to label live cercariae and to investigate interactions between schistosome antigens and innate immune cells. Conventionally, CFDA-SE passively diffuses into cells where it is cleaved by free esterases and binds covalently to free amines on proteins as a fluorescent product [26] and has been used to label various bacteria and protozoa [27]–[29].
In this study, we are the first to label live S. mansoni cercariae with the fluorescent tracer CFDA-SE and to visualise the penetration of host skin in real time by labelled larvae. CFDA-SE was observed to preferentially label the contents of the acetabular glands but did not alter the immune stimulatory capacity of 0-3hRP released by CFDA-SE labelled invading larvae. Both MΦ and DC incorporated labelled 0-3hRP (pro-Th2 [15]) by phagocytosis, but the rate of translocation to lysosome-associated membrane protein-1 (LAMP-1+) phagosomes was retarded compared to that of E. coli (pro-Th1) bioparticles. Moreover, the rate of 0-3hRP uptake was faster and the extent of activation greater in DC than in MΦ. These observations provide insights into how schistosome infection may impact upon phagocytic cells of the innate immune response in the skin, and how this may affect the priming of the adaptive immune response in the skin-draining lymph nodes (sdLN).
Female C57BL/6 mice (8–12 weeks old) were bred and maintained at the University of York and housed under specific pathogen free (SPF) conditions in filter topped cages. All experiments were carried out within the guidelines of the United Kingdom Animal's Scientific Procedures Act 1986. All the research that involved the use of animals was approved by the University of York Ethics committee.
A Puerto Rican strain of S. mansoni was maintained by routine passage through outbred NMR-I mice and Biomphalaria glabrata snails. Cercariae were shed from snails harbouring patent schistosome infections by exposure to light for up to 2 hours. Isolated cercariae were washed ×3 by pulse centrifugation at 200 g in 10 ml of sterile aged tap water (ATW) and re-suspended.
Cercariae (∼1–5×104/ml) were incubated with various concentrations of the amine reactive tracer Vybrant CFDA-SE (Invitrogen Ltd, Paisley, UK) diluted with ATW at 28°C for 60 mins. Cercariae were concentrated by pulse centrifugation at 200 g followed by 3× washes in ATW prior to re-suspension in ATW and incubation for a further 60 mins to allow unconjugated dye to diffuse out of the parasite. Parasites were again washed 3× prior to measurement of fluorescence, or were fixed with 2% paraformaldehyde for 20 mins prior to imaging.
For collection of labelled 0-3hRP released from transforming cercariae, the protocol of Jenkins et al. [15] was modified. Suspensions of CFDA-SE labelled cercariae were mechanically-transformed [30] to separate heads from tails and then cultured in serum free RPMI 1640 (Invitrogen Ltd) containing 200 U ml−1 penicillin and 100 µg ml−1 streptomycin (Invitrogen Ltd) for 3 hrs. The supernatant containing 0-3hRP was concentrated in Vivaspin 15 tube (Sartorius Stedim Ltd, Epsom, UK) with a 5-kDa membrane. The protein concentration was determined using a Coomassie Plus-200 assay (Perbio Science Ltd, Cheshire, UK).
Aliquots of CFDA-SE labelled and unlabelled cercariae were placed in black 96 well clear bottom plates (Camlab Ltd, Cambridge, UK). Fluorescence was measured on a POLARstar OPTIMA microplate reader (BMG Labtech, Saitama City, Japan) (492±5 nm excitation; 520±5 nm emission). A manual count of cercariae per well was performed and data expressed as relative fluorescent units (RFU) per live cercaria.
Confocal or fluorescent microscopy was performed on both live and fixed parasites, or fixed cells, using a Zeiss confocal LSM 510 meta (Carl Zeiss Ltd, Welwyn Garden City, UK) or a Nikon Labophot fluorescent microscope equipped with a Nikon Coolpix 995 (Nikon Corp, Tokyo, Japan). All images were captured at ×10, ×20 or ×100 using identical laser settings at 488 nm excitation; 520 nm emission wavelengths with a pinhole setting of between 2–50 µm and parasites were manually kept in focus during skin penetration or movement. Photographic analysis was performed using Adobe photoshop or LSM image browser 4.2 (Carl Zeiss Ltd, UK) and 3D images reconstruction was performed and analysed using Volocity 4.3.2 (Improvision, Coventry, UK) from LSM Z stacks.
Anesthetised mice were infected via the pinnae [31] with 500 unlabelled or CFDA-SE labelled cercariae. After 30 mins, the remaining parasite suspension was collected and the number of non-penetrant cercariae established.
Peritoneal exudate cells (PEC) were extracted from mice by peritoneal lavage, 5 days post-injection with 0.5 ml sterile 3% Brewers thio-glycollate medium (Sigma–Aldrich) [14]. PEC were separated into adherent and non-adherent populations by adherence to plastic after culture for 2 hours at 37°C, 5% CO2.
BMMΦ were derived as follows. Femurs of naïve mice were removed, flushed with chilled PBS and the resulting cell suspension washed and re-suspended in Dulbecco's Modified Eagle's Medium (DMEM, Invitrogen Ltd) supplemented with 10% of heat-inactivated low endotoxin foetal bovine serum (Biosera, Ringmer, UK), 2 mM L-glutamine, 200 U ml−1 penicillin and 100 µg ml−1 streptomycin, 50 µM 2-mercaptoethanol (Invitrogen Ltd) and 20% L929 cell-conditioned medium (Gift P. Kaye, University of York). Cells were plated at 1×106 per well in 24 well plates (VWR, Luttworth, UK) and incubated at 37°C in 5% CO2 for 7 days. BMDC were obtained as previously described [15] following 7 days culture in the presence of 20 µg/ml GM-CSF (Peprotech, London, UK).
Adherent PEC, non-adherent PEC, BMMΦ, and BMDC were then cultured with the following; 1000 unlabelled or CFDA-SE labelled cercariae, or 40 µg/ml unlabelled or CFDA-SE labelled 0-3hRP. Supernatants from cell cultures were removed and stored at −20°C prior to cytokine analysis. The remaining cells were removed using chilled PBS, washed and re-suspended in chilled culture media prior to labelling with antibodies and analysis by flow cytometry.
Flow cytometric analysis of in vitro cultured cells, or those recovered ex vivo, was performed on a DakoCytomation Cyan ADP analyser (Dako, Ely, UK). Cells were initially blocked for 30 mins with anti-CD16/32 mAb in PBS containing 1% FCS,and 2 mM EDTA. Subsequently, cells were labelled with directly conjugated antibodies; F4/80 Pacific Blue (#BM8), CD40 allophycocyanin (#1C10), CD86 allophycocyanin (#GL1), IA/IE allophycocyanin (#M5/114.15.2) (Insight Biotechnology Ltd, Wembley, UK). Biotin conjugated antibodies against CD11c (#N418) and GR-1 (#Ly-6C) were probed with streptavidin allophycocyanin or streptavidin Pacific Blue (Invitrogen Ltd). All antibody concentrations were optimised and all analyses performed alongside irrelevant isotype controls. Data was analysed using Summit v4.3 (Dako, UK).
Cytokine levels were measured by ELISA. IL-6 was captured with anti-IL-6 mAb (#MP5-20F3) and probed with biotinylated anti-IL-6 mAb (#MP5-32C110) detected with streptavidin peroxidase conjugate (BD Pharmingen, Oxford, UK). IL-12p40, IL-10 and TNF-α were measured using kits (Invitrogen Ltd) according to the manufacturer's protocol. The lower sensitivity of the assays were 15 pg/ml (TNF-α), 20 pg/ml (IL-6), and 32 pg/ml (IL-12p40, IL-10).
Cell samples were re-suspended in Trizol (Invitrogen Ltd) and RNA extracted following the manufacturer's protocol. Extracted RNA was reverse transcribed into cDNA using Superscript II Reverse Transcriptase (Invitrogen Ltd), checked for quality and genomic DNA contamination, and 10 ng (5 µl) of each resulting cDNA sample analysed by real time PCR on an ABI PRISM 7000 sequence detection system (Applied Biosystems, Warrington, UK). Relative quantities of RNA were determined using Taqman probes (Sigma–Aldrich, UK). The specific primer pairs and probes were; Arginase, 5′- TCACCTGAGCTTTGATGTCG, 3′CTGAAAGGAGCCCTGTCTTG, Probe: 5′ GTTCTGGGAGGCCTATCTTACAGAGAAGGTCTCTAC, iNOS 5′- CTGCATGGACCAGTATAAGG, 5′- CTAAGCATGAACAGAGATTTCTTC, Probe: 5′ –AGTCTGCCCATTGCTG. The relative expression of each gene was normalised to the values for the GAPDH housekeeping gene before statistical analysis. GAPDH 5′ - CCATGTTTGTGATGGGTGTG, 5′- CCTTCCACAATGCCAAAGTT Probe: CATCCTGCACCACCAACTGCTTAGC.
Mice were infected with 1000 unlabelled or CFDA-SE labelled cercariae for 30 mins on each ear [31]. Pinnae from naïve and infected (unlabelled or CFDA-SE labelled cercariae) mice were collected at 3, 24, 48 and 72 hours. Pinnae were split and then floated on 50 µg/ml Liberase (Roche Products Ltd, Welwyn Garden City, UK) in RPMI 1640 and incubated at 37°C for 30 mins. Pinnae were then torn into large pieces using tweezers and incubated with shaking for a further 30 mins. Auricular lymph nodes (sdLN) that drain the pinnae were also removed from the infected mice. They were cut into small pieces and incubated with 0.2 mg/ml DNAse (Sigma–Aldrich, UK) and 0.5 mg/ml collagenase D (Roche Products Ltd) for 20 mins. Pinnae and sdLN cell suspensions were filtered through 100 µm metal gauze, washed in PBS pH 7.2 and enumerated prior to being labelled with antibodies and analysed by flow cytometry.
BMMΦ and BMDC were cultured as previously described and seeded at 0.2×106 onto circular cover slips. Cells were then stimulated for up to 18 hrs with CFDA-SE labelled 0-3hRP, and compared to the uptake of Alexa Fluor488 or 594-labelled E. coli bioparticles (1 µm ; Invitrogen Ltd) representing a control microbial material and known to be a classical pro-Th1 stimulant. Cells attached to the coverslips were then fixed in 2% paraformaldehyde and permeabilised using 0.2% saponin (Sigma–Aldrich) for 30 mins. Cells were stained with DAPI (Sigma–Aldrich), polyclonal anti-rabbit antibody against EEA-1 (Abcam plc, Cambridge, UK) and biotin conjugated mAb against LAMP-1 (#1D4B, Insight Biotechnology, UK). Cells were then washed ×3 in PBS pH 7.2 and incubated with anti-rabbit Alexa Fluor547 and streptavidin Alexa Fluor633 (Both Invitrogen Ltd). The cell coated cover slips were finally washed ×3 and fixed to a glass microscope slide with colourless nail varnish and Vectarshield (Vector laboratories, Peterborough, UK).
Z series images were collected using a Zeiss LSM 510 meta confocal microscope on four channels Ex/Em 420/480 (DAPI), 488/520 (CFDA-SE), 560/595 (Alexa Fluor547) and 633/640 (Alexa Fluor633). All multicolour samples had identical settings and were imaged sequentially; controls showed no bleed-through. Z-series were then converted to 3D images using Volocity 4.3.2 (Improvision®, UK). Co-localisation of CFDA-SE labelled 0-3hRP or Alexa 488 E. coli bioparticles with the intracellular markers were analysed using Volocity 4.3.2 software to generate a co-localisation coefficient Mx.The coefficient ranges from 0 to 1, with 1 indicating that the entire signal from one channel is co-localised with the other and 0 representing no co-localisation between channels. The threshold for each channel was generated automatically to exclude voxels for which Pearson's correlation between the channels is less than or equal to 0, based on a technique from Costes et al [32].
Changes in CFDA-SE labelled material uptake, cytokine production and differences in co-localisation were evaluated using Students t-test or one-way ANOVA (***,P<0.001; **, P<0.01; *, P<0.05). Differences were considered significant when P<0.05.
CFDA-SE dye preferentially labels material localised within the cercarial pre- and post-acetabular glands and their associated ducts as shown by the fluorescent 2D and 3D confocal images (Fig. 1A–B and supplementary Video S1). The relative absence of labelling on the outer surface of the cercaria, or within the body and tail, indicates that CFDA-SE does not cross the surrounding glycocalyx and suggests that the dye travels up the ducts to enter the acetabular glands which are rich in proteases [17],[18]. Optimal concentrations and incubation conditions for labelling live cercariae with CFDA-SE were established (see supplementary Figure S1 A–E). Furthermore, labelling parasites with CFDA-SE under these optimal conditions did not adversely affect the infective potential or viability of cercariae compared to unlabelled parasites since both sets of cercariae had almost identical penetration efficiencies of approximately 70% (see supplementary Figure S1); it also does not affect their ability to mature into adults, or lay eggs (data not shown)
Transformation of CFDA-SE labelled cercariae into schistosomula leads to a decrease in their fluorescence (Fig. 2A). This results from the release of CFDA-SE labelled acetabular gland contents (ES material) which accompanies the transformation of cercariae. The contents of the acetabular glands contain numerous proteases that conventionally facilitate penetration of host skin [17],[18]; this ES material was detected in the culture media, in which the parasites had transformed over 3 hours, as an increase in fluorescence (Fig 2A). CFDA-SE labelled ES material is clearly visible as discreet vesicles being released from the acetabular gland duct openings (Fig. 2B & C). CFDA-SE labelled material was only evident within the contents of the vesicle and not the surrounding material (Fig. 2C).
Live CFDA-SE labelled cercariae were cultured with PEC to examine the ability of phagocytic cells to internalise larval ES products. The majority of PEC obtained at this time point (day 5) were CD11b+ MΦ rather than neutrophils (data not shown). Uptake of CFDA-SE material was significantly greater by plastic adherent compared to non-adherent PEC (35.6±1.5% versus 9.8±0.5%; P<0.001; Fig. 3A), implying that the material had indeed been phagocytosed. The lack of uptake by non-adherent cells is confirmatory evidence that labelled material is not taken up as a non-specific event, or as free dye. Further evidence that uptake was a specific process is that CFDA-SE label localises within distinct intracellular components, whereas cells directly exposed to an equivalent concentration of CFDA-SE dye alone exhibited even distribution (Fig. 3B. Within the adherent population, ES material released by live cercariae enhanced the frequency of MHC-II+ cells (from 23.5% to >70%; Fig. 3C). Moreover, over 40% of MHC-II+ cells were also positive for CFDA-SE demonstrating that cells with potential antigen presenting function had taken up the ES material.
The fluorescence of bone marrow-derived macrophages (BMMΦ) activated with transformed CFDA-SE labelled cercariae and labelled 0-3hRP was significantly greater (P<0.001) compared to cells cultured with unlabelled controls, and untransformed labelled cercariae (Fig 3D). This is further evidence that the CFDA-SE material represents ES released by parasites as they transform from cercariae into schistosomula. The addition of either EGTA (5 mM) or cytochalasin D (10 µg/ml) significantly inhibited the uptake of CFDA-SE labelled 0-3hRP by 77.43% and 67.89% respectively (both P<0.001; Fig. 3E) showing that Ca+ and actin dependent receptor(s) are responsible for the majority uptake of cercarial ES products by an active phagocytic mechanism.
No difference was observed between the production of cytokines, expression of co-stimulatory markers and MHC-II between unlabelled and labelled cercariae or 0-3hRP, demonstrating that the presence of CFDA-SE on labelled proteins does not affect the stimulation of BMMΦ (see supplementary Figure S2).
Phagocytosis of ‘foreign’ molecules by host MΦ depends upon efficient endosomal trafficking of this material to phagosomes where it is degraded. The speed at which this occurs has been linked to the development of inflammatory (rapid) versus regulatory (delayed) processes [33],[34]. Therefore, the compartmentalisation of CFDA-SE labelled pro-Th2 0-3hRP [15] was compared to that of E. coli bioparticles labelled with Alexa Fluor488 that are a classical pro-Th1 stimulant. Using confocal microscopy, Z-stack images were acquired and analysed for the co-localisation (Mx coefficient) of labelled material with the early (EEA-1+) or late (LAMP-1+) phagolysosomes at different times.
0-3hRP initially trafficked into EEA+ phagosomes at a rate similar to that observed for E. coli bioparticles (15 mins; Fig. 4A). After 30 mins, 0-3hRP was still located within the EEA-1+ compartment although small amounts were also present in LAMP-1+ phagosomes (Fig. 4B; see supplementary Video S2). In contrast, at 30 mins, E. coli was almost exclusively located in LAMP-1+ compartments. By determining the co-localisation coefficients, while E. coli rapidly transferred out of EEA-1+ (Fig. 4C) into LAMP-1+ phagosomes (Fig. 4D; see supplementary Video S3), 0-3hRP was slower to translocate to the phagosome, suggestive of a reduced response by MΦ to 0-3hRP compared with E. coli.
To explore which cells in the skin interact with ES molecules released by larvae in vivo, CFDA-SE labelled parasites were used to infect the pinnae of C57BL/6 mice and invading parasites imaged by time lapse confocal microscopy. In the videos and accompanying stills, (Fig. 5A and supplementary Video S4 and Video S5), an infecting cercaria is observed to attach to the stratum corneum and then burrow into the upper layer of the epidermis with its tail detaching by approximately 20 mins. The brightness of cercarial tail is an artefact caused by the wide pinhole diameter and its proximity to the camera. CFDA-SE labelled material released from the acetabular glands is deposited first at (∼10 mins), and then surrounding (∼25 mins), the point of entry into the epidermis revealed as a ring of fluorescence. As the parasite burrows further within the epidermis, CFDA-SE labelled material is released via the oral sucker (between 10 to 120 mins), presumably aiding migration by depositing tissue digesting proteases ahead of the parasite's line of movement. As the parasite continues to migrate, fluorescence associated with the larval head progressively declines in the acetabular glands, compatible with the notion that the gland contents are released in order to facilitate parasite migration. Moreover, the migration path of the parasite is revealed as a trace of CFDA-SE+ material left in its wake.
Skin cells extracted from the pinnae of mice infected with labelled cercariae, and analysed by flow cytometry revealed that up to ∼7% of CD45+ cells were CFDA-SE+ 3 hrs after infection (Fig. 5B). By 48 hrs, there was a significant decrease (P<0.01) in the percentage of CD45+ cells that were CFDA-SE+, which was followed by a further decline by 72 hrs when the majority of parasites should have left the skin [4]. Phenotypic analysis of the cells showed that the labelled ES material was initially taken up by GR-1+MHC-II− cells (neutrophils), but by 48 hrs far fewer CFDA-SE+ GR-1+ MHC-II− cells were detected (Fig. 5C). Both F4/80+ MHC-II+ and CD11c+ MHC-II+ cells, predicted to be skin-derived MΦ and DC respectively, were also CFDA-SE+ demonstrating that antigen presenting cells (APC) in the skin had taken up ES material released in vivo by invading larvae, or had taken up apoptosing neutrophils that had previously taken up CFDA-SE ES material. The number of DC and MΦ recovered from the skin that were CFDA-SE+ peaked at 24 hrs but declined thereafter (P<0.01) possibly reflecting their onward migration to draining lymphoid tissues. Our data also show that the proportions of MΦ and DC in the skin that were CFDA-SE+ were similar at each time point (P>0.05; Fig. 5C). However, as the total number of MΦ in the pinnae after infection (CFDA-SE− and CFDA-SE+ cells combined) is approximately twice that of DC (5.37±0.39×105 cf. 3.01±0.27×105 at 24 hrs), we infer that MΦ are less efficient than DC at taking up CFDA-SE material.
CFDA-SE+ cells were also detected in the sdLN that drain the infection site. Although only negligible numbers of CFDA-SE+ cells were recorded in the sdLN by 24 hrs (0.74±0.37% of the large granular cells), a peak of 3.36±0.6% was detected at 48 hrs (Fig 5D). Virtually no CFDA-SE+ GR-1+ MHC-II− cells were detected in the sdLN at any time (data not shown) implying the lack of recruitment of neutrophils to this location, or that they had rapidly been removed following apoptosis. The vast majority of CFDA-SE+ cells in the sdLN were either F4/80+ MHC-II+ or CD11c+ MHC-II+ (Fig. 5E) indicating that MΦ and DCs which had taken up labelled parasite molecules in the skin had migrated to the sdLN. Alternatively, CFDA-SE labelled parasite antigen released by cercariae within the first 2 hrs as they penetrate (Fig. 5A and supplementary Video S5) may have drained freely to the sdLN and was processed by cells in situ. However, the lack of CFDA-SE+ cells at the earliest time point (3 hrs) in the sdLN would suggest that the incorporation of freely draining CFDA-SE released parasite material in the sdLN does not occur, although free fluorescein isothiocyanate painted directly on the skin could be detected in the sdLN by 3 hrs (data not shown). Rather, as the peak numbers of CFDA-SE+ MΦ and DC in the sdLN was reached at 48 hrs, we believe that this reflects the migration of antigen laden cells to the sdLN.
As cells expressing surface markers characteristic of MΦ and DC were both observed to take up CFDA-SE labelled molecules in the skin of infected mice, the relative reactivity of these two cell types to stimulation with molecules released by live cercariae was compared using parallel cultures of BMMФ and bone marrow derived dendritic cells (BMDC).
A similar number of BMDC and BMMΦ internalised ES material released from CFDA-SE labelled cercariae but significantly more BMDC internalised CFDA-SE labelled 0-3hRP (Fig. 6A; P<0.05). BMDC also internalised greater amounts of both CFDA-SE ES and 0-3hRP as reflected in the significantly greater MFI (Fig. 6B). In addition, BMDC incorporated 0-3hRP initially at a faster rate than BMMФ, and a greater proportion of BMDC had taken up CFDA-SE 0-3hRP at all time points (e.g. 50.27%±2.4 versus 32.47%±3.8 at 2 hrs; Fig. 6C). BMDC also expressed much higher MFI levels of CD40 CD86 and MHC II expression in response to both cercariae and 0-3hRP (Fig. 6D). Inflammatory cytokine (i.e. IL-6, IL-12p40 and TNF-α) output from BMDC in response to 0-3hRP was much greater than from BMMΦ whereas the production of regulatory IL-10 was significantly lower (P<0.01; Fig. 6E). This suggests that MΦ exhibit a more regulatory phenotype than DC exposed to ES products released by transforming cercariae. To examine this question, we used qPCR to reveal that BMDC expressed significantly higher levels of inducible nitric oxide synthase (iNOS) transcript than BMMΦ (P<0.01; Fig. 6F). This marker of ‘classical activation’ indicates that BMMФ are less active than BMDC at responding to 0-3hRP. In contrast, BMMФ had significantly elevated levels of arginase 1 mRNA (P<0.01) which converts L-arginine via an alternative pathway that does not yield toxic nitrogen products. This supports the idea that the cercarial ES promote the development of BMMΦ that are regulatory.
Both BMMФ and BMDC initially internalised CFDA-SE labelled 0-3hRP into the EEA-1+ compartment but BMDC translocated 0-3hRP into the LAMP-1+ phagosome faster than BMMФ (Fig. 7A). The rate of translocation of AF594 E. coli bioparticles was faster compared to CFDA-SE 0-3hRP for both BMDC and BMMΦ but did not significantly differ between the two cell types (Fig. 7B). The reduced translocation rate by BMMФ for 0-3hRP was visualised by co-localisation with LAMP-1+ phagosomes at 15 mins in BMDC, whilst it was still present in EEA-1+ compartments in BMMФ (see supplementary Figure S3 A). By 30 mins, almost all the CFDA-SE labelled 0-3hRP was observed in LAMP-1+ compartments in BMDC but in BMMФ, it was still found in the EEA-1+ compartments (see supplementary Figure S3 B). Co-culture of AF594 E. coli bioparticles and CFDA-SE 0-3hRP revealed that the two antigens were internalised into different phagosomes within the same cell (Fig. 7C) and did not affect each others translocation from the early phagosome to the phagolysosome (data not shown).
The skin and its associated immune cells is the first barrier to schistosome cercariae during infection but the impact of the dermal innate immune response on parasite survival and the development of adaptive immunity is largely unknown. In this study, cercariae were labelled with a fluorescent tracer in order to facilitate the visualisation of the parasite in the skin, the release of ES products, and uptake of parasite material by MΦ and DC in vitro and in vivo. Intriguingly, although both MΦ and DC take up labelled ES material released by cercariae (aka 0-3hRP), which is known to favour the development of pro-Th2 DC [15], it is processed at a slower rate than a typical pro-Th1 agent, and it is processed slower by MΦ than by DC. This suggests that the processing of parasite material in the skin by these two cell types might have significant effect on the balance of the immune environment to favour immune priming or immune regulation.
The amine reactive tracer CFDA-SE specifically labelled the contents of the cercarial acetabular glands which contain an abundance of tissue degrading protease enabling penetration of the skin [17],[19]. Indeed, the presence of esterases or proteases in these glands presumably facilitates the efficient cleaving of the dye from its inactive to its fluorescent state [26]. Importantly, we determined that CFDA-SE did not alter the ability of acetabular gland products to activate innate immune MHC-II+ cells. Indeed, there was no change in the expression of MHC-II, CD40 and CD86, or the production of IL-6, IL-12, IL-10 and TNF-α in response to labelled compared with unlabelled cercariae or 0-3hRP. Therefore, we conclude that CFDA-SE is an ideal fluorescent label with which to track the fate of parasite released molecules in relation to cells of the innate immune response.
CFDA-SE labelled material was released by cercariae only upon transformation into schistosomula. It was released as membrane-less vesicles [35] implying that the contents of the glands which contain numerous antigenic proteins and glycoproteins [18] are effectively labeled by CFDA-SE whilst still in the acetabular glands. As there is no de novo synthesis of protein by cercariae [30], the contents of the acetabular glands are pre-formed; as such it is the first antigenic material detected by the host's innate immune response. However, the lack of labeling on the outer portion of the vesicle suggests that other non-protein molecules such as lipids and/or glycans surround the protein rich contents as they are expelled from the sucker during skin penetration, and consequently may represent additional source of ligands for innate immune cells [36].
Fluorescent labelling of the parasites enabled for the first time detailed real time imaging of parasites as they penetrate into and through the skin. Cercariae were observed to firstly attach to the outer stratum corneum and then burrow into the epidermis. In every case, the cercarial tail detached from the parasite body at penetration and moved out of the field of view propelled by their continued movement. Cercarial tails (see supplementary Video S4 and Video S5) never entered the skin and therefore do not provide a source of material to modulate the immune response as suggested by others [37].
Shortly after attachment (∼10 min), the acetabular gland contents were released supporting the idea that gland material is used not only as an aid for attachment but also that it is required for entry into the stratum corneum [38]. Others have postulated that the stratum corneum offers little barrier to cercariae since in aqueous conditions its structural integrity is lost and the parasites simply push through [1],[3]. However, evidence shown here supports the view that acetabular gland material is released at the point of infection to aid penetration shown as a thick ring of fluorescent material. Whether this material is used to lyse cells, or the extra cellular matrix, is not clear. The progressive reduction in the fluorescence of the glands as the parasite migrates through the epidermis is revealed as a trace of fluorescence marking a ‘penetration tunnel’ [35],[39]. By 2 hrs, most of the acetabular gland content appears to be spent [40],[41]. However, some CFDA-SE material persists indicating that gland contents remain available for digesting the epidermal basement membrane [4],[10], the dermis [19], or even a blood or lymphatic vessel, facilitating the parasite's onward migration. Moreover, the presence CFDA-SE material in the skin until at least 48 hrs shows that some ES material persists in the skin, possibly as material just released by the invading larvae.
The uptake of CFDA-SE labelled material released from cercariae is largely an active actin and Ca+ dependent phagocytic process, as uptake was inhibited by EGTA and cytochalasin D. Although the receptors responsible for uptake of cercarial ES are presently unknown, they are likely to include CD206 and CD209 which could recognise mannose- and fucose-rich glycoproteins abundant in 0-3hRP [22]. As the cells that phagocytosed labelled ES material were MHCII+, we conclude that that the ultimate fate of phagocytosed gland contents is to be processed and then presented to the adaptive immune system. Cercarial ES and 0-3hRP were effective at activating these MHC-II+ cells through increased expression of MHC-II, co-stimulatory markers (CD40, CD86) and pro-inflammatory cytokines (IL-12p40/23, IL-6 and TNF-α) which would all promote their phagocytic activity, migratory capacity, and ability to act as effective APC.
The phagocytic machinery used to internalise foreign particles results in the formation of a phagosome that matures and plays a key role in initiation of the immune system. However, the endosomal processing pathway for pro-Th2 0-3hRP was retarded compared to a typical pro-Th1 stimulus E. coli. Some microbes aid their survival [42] by disrupting the TLR signalling pathway involved in phagosome development [33],[43]. For example, Mycobacterium tuberculosis arrests phagosome maturation by retaining EEA-1 on the phagosome [44]. Enhanced phagosome maturation (i.e. in response to E. coli) leads to increased processing of antigen to MHC-II molecules through the engagement of TLRs [45]. In the case of 0-3hRP, the reduced rate of phagosome maturation compared to E. coli could suggest that it has a reduced stimulatory response by being less efficient at binding and activating TLRs. Alternatively, 0-3hRP may trigger a different signalling pathway which does not efficiently promote phagosome maturation. For example, schistosome egg antigen (SEA) which also induces potent pro-Th2 DC [46] is reported to stimulate DC independent of TLR2, TLR4 and MyD88 [47],[48], whilst filarial ES-62 appears to use a non-conventional signalling pathway [49],. Evidence that 0-3hRP and E. coli bioparticles do not co-localise within the same LAMP-1+ phagolysosome, supports the hypothesis that each phagosome is independent of each other [45]. A similar phenomenon occurs in BMDC co-pulsed with SEA (pro-Th2) and Proprionibacterium acnes (pro-Th1) whereby the two stimulants occupy different locations within the same cell and induce contrasting Th subsets [51]. However, increased concentrations of P. acnes are suggested to enhance antigen processing and induce weak Th1-specific SEA specific responses [51]. The delayed transfer of 0-3hRP to the mature phagosome is one possible explanation for the limited maturation phenotype of DC stimulated with 0-3hRP shown by proteomic analysis [16]. Moreover, these DC exhibited limited expression of MHC-II and other co-stimulatory molecules, but were potent inducers of Th2 responses in vitro and in vivo [15].
By infecting the pinnae with CFDA-SE labelled cercariae, we determined that dermal-derived MΦ, DC, and neutrophils phagocytosed labelled ES in vivo. Eosinophils are rare in skin exposed to a single dose of cercariae as in this study but are highly abundant following multiple infections and appear to have an important role in defining an IL-4/IL-13 rich cytokine environment of the infection site (PC Cook & AP Mountford; manuscript in preparation). Neutrophils quickly influx into the infection site [7],[10],[25] and were the most abundant (∼50%) cell type to have phagocytosed CFDA-SE labelled ES at 3 hrs. Neutrophils are an important source of chemokines which attract monocytes, MΦ and DC to the site of infection [52]. Indeed CCL3 and CCL4 are present at increased levels immediately after cercarial penetration [7]. The decline in CFDA-SE+ neutrophils after 24 hrs is likely to reflect rapid degradation of labelled ES material due to their potent proteolytic activity [53] followed by their rapid clearance from the skin; none were observed in the sdLN. The other dominant CFDA-SE+ cell populations in the skin at 3 hrs were MHC-II+ MΦ and DC which each accounted for ∼25% of the total CFDA-SE+ cell population. As MΦ are more populous than DC in naive mouse skin [8], and in our study on infected pinnae are also much more abundant, we appear to show that DC take up labelled material more efficiently than MΦ.
As the numbers of CFDA-SE+ MΦ and DC in the skin peak at 24 hrs but declined thereafter, we infer that both cell types migrate to the sdLN. Indeed, LC emigrate from the epidermis to the sdLN following exposure to schistosome larvae [11], although their migration can be delayed by up to 48 hours in response to parasite-derived prostaglandin D2 [13]. A similar interpretation could be argued for the data presented here as only a very small number of CFDA-SE+ cells were detected in the local sdLN up to 48 hrs. Delayed migration could affect MΦ as well as DC, although it may also reflect differences in the temporal migration rates of the two types of cell. This delayed cell migration could aid parasite escape from the skin.
Although both MФ and DC internalised CFDA-SE labeled 0-3hRP, DC phagocytosed greater amounts of antigen and at a faster rate. DC also expressed higher levels of activation markers, increased levels of IL-6, TNF-α and IL-12p40/23, and had significantly greater expression of iNOS. On the other hand, MФ secreted significantly increased levels of regulatory IL-10 and had far more transcripts for arginase 1. In this context, schistosome larvae are known to induce the production of many different mediators with immunoregulatory function which serve to protect the parasite from immune attack but also to limit damage to the host caused by inflammation [6]. The production of IL-10 by the skin and skin-derived cells in response to schistosomes is critical in limiting IL-12 driven pathology in the skin [7],[12],[13],[54]. Prostaglandin E2 which is released by cercariae upon transformation [54] is a potent inducer of IL-10 secretion from MΦ [55] and could be important in our model. The observation that BMMΦ but not BMDC produce abundant IL-10 in response to cercarial ES may implicate MΦ as the possible source of this cytokine in vivo which in turn could mediate the actions of DC. In fact was recently reported that dermal-derived MΦ in the sdLN can produce IL-10 which directly suppresses the activity of DC [56]. High levels of IL-10 can cause reduced phagosome maturation [57] and may help explain the limited maturation of MФ in response to our ES material. The elevated levels of arginase-1 in our studies are also indicative of the MΦ having an ‘alternatively activated’ phenotype which is a feature of many helminth infections [58]–[60]. The balance of arginase/iNOS production is central in controlling the function of MΦ with arginase countering the pro-inflammatory cascade and production of NO [61]. Arginase-1 production by MΦ is also important in wound healing [62] and is a feature of tissue remodeling after repeated infection of the skin by schistosome cercariae (PC Cook & AP Mountford; manuscript in preparation). Our data here indicate that cercarial ES products directly drive MФ to take on an ‘alternatively activated’ phenotype independent of other host derived immune mediators (e.g. IL-4 and IL-13).
Finally, the increased kinetics of antigen translocation through the endosomal pathway of BMDC is indicative of a higher activation rate and increased activation of these cells [33],[45]. As DCs secrete higher quantities of IL-12 compared to MФ in other infection models [63],[64] and are more potent APC [65], our data suggest that DC favour pro-inflammatory responses but that MФ have the capacity to regulate this response. The greater uptake of ES material by DC relative to MΦ in vitro and the greater proportion of DC that were CFDA-SE+ in the skin is evidence that DC are more important than MΦ as APC. However, as the skin comprises both cell types, the relative abundance of MΦ versus DC within the inflammatory foci which form around schistosome larvae in the skin may explain why there is a balanced immune phenotype of stimulation and regulation [9]. It would be instructive to determine whether MΦ and DC in the skin of schistosome infected mice differ in their expression of various TLRs and C-type lectins that might explain their differential rates of processing of schistosome ES products and thus their function as APC. Manipulation of the skin's immune response to promote the development of anti-parasite immune responses must therefore take account of DC populations to maximise presentation of parasite antigens but also to consider the regulatory role of skin-derived MΦ.
|
10.1371/journal.ppat.1002843 | Bap, a Biofilm Matrix Protein of Staphylococcus aureus Prevents Cellular Internalization through Binding to GP96 Host Receptor | The biofilm matrix, composed of exopolysaccharides, proteins, nucleic acids and lipids, plays a well-known role as a defence structure, protecting bacteria from the host immune system and antimicrobial therapy. However, little is known about its responsibility in the interaction of biofilm cells with host tissues. Staphylococcus aureus, a leading cause of biofilm-associated chronic infections, is able to develop a biofilm built on a proteinaceous Bap-mediated matrix. Here, we used the Bap protein as a model to investigate the role that components of the biofilm matrix play in the interaction of S. aureus with host cells. The results show that Bap promotes the adhesion but prevents the entry of S. aureus into epithelial cells. A broad analysis of potential interaction partners for Bap using ligand overlayer immunoblotting, immunoprecipitation with purified Bap and pull down with intact bacteria, identified a direct binding between Bap and Gp96/GRP94/Hsp90 protein. The interaction of Bap with Gp96 provokes a significant reduction in the capacity of S. aureus to invade epithelial cells by interfering with the fibronectin binding protein invasion pathway. Consistent with these results, Bap deficient bacteria displayed an enhanced capacity to invade mammary gland epithelial cells in a lactating mice mastitis model. Our observations begin to elucidate the mechanisms by which components of the biofilm matrix can facilitate the colonization of host tissues and the establishment of persistent infections.
| Staphylococcus aureus is a pathogen responsible for a wide variety of infections, some of which become chronic due to the capacity of this bacteria to form multicellular communities that grow embedded in a self-produced extracellular matrix, referred to as biofilms. Numerous evidences have demonstrated that growing in the biofilm protects bacteria from the immune system and antimicrobial treatments. However, less attention has been paid to the role that the biofilm extracellular matrix plays in the interaction with host cells. Here, we investigate this issue through the use of the proteinaceous biofilm matrix assembled by the Bap protein as a model. Our results show that the Bap biofilm matrix triggers the adhesion to epithelial cells. After adhesion, Bap binds directly to the host receptor Gp96 and this interaction inhibits the entry of the bacteria into the cells by interfering with the fibronectin-binding protein mediated invasion pathway. As a result, the expression of Bap decreased cell invasion and increased bacterial persistence in lactating mice mammary glands. Thus, our findings revealed a dual role for the Bap‐dependent biofilm matrix during the establishment of persistent infections, promoting adhesion of S. aureus to epithelial cells and impairing host cell invasion.
| Staphylococcus aureus is a regular commensal of the skin of animals and human population and it persistently colonizes the anterior nares of around 25% of human adults [1]. S. aureus is harmless in these locations, but it turns into an extremely threatening pathogen when it traverses the epithelial barrier and gains access to internal tissues from where it can infect almost any organ and cause a broad spectrum of infections including abscesses, pneumonia, endocarditis, osteomyelitis, sepsis and infections associated with foreign-body implants [2]. Once in the internal tissue, S. aureus remains mainly extracellular, in the interstitial space between the cells [3], [4], where bacteria encounter cellular, humoral and complement compounds of the host innate immune system. To succeed in this environment, S. aureus produces a large variety of virulence factors that mediate cell and tissue adhesion (surface proteins), contribute to tissue damage and spreading (proteases, coagulase, DNAse, lipases, toxins) and protect bacteria against the host immune defense system (superantigens) [5], [6]. In some cases, S. aureus proliferates producing bacterial aggregates that grow encased in a self-produced extracellular polymeric matrix, known as a biofilm [7]–[11]. Based on the susceptibility of the biofilm matrix to the disaggregation with glycoside hydrolases (dispersin B), proteases or DNAse, it is recognized that the S. aureus biofilm matrix can be made of exopolysaccharides, proteins and DNA [12]–[17]. The exopolysaccharidic biofilm matrix is composed of a polymer of poly-N-acetyl-β-(1–6)-glucosamine, termed polysaccharide intercellular adhesin (PIA) or poly-N-acetylglucosamine (PNAG) [18]–[21]whereas the proteinaceous biofilm matrix can be assembled with different surface proteins, namely Bap, FnBPs, SasG and Protein A [16], [17], [22]–[26].
The first example of a surface protein able to induce biofilm development was Bap. It is a large protein of 2,276-aminoacids with a series of identical repeats of 86 amino acids that accounts for more than half of the protein [22], and two canonical calcium binding EF-hand motifs able to control Bap functionality in response to the calcium present in the growth media [27]. The bap gene in S. aureus was initially identified in a mobile pathogenicity island (SaPIbov2) whose mobility depends on the activity of a self-encoded recombinase protein [28]. So far, the bap gene has never been found in S. aureus human isolates. However, a bap ortholog gene is present in the core genome of several coagulase-negative staphylococcal species that are frequent colonizers of human skin [29].All the S. aureus strains harbouring the bap gene are strong biofilm producers and Bap-mediated biofilm formation process occurs independently of the presence of the PIA/PNAG exopolysaccharide [30].A particularly interesting issue concerning this Bap assembled matrix is that functionally related proteins homologous to Bap exist in many phylogenetically unrelated bacteria including Enterococcus faecalis, Acinetobacter baumanii, Pseudomonas aeruginosa, Salmonella enteritidis, Lactobacillus reuteri, Bordetella pertussis and Escherichia coli (for a review see [31]). All the Bap homologous proteins show high molecular weight, contain a core domain of repeats and promote bacterial aggregation and biofilm development.
It is currently clear that matrix production and subsequent biofilm formation is very often associated with the establishment of persistent infections due to the highly resistant nature of the embedded bacteria to both the host immune defences and the antimicrobial therapy. In this respect, several studies have demonstrated that the PIA/PNAG matrix aids S. aureus in the evasion of host immune defences by protecting bacteria from macrophage phagocytosis and attenuating host proinflammatory responses [32]–[35].Furthermore, phenol-soluble modulins (PSMs) surfactant peptides secreted to the biofilm matrix of S. aureus act as biofilm structuring factors but also have multiple functions in immune system evasion [35], [36].
In addition to its protecting role, it is feasible to envision that the biofilm matrix may mask important bacterial surface antigens, and in this context, the interaction between bacteria inside the biofilm and the host might rely on the specific binding of extracellular matrix components to host cell receptors. Accordingly, various studies have proposed a role of the Bap-mediated biofilm matrixes, including BapA of S. Enteritidis, Esp of E. faecalis, Lsp of L. reuteri and Bap of A. baumanii, in the adhesion to host cells [37]–[40]. With regard to Bap of S. aureus, we have previously shown that the presence of a Bap-mediated biofilm matrix interferes with the binding of several S. aureus adhesins (fibronectin-binding protein and clumping factor) to their targets (fibrinogen and fibronectin) in host tissues [41]. Despite this masking effect caused by the Bap matrix, S. aureus strains producing Bap display an enhanced capacity to colonize and persist in the mammary gland [30].However, the underlying molecular mechanisms of the interaction between the Bap related proteins and eukaryotic cells remain unknown.
In this study, we used the staphylococcal Bap mediated matrix as a model to investigate the role that components of the biofilm matrix play in the interaction with host cells. Our results revealed that Bap enhances the adhesion but inhibits the entry of S. aureus into the epithelial cells. We also identified a direct interaction between Bap and the Gp96 chaperone protein from host cells. Binding of Bap to Gp96 was responsible for the inhibition of bacterial invasion into nonprofessional phagocytic cells by interfering with the fibronectin-binding protein mediated internalization pathway. Overall, our results reveal new facets of the roles that the biofilm matrix plays during the establishment of persistent infections.
To analyze the involvement of the Bap matrix on the adherence capacity of S. aureus, we tested the ability of S. aureus V329 strain and Δbap mutant to adhere to two different cell lines, a bovine mammary epithelial (MAC-T)and a human hepatocyte (Hep-3B) cell line. The bacterial inoculum used in cell assays came from an overnight culture in which S. aureus V329 strain aggregated at the bottom of the tube and also formed a Bap-dependent biofilm adhered to the glass wall, whereas Δbap mutant grew planktonically (data not shown). The results revealed that the V329 strain adhered 4-fold more efficiently (P<0.05) than the Δbap mutant to both cell lines (Figure 1A). It is worth noting that Bap-negative bacteria also showed a reduce capacity to adhere to a HEK293 cell line, despite the fact that these differences passed unnoticed in a previous study of our group [41].We then compared the adhesion of a natural bap-negative strain, S. aureus Newman, and its isogenic derivative containing a chromosomal copy of the bap gene (S. aureus Newman_Bap) [27]. S. aureus Newman_Bap showed a5 times (P<0.05) higher capacity to adhere to both MAC-T and Hep-3B cell lines than its parental bap-deficient S. aureus Newman strain (Figure 1B). These results indicated that the presence of Bap enhances the capacity of S. aureus to bind to epithelial cells, but they did not answer the question as to whether the Bap protein was sufficient to promote adhesion. To elucidate this point, the Bap protein was expressed in a heterologous surrogate bacterium, Enterococcus faecalis and its adherence capacity was tested. E. faecalis producing the Bap protein adhered significantly more efficiently to both MAC-T and Hep-3B cell lines (P<0.05) than the corresponding wild type strain (Figure 1B). Taken together, these results demonstrated that the Bap protein confers the capacity to adhere to epithelial cells.
Next, taking into account that S. aureus strain V329 invaded human embryonic kidney cells (HEK293)less efficiently than its corresponding Bap-deficient strain [41],we decided to test whether Bap could also block S. aureus entry in our cellular models, MAC-T and Hep-3B cells. Quantification of intracellular bacteria, after invasion assays with S. aureus V329 and Δbap revealed that S. aureus Δbap, despite its deficiency in the adhesion capacity to epithelial cells, was able to invade more efficiently MAC-T and Hep-3B cells than the wild type strain(Figure 1C). To ensure that inhibition of S. aureus entry was due to the presence of Bap rather than to the presence of a biofilm matrix, we compared the invasion capacity of S. aureus V329 and Δbap strains complemented with a plasmid carrying the icaADBC operon (pSC18) as well as S. aureus ISP479r, a strain that constitutively produces large amounts of PIA/PNAG exopolysaccharide, and its isogenic Δica mutant. The results showed that the presence of the PIA/PNAG matrix does not have any effect on the capacity of bacteria to invade epithelial cells (Supplementary Figure S1).
Taken together, these results showed that the Bap-mediated matrix promotes bacterial adhesion to epithelial cells, and on the other hand, interferes with S. aureus cell entry.
We then aimed to identify putative host cell receptors interacting with Bap by means of a ligand overlay approach. MAC-T and Hep-3B total cell extracts were separated by SDS-polyacrylamide gel electrophoresis, transferred onto a nitrocellulose membrane and incubated in the presence of purified recombinant Bap protein containing a 6xHistidine tag replacing the LPXTG motif. Then, proteins that bound specifically to Bap were detected by probing the membrane with an anti-Bap serum. A Coomassie blue stained gel of the total membrane protein profile of each strain is shown for reference (Figure 2A). We focused on Bap-reactive bands that were present in both cell extracts and absent when the membranes were not incubated with the Bap protein (Figure 2B). A prominent band in the range of ∼100 kDa was apparent in both MAC-T and Hep-3B cell extracts. To exclude that the 6xhistidine tag present in the recombinant Bap protein could be responsible for the interaction with the eukaryotic protein, we performed a similar ligand overlayer assay using the unrelated6xHis tagged dispersin protein. As expected, no specific bands reacting with the cell extracts were detected when this protein was used as a bait (data not shown).
To determine the identity of the ∼100 kDa Bap binding protein, the region corresponding to the location of the ∼100 kDa band was excised from a parallel Coomassie stained gel and this sample was subject to trypsin digestion and MALDI-TOF analysis followed by peptide mass fingerprinting that was compared with the human proteome. The results showed that the protein band corresponded to the endoplasmic reticulum chaperone Gp96 (GRP94), a member of the Hsp90 family of molecular chaperones [42], [43].
Although Gp96 is recognized as an endoplasmic reticulum chaperon for Toll-like receptors [44], numerous evidences indicate that it is also expressed on the plasma membrane of different cell types [45]–[55].Thus, we decided to analyze whether Gp96 protein is expressed at the cell surface of MAC-T and Hep-3B cells. As a negative control, we also included in the assay two cell lines (Vero and GPC-16) that have been previously shown to poorly express extracellular Gp96 [52]. Firstly, we confirmed the expression of Gp96 in whole MAC-T and Hep-3B cell extracts by immunoblotting using anti-Gp96 antibodies (data not shown). Secondly, we investigated if Gp96 was localized in the plasma membrane of MAC-T and Hep-3B cells, by means of labelling surface exposed proteins of intact cells using the membrane-impermeable biotinylation reagent sulpho-N-hydroxysuccinimide (NHS) biotin (Pierce). After harvesting the cells, the biotinylated proteins were purified on streptavidine-beads. Upon reduction, the biotinylated proteins were released from the beads and analyzed by western-blot using anti-Gp96 antibodies. The results revealed the presence of Gp96 in the biotinylated protein fraction of MAC-T and Hep-3B cells. Accordingly, a faint band was detected in Vero and GPC-16 cells. In addition, to verify that membrane impermeability was not disrupted during the assay and exclude that cytoplasmic Gp96 could be labeled during the experiment, the biotinylated protein fractions were interrogated using anti-α-catenin antibodies. The absence of α-catenin, an abundant protein in the cytoplasm, confirmed that the biotinylated fraction did not contain cytoplasmic proteins (Figure 3A).
Lastly, we used immunofluorescence staining to localize Gp96 distribution at the cell surface. Nonpermeabilised MAC-T, Hep-3B, Vero and GPC-16 cells were incubated with polyclonal anti-Gp96 followed by incubation with a secondary antibody conjugated with Alexa-488. Then, cells were permeabilized and F-actin was labelled with Alexa-Fluor 647-phalloidin. In agreement with previous results, Gp96 was clearly detected at the cell surface of nonpermeabilized MAC-T and Hep-3B cells whereas it was absent from the cell surface of GPC-16 and Vero cells (Figure 3B).
One caution of the ligand overlay assay is that proteins presented in non-native conformations may interact in artificial ways with a ligand and lead to the detection of “false positive” interactions. Several lines of evidences were used to independently verify that Gp96 is a ligand of Bap. First, we assessed the interaction of purified Bap with recombinant Gp96 by coimmunoprecipitation using anti-Bap polyclonal antibodies, and as a result, we found a band corresponding to Gp96 after immunoprecipitation of the complex (Figure 4A). Second, we analyzed the binding of Gp96 to live, intact S. aureus bacteria. For that, S. aureusV329 (Bap+) and it isogenic S. aureus Δbap mutant were incubated with recombinant Gp96 protein. Binding of Gp96 to S. aureus was detected by western-blot using monoclonalanti-Gp96 antibodies. As shown in figure 4B, the presence of Gp96 was only detected in cell extracts of S. aureus producing Bap. These results confirmed that Gp96 serves as a ligand for native Bap in intact bacteria. Third, the specificity of the Gp96binding to Bap was further validated by the ability of polyclonal anti-Bap antibodies to block the binding of Gp96 to Bap producing bacteria (Figure 4B).Fourth, as Bap homologues have been identified in several staphylococcal species including S. epidermidis, S.chromogenes and S. hyicus [29], we investigated whether Bap homologous proteins were also able to interact with Gp96. For that, live intact S. epidermidis C533, S. hyicus 12and S. chromogenes C483 bacteria were incubated with recombinant Gp96 protein and binding of Gp96 was detected by western-blot using anti-Gp96. As shown in Figure 4C, all the coagulase negative staphylococcal strains producing Bap proteins were capable of pulling down Gp96 indicating that all Bap homologous proteins produced by different coagulase negative staphylococcal species interact with Gp96.
To investigate whether the interaction of Bap with Gp96 was responsible for the Bap-mediated adhesion of S. aureus to epithelial cells, we measured the adhesion of S. aureus to Hep-3B cells after knocking down the Gp96 message by siRNA. The efficiency of Gp96 down regulation was determined by densitometry of Gp96 immunoblots using anti-Gp96 antibodies. As is shown in figure 5A, transfection of Gp96 specific siRNA duplexes consistently resulted in a 90% reduction in Gp96 protein levels compared to the nontransfected cells or to the cells transfected with control siRNA. The specific silencing of Gp96 expression was further confirmed by the lack of effect of Gp96 siRNA on the fibronectin protein. However, inhibition of Gp96 expression had no significant effect on the adhesion of Bap producing bacteria, suggesting that Gp96 was not required for Bap-mediated adhesion (Figure 5B).Additional evidence that Gp96 was not required for S. aureus adhesion to epithelial cells was obtained by assessing the adhesion of S. aureus V329 and Δbap to the Gp96 deficient cell lines, Vero and GPC-16. The results revealed that the V329 strain still adhered significantly more efficiently (P<0.05) than the Δbap mutant to both cell lines, indicating that the interaction of Bap with Gp96 was not required for the Bap-mediated adhesion to epithelial cells (Figure 5C).
We next examined a plausible role of the interaction of Bap with Gp96 in the invasion of S. aureus into epithelial cells. For that, we first determined the entry of S. aureus V329 and Δbap to the Gp96 deficient cell lines, Vero and GPC-16. The results showed no significant differences between bacteria expressing Bap and the Δbap mutant (Figure 6A), suggesting that the presence of Gp96 on the cell membrane is necessary for Bap-mediated inhibition of cell invasion. To confirm this, we carried out invasion assays on Vero cells producing Gp96 from a pcDNA3 vector containing gp96 cDNA [56] (Figure 6B).The presence of Gp96 significantly reduced (P<0.05) the capacity of S. aureus V329 (Bap+) to invade the cells whereas it did not have any significant effect on the entry of Δbap deficient bacteria(Figure 6C).These results confirmed that it is the interaction of Bap with Gp96 and not the fact that Bap producing bacteria are merely coated in a matrix, which reduces the capacity of S. aureus to invade epithelial cells. Finally, we determined the entry of S. aureus in Hep-3B cells after knocking down the expression of Gp96 by siRNA (Figure 6D). The number of intracellular S. aureus V329 increased significantly (P<0.05) when the expression of Gp96 was inhibited(Figure 6E). Taken together, these results indicate that the interaction of Bap with Gp96 inhibits the entry of S. aureus into epithelial cells.
We next investigated the mechanisms by which the Bap-Gp96 interaction was inhibiting the entry of S. aureus into the host cell. To this end, we first analyzed whether Bap-Gp96 interaction was affecting the signaling pathway downstream Gp96.For that, we blocked Gp96activityand therefore the downstream signaling pathway by incubating the cells overnight in the presence of 17-AAG (17-(Allylamino)-17-demethoxydeldanamycin), which binds with high affinity into the ATP binding pocket of Gp96. The results revealed that the presence of 17-AAG (1 µM) did not affect S. aureusV329 entry (Figure 7A). To confirm these results and also to rule out the possibility of an indirect effect of the Gp96 absence on the expression of other Bap receptors we transcomplemented Gp96 deficient cells with purified soluble Gp96 and analyzed the infection rates of S. aureus V329 and Δbap. Addition of soluble Gp96 prior to infection significantly reduced the entry of S. aureus V329 into Vero cells and this reduction was even higher when bacteria were preincubated with recombinant Gp96 before infection (Figure 7B). In contrast, the entry of the Bap deficient strain was not affected by the preincubation with Gp96 (Figure 7B). Together, these results indicate that the inhibition of S. aureus entry caused by Bap-Gp96 interaction is neither due to an interference with the Gp96 signaling pathway nor to a blockage of Bap binding to other cell receptors.
In cell culture models, invasion of non-professional phagocytic cells by S. aureus depends on the presence of FnBPs on the bacterial surface, and fibronectin and integrins on the host cell. FnBPs are important not only for adhesion but also for activating host-cell cytoskeletal remodeling via integrin-coupled signaling [57]–[63].We thus investigated the hypothesis that Bap-Gp96 interaction might be somehow interfering with the FnBPs-mediated invasion process. To this end, we deleted both fnbA and fnbB genes(Δfnb) in both S. aureus V329 and in the Δbap mutant. As shown in figure 7C, the wild type strain and its mutant in fnbAB showed a similar infection rate, whilst deletion of FnBPs counteracted the increased invasion capacity shown by the S. aureus V329 Δbap. To further explore this question, we made use of S. aureus Newman strain that is deficient in the production of FnBPs [64] and consequently presents a very low invasion rate. As expected, complementation of this strain with plasmid pFNBA4 that expresses the fnbA gene significantly enhanced its entry capacity into MAC-T cells. Notably, when the bap gene was expressed from the chromosome of this fnbA complemented strain, the invasion rate decreased significantly (Figure 7D). To confirm the requirement of Gp96 in the masking effect on FnBPs-mediated internalization process activity, similar experiments were carried out in Vero cells transfected with pcDNA3gp96.Again,reductionof FnBPs-mediated invasion of Bap producing bacteria only occurred when Vero cells were producing Gp96 (Figure S2). Taken together, these results suggest that Bap-Gp96 interaction interferes with FnBPs-mediated entry of S. aureus in epithelial cells.
As Bap is a large protein of 2276-amino-acid we wondered whether the interaction of Bap with Gp96 could act as a steric hindrance, limiting the accessibility of FnBPs to its target, fibronectin, and in consequence minimizing cell entry. To further assess this hypothesis we constructed a S. aureus strain that produces a recombinant short Bap protein, containing a single repetition (ΔrepBap), and leading to a Bap derivative about half the size of the wild type protein (Figure 8A). Western-blot analysis using anti-Bap antibodies confirmed that S. aureus ΔrepBap strain produced similar levels of the short Bap protein compared to wildtype strain (Figure 8B). Moreover, the biofilm formed by S. aureus ΔrepBap strain was indistinguishable from that produced by V329 strain indicating that the short Bap variant is functional (Figure 8 C).
Then, we investigated whether ΔrepBap still retained the capacity to interact with Gp96 using the pull-down assay. As shown in figure 8D, Gp96 was pulled down by intact cells producing ΔrepBap as efficiently as bacteria producing wildtype Bap, indicating that a Bap allele containing a single repetition can interact with Gp96. S. aureus ΔrepBap strain was then tested for its ability to invade epithelial cells. The efficiency of entry of S. aureus producing a ΔrepBap-mediated biofilm matrix was significantly higher than that of wildtype bacteria and very similar to that of the Bap deficient strain (P<0.05) (Figure 8E). Together these data indicate that a short version of Bap, although still able to interact with Gp96, is unable to block the infection capacity of wild type bacteria, and thus support the hypothesis that the interaction of Bap with Gp96 might cause a stearic impediment that interferes with the FnBPs binding to fibronectin.
To investigate the relevance of the Bap-dependent biofilm matrix in the prevention of in vivo host cell invasion we used a lactating mouse mastitis model. We first confirmed that Gp96 is expressed in mammary glands of lactating mice by western-blot using anti-Gp96 antibodies (Figure 9A). Two mammary glands (L4 and R4) of a group of 7 lactating mice were inoculated with a bacterial solution containing 106 CFU of S. aureus V329 (Bap+) and bap::tet strains. After 18 h post-infection mammary glands were treated 3 h with a solution of PBS containing gentamicin to remove all the extracellular bacteria. Then, the mammary glands were removed aseptically, homogenized and several dilutions were plated on selective agar. To evaluate the invasion capacity of each strain we used the competition index method. As a previous control, we first verified that the invasion differences detected in-vitro between the wild-type and the bap-deficient strain were maintained when both strains were used to co-infected MAC-T cells (Figure S3).In agreement with in vitro assays, wild type bacteria producing Bap showed a significantly lower capacity to invade the mammary gland cells compared with Bap deficient bacteria (Figure 9B). Next, the experiment was repeated comparing the in vivo invasion capacity of Bap deficient bacteria and the strain expressing a short Bap derivative. These two strains showed a very similar capacity to invade mammary gland cells (Figure 9B). These results again confirmed in vitro results showing that the entry efficiency of S. aureus producing a ΔrepBap-mediated biofilm matrix is equal to that of the Bap deficient strain. Together, these results strongly suggest that full-size Bap acts as anti-invasion factor of the mammary gland in vivo.
Biofilm formation is recognized as causing or exacerbating several S. aureus chronic infections such as osteomyelitis, endocarditis and device related infections. The molecular mechanisms underlying the persistence of biofilm infections have been mainly associated with the protection barrier that the biofilm matrix provides against the host immune response and the antibiotic treatments to the embedded bacteria. For example, exopolysaccharidic PIA/PNAG biofilm matrix protects S. aureus from phagocytosis by polymorphic neutrophils and from antibodies mediated opsonisation and retards the rate of antibiotic penetration enough to induce the expression of genes that mediate resistance [32], [65]–[67].
With the aim to identify new functions for the biofilm matrix that may facilitate the development of persistent infections, we have investigated the role of the proteinaceous Bap-mediated biofilm matrix in the interaction with host cells. Our results provide evidences that the Bap biofilm matrix promotes the adhesion of S. aureus to different types of epithelial cells. For this function, Bap does not require the participation of any other staphylococcal factors, because production of Bap in E. faecalis is sufficient to bestow the capacity to adhere to epithelial cells. Screening for the specific molecular target on the epithelial cells using a ligand immunoblot overlay approach revealed that Bap binds to Gp96, also known as GRP94 or endoplasmin, which is a major chaperon of the lumen of the endoplasmic reticulum (ER) [42], [43], [68]. We initially received this result with caution because, using the same approach, Gp96 had been previously identified as the ligand for Vip, a surface protein of Listeria monocytogenes without homology with Bap [52]. To exclude that Bap might be interacting with Gp96 in an artificial way, confirmatory pull down studies were carried out. The results revealed that (i) recombinant Gp96 was pulled down with native Bap protein anchored to the cell wall of intact bacteria; (ii) recombinant Gp96 was pulled down with soluble Bap and anti-Bap antibodies; and (iii)anti-Bap antibodies were able to inhibit the interaction between Bap and Gp96. Another caution about Gp96 being the ligand for Bap was that Gp96 was initially described as an endoplasmic reticulum (ER) protein based on the presence of the KDEL motif in the carboxy-terminal domain of the protein [47], [69], [70]. However, in agreement with other studies [45]–[50], [52], [54], [55], we have found that Gp96 is not restricted to the ER and is also present on the surface of some epithelial cell lines (Hep-3B and MAC-T).
Given that Bap was interacting with Gp96 and promoting adherence to epithelial cells, we hypothesized that binding of Bap to Gp96 might be responsible for the enhanced adherence of Bap producing bacteria to the surface of the epithelial cells. In contrast to our assumption, we found that depletion of Gp96 expression by siRNA in Hep-3B cells did not reduce the adhesion capacity of Bap positive bacteria. Furthermore, bacteria producing Bap also displayed a higher capacity to adhere to cells that do not produce Gp96 than Bap deficient bacteria, indicating that Bap was interacting with another factor, different from Gp96, to promote the adhesion to the epithelial cells. These results raised the question as to why this receptor was not identified with the ligand overlay experiment. At least two reasons can be envisioned to explain this failure. Either the ligand is not a protein and/or Bap might recognize its ligand only in its native folded structure. Further experiments will be needed to identify this additional cellular ligand.
On the other hand, we have shown that the interaction of Bap with Gp96 inhibits bacterial entry into epithelial cells. Several results support this conclusion: (i) Bap deficient bacteria showed higher levels of invasion than the corresponding wildtype strain in cells producing Gp96; (ii) the invasion differences between Bap positive and negative strains disappeared when invasion was tested inGp96 negative cells; (iii) the entry of Bap positive bacteria decreased in cells overproducing Gp96 and increased in cells depleted of Gp96 by the expression of siRNA; (iv) preincubation of cells or bacteria with soluble Gp96 inhibited the S. aureus entry into the cells. These former results also indicate that Bap does not need to interact with membrane anchored Gp96 to inhibit bacterial entry. How does the interaction between Bap and Gp96 interfere with S. aureus invasion? We initially explored the possibility that the binding of Bap with Gp96 might interfere with the signaling pathway downstream Gp96. Against this hypothesis, we have shown that treatment of MAC-T and Hep-3B cell with 17-AAG, a drug that affects the signaling pathway regulated by Gp96, does not affect the invasion levels of Bap producing bacteria. Furthermore, incubation of Vero cells with soluble Gp96, which is not linked with the cytoplasmic partners, still reduced the capacity of S. aureus to invade the cells. Alternatively, the interaction of Bap with Gp96 could be directly interfering with the recognition of another host ligand. The main known mechanism of S. aureus invasion into the host cell is mediated by the bacterial fibronectin binding proteins, FnBPA and FnBPB and host cell fibronectin andα5β1 integrins [57]–[63].Our results suggest that the interaction of Bap with Gp96 interferes with the FnBPs-fibronectin-integrin invasion pathway because deletion of FnBP proteins counteracted the increased invasion rates of Bap deficient bacteria. In addition, overexpression of FnBPs restored the invasion capacity of Bap negative bacteria but did not change the decreased invasion capacity of Bap producing bacteria. It is worth noting that the interference of the FnBPs invasion pathway depends on the length of Bap, because a short but still functional allele of Bap displayed a significantly lower efficiency than the full-length protein to inhibit bacterial entry.
Gp96 has been reported to be the ligand for various bacterial surface proteins, though the consequences of this interaction seem to be different depending on the bacteria [52], [53], [55], [56], [71]–[73]. The bacterial Outer membrane protein A (OmpA) of Escherichia coli interacts with Ecgp96, an homolog of Gp96 that is highly expressed in brain microvascular endothelial cells during meningitis infection, and induces bacterial invasion [56], [71], [72]. Also, Cabanes et al. [52]identified that the interaction of Vip with Gp96 promotes Listeria monocytogenes invasion of the cell. Gp96 has also been described to interact with other bacterial products such as the exotoxin A (TxA) of Clostridium difficile [74]and the outer membrane vesicles (OMVs) produced by adherent-invasive E. coli (AIEC) promoting bacterial invasion [55]. In contrast, the interaction of the outer membrane porin PorBIA of Neisseria gonorrhoeae with Gp96 inhibits bacterial invasion [53]. Moreover, the interaction of Hsp90 (a chaperone homolog to Gp96) with the trimeric surface protein NadA also interferes with bacterial adhesion and invasion [73]. It is worth noting that OmpA, Vip, PorB and NadA do not show any homology with Bap that would explain the interaction with the same ligand. One may speculate that Gp96 is able to bind unspecifically to these bacterial proteins due to its chaperone structure. However, pull down and ligand overlay assays with bacteria producing two other members of the Bap family (Esp protein of Enterococcus faecalis and BapA from Salmonella Enteritidis) or an unrelated bacterial protein (dispersin)did not show any interaction with Gp96 (data not shown), excluding the hypothesis that Gp96 can interact with any bacterial protein. Also, these results indicate that not all the members of the Bap family can interact with Gp96 protein.
The presence of Bapin S. aureus is specifically enriched, by yet unknown reasons, in mastitis-derived isolates where Bap facilitates the persistence of the bacteria in lactating ewes mammary glands [30]. Bap homologous proteins are also encoded in different coagulase negative staphylococcal species associated with chronic mastitis infections [75]–[77].Gp96 is expressed in the bovine mammary glands during the lactating period (Figure 9A) and its presence in the milk has been suggested to be a host defense mechanism [78]. In the present study, we have found a connection between these two findings using a lactating mice infection model, where the Bap expression had a profound impact on the capacity of the bacteria to adhere and invade the mammary gland epithelial cells. Overall, our results support the view that Bap mediated biofilm development facilitates the formation of bacterial aggregates that survive attached to the epithelial cells of the mammary gland by impairing the bacterial internalization through the interaction with Gp96. In this situation, the Bap biofilm matrix promotes the establishment of long-term persistent infections and mediates immune evasion by masking surface antigens. Whether the proteinaceous Bap-dependent biofilm matrix of different bacterial species is playing a similar role in the interaction with the host cells is worthy of further exploration.
All animal studies were reviewed and approved by the “Comité de Etica, Experimentación Animal y Bioseguridad” of the Universidad Pública de Navarra (approved protocol PI-6/10). Work was carried out at the Centro de Agrobiotecnología building (Idab) under the principles and guidelines described in the “European Directive 86/609/EEC” for the protection of animals used for experimental purposes.
Escherichia coli DH10B was cultured in Luria-Bertani (LB) media or on LB agar (Pronadisa, Spain) with appropriate antibiotics. Staphylococcal strains were grown in Trypticase soy broth (TSB) or Trypticase soy agar (TSA) supplemented with glucose 0.25% (TSB-glu) when indicated. Media were supplemented with the following appropriate antibiotics at indicated concentrations: erythromycin 20 µg ml−1 or 1.5 µg ml−1, chloramphenicol 20 µg ml−1, ampicillin 100 µg ml−1, tetracycline 10 µg ml−1.Bovine mammary gland epithelial cell line (MAC-T) was used because bap gene is frequently present in S. aureus isolates causing mastitis infections. Human hepathoma cell line (Hep-3B, ATCC number CCL-2) was used as an example of any other cell line different from mammary gland epithelial cells. Vero epithelial cells (ATCC number CCL-81) and guinea–pig GPC-16 epithelial cells were used due to their low extracellular Gp96 expression [52]. Cells were maintained in Dulbecco's modified Eagle's medium (DMEM) (Gibco-BRL) supplemented with 10% heat-inactivated fetal bovine serum (Gibco-BRL).
To generate the deletion in the bap gene we amplified by PCR two fragments of 500 bp that flanked the left sequence of the gene using primers Δbap-A (ggatccgacatacattagatatttgg) and Δbap-B (ctcgagcaattttatgacgcactatt) and the right sequence of bap using primers Δbap-C (ctcgagcccattttattattggttctg) and Δbap-D (gaattcgccgaaatgttggccgtattc). Fragments were then fused by ligation into the shuttle vector pMAD, and the resulting plasmid was transformed in V329 strain by electroporation. Allelic exchange in the absence of a selection marker was performed as previously described [27]. FnBPs mutant in V329 and Δbap strains was performed as previously described [17]. To generate the ΔrepBap strain, the bap gene of S. aureus V858 whose bap contains a single C-repeat was amplified using DNA polymerase KOD XL (Merck) with primers BapXho-5 (ctcgagtaaaaaaatttattttgaggtgag) and BapXho-3 (ctcgagctctccacctttgtaagtg). The gene was cloned into the pMAD plasmid and the resulting plasmid was transformed into Δbap strain by electroporation. Allelic exchange in the absence of a selection marker was performed as previously described [27]. ΔrepBap strains were verified using primers Bap-6m (cctatatcgaaggtgtagaattgcac) and Bap-7c (gctgttgaagttaatactgtacctgc).
S. aurues bap::tet strain was constructed using plasmid pJP188 [79].For complementation experiments, the multicopy plasmid pFNBA4 [80] that carries the wild-type fnbA gene of S. aureus8325-4 was used. E. faecalis 23 was complemented with plasmid pBT2-bap.
For Bap purification we constructed a recombinant Bap protein in which the LPXTG motif of Bap was replaced by the 6-histidine tag. For that, a 281 bp fragment upstream the LPXTG was amplified using primers BapEco-A (gaattcaattcaggtgctggagacac) and BapBam-B (ggatcctcagtggtggtggtggtggtgttctggtaattcattttg). A 600 bp fragment downstream Bap was amplified using primers BapBam-C (ggatccatgtttaaattattgtaaat) and BapEco-D (gaattcgccgaaatgttggccgtattc). Fragments were cloned into the EcoRI site of the pMAD plasmid [81]and the resulting plasmid was transformed into V329 strain by electroporation using a previously described protocol [30]. The construction was verified by sequencing the insert. Allelic exchange in the absence of a selection marker was performed as previously described [27]. As the LPXTG motif was replaced by the 6xHis tag, the recombinant Bap-6xHis protein was obtained at the supernatant. One liter of the supernatant was concentrated using centrifugal filter units 10,000 wco (Millipore). Bap-6xHis tagged protein was purified from the supernatant using His GraviTrap affinity columns (GE Healthcare).
MAC-T and Hep-3B epithelial cells were lysed in RIPA-buffer (150 mM NaCl, 50 mM Tris pH7.5, 0.1% SDS, 1% Triton X-100, 0.5% sodium deoxycholate, 1 mM sodiumorthovanadate, 10 mM NaF, β-glycerophosphate 100 mM and protease inhibitor cocktail (Roche)). Lysates were clarified and protein concentration was determined. 40 µg of cell lysates were resolved by 10% sodium dodecyl sulphate (SDS)-polyacrylamide gel electrophoresis. Proteins were transferred onto a nitrocellulose membrane and blocked overnight. The membrane was then incubated with 50 µg/ml of purified Bap, washed and incubated with anti-bap antibodies diluted 1∶20,000 [30]. Alkaline phosphatase-conjugated goat anti-rabbit immunoglobulin G diluted 1∶10,000 was used as secondary antibody. Protein identification was analyzed by MALDI-TOF analysis followed by peptide mass fingerprinting [16].
10 µg of the recombinant Gp96 protein (SPP-766 Stressgen) were mixed with 10 µg of purified Bap. The mix was incubated at 4°C for 2 h in slow agitation. Gp96 was immunoprecipitated using 1.5 µl of anti-Bap antibodies, for 2 h at 4°C and then with 50 µl of Protein G sepharose beads (GE Healthcare). Immunoprecipitated proteins were boiled in Laemmli buffer and analyzed by SDS-polyacrylamide gel electrophoresis, inmmunoblotted with primary antibodies anti-Bap (1∶2500) or anti-Gp96 (1∶1000)monoclonal antibodies (SPA-850 Stressgen)and with secondary antibodies goat anti-rabbit immunoglobulin-G HRP (Thermo) or anti-rat immunoglobulin-G HRP (SAB-200 Stressgen). Pull down experiment using bacteria was performed as follows. S. aureus V329 expressing Bap, Δbap, ΔrepBap and CNS(S. epidermidis C533, S. hyicus 12, and S. chromogenes C483) were grown overnight at 37°C. A volume of cells corresponding to an OD600 value of 5 was centrifuged and washed twice with PBS buffer. Bacteria were then incubated with 5 µg/ml of recombinant Gp96 protein in the absence or in the presence of anti-Bap antibodies for 2 h at 4°C and slow agitation. Unbound Gp96 was removed by washing the bacteria 4 times with 1 ml of PBS buffer. Immunoprecipitated Gp96 was detected using anti-Gp96 antibodies and anti-rat immunoglobulin-G HRP.
Biofilm formation assay in microtiter wells was performed as described [18]. Briefly, strains were grown overnight at 37°C and were diluted 1∶40 in TSB-gluc. Cell suspension was used to inoculate sterile 96-well polystyrene microtiter plates (IWAKI). After 24 hours at 37°C wells were gently rinsed three times with water, dried and stained with 0.1% of crystal violet for 15 min. Colony morphology of S. aureus was analyzed using Congo red agar plates [82]–[84].Congo red agar was prepared as follow: 30 g/l of trypticase soy (Pronadisa), 15 g/l of agar (Pronadisa), 0.8 g/l of Congo red stain (Sigma) and 20 g/l of sucrose. The Congo red stain and the sucrose solution were autoclaved separately (121°C for 20 minutes) and (115°C for 15 minutes) respectively. S. aureus strains were streaked on congo red agar and were incubated at 37°C for 24 hours. Rough colonies are being indicative of biofilm formation.
Adherence and invasion experiments were performed as described previously [57]. Briefly, prior to use, wells were seeded with 0.3×106 cells in 6-well tissue culture plates and 0.5×105 cells in 24-well tissue culture plates. Once cells were confluent (1.2×106 or 0.2 106 cells per well) the culture medium was removed and cells were washed with DMEM plus 10% heat-inactivated fetal bovine serum. For adherence assays, overnight bacterial cultures were mixed vigorously and added to the monolayer cells in a multiplicity of infection of 10 in DMEM. Incubation was carried out 1 hour at 37°C in 5% CO2. To remove non-adherent bacteria, cells were washed three times with sterile PBS. Eukaryotic cells were lysed with 0.1% Triton X-100. Before plating extracts were mixed vigorously by vortexing and sonication. The number of adherent bacteria were determined by serial dilution and plating. For invasion assays, bacteria were added to the monolayer cells in a multiplicity of infection of 40 in DMEM. Incubation was carried out for 1 hour at 37°C in 5% CO2. To kill extracellular bacteria, media was replaced with 2 ml of DMEM containing 50 µg ml−1 of gentamicin (SIGMA) for 2 hour. Cell monolayers were washed three times with sterile PBS and lysed with 0.1% Triton X-100. Before plating extracts were mixed vigorously by vortexing and sonication. The number of intracellular bacteria was determined by serial dilution and plating. Experiments were performed in triplicate.
Cell surface biotinylation was performed using Pierce Cell surface protein isolation kit according to the manufacturer's protocol. Briefly, 4 flasks of 75 cm2 of live confluent cells were incubated with Sulfo-NHS-SS-Biotin for 30 min at 4°C. Sulfo-NHS-SS-Biotin was quenched and biotinylized cells were lyzed. For isolation of labelled proteins lyzed cell were incubated with NeutrAvidin Agarose. Eluted proteins were resolved by 10% sodium dodecyl sulphate (SDS)-polyacrylamide gel electrophoresis. Immunodetection was performed following protein transfer onto nitrocellulose membrane and incubation with anti-Gp96 antibodies. To control membrane impermeability α-catenin was detected using anti-α-catenin antibodies (H-297 Santa Cruz Biotechnologies).
Cells grown on coverslips, fixed with paraformaldehyde 3.5% (SIGMA) and stained with anti-Gp96 (H-212; Santa Cruz Biotech) diluted 1∶100 and stained secondary antibody Alexa Fluor 488-conjugated goat anti-rabbit (Invitrogen). Cells were then permeabilized (0.1% triton X-100 for 5 min in PBS). Alexa-Fluor 647-Phalloidin (Invitrogen) diluted 1∶200 was used to label actin filaments. Preparations were observed with an epifluorescence microscope and images were acquired and analyzed with EZ-C1 software (Nikon).
Transient transfection of pcDNA3-gp96 in Vero cells was performed using Lipofectamine 2000 (Invitrogen) following the manufacturer's protocol. Briefly, lipofectamine was diluted 1∶25 in DMEM media. Diluted lipofectamine was then mixed with 0.8 µg of DNA in a 2∶1 ratio and the mixture was incubated for 20 min at room temperature. After incubation, 100 µl were added to a 24-well culture vessel containing Vero confluent cells. To silence gene expression by siRNA, Hep-3B cells were transfected with Gp96 silencer siRNA (Hs_TRA1_9) and control siRNA (Allstars Negative Control) purchased from Qiagen (Valencia, CA). Transfection was performed using a Ready-to-Use mix from Qiagen as described by the manufacturer. Cells were collected for immunodetection of Gp96 using monoclonal anti-Gp96 antibodies and fibronectin using anti-fibronectin antibodies (SIGMA).Gp96 expression was determined by density measurement of digital images using ImageJ software.
MAC-T and Hep-3B monolayers were treated with 1 µM of 17-(Allylamino)-17-demethoxydeldanamycin (17-AAG) or the vehicle control DMSO, one day before the experiments. Overnight treatments of MAC-T and Hep-3B had no effect on cell morphology. Then, cells were washed and incubated with bacteria for invasion assays. Transcomplemetation of Vero cells were performed as described by Rechner et al., 2007 [53]. Briefly, 0.5×106 Vero cells were incubated with 10 µg/ml of Gp96 for 30 min (cells+Gp96) and infected with S. aureus V329 and Δbap strains (MOI = 40) to perform invasion assays. Additionally, S. aureus V329 and Δbap strains were incubated with 5 µg/ml of recombinant Gp96 protein for 2 h at 4°C. After washing bacteria were used for infection of Vero cells (0.5×106) (bacteria+Gp96).
CD1 mice were maintained in the animal facility of the Institute of Agrobiotechnology, Public University of Navarra. Seven to 10 days after parturition, pups of a group of 7 lactating female mice were removed 2 h before bacterial inoculation. A mixture of ketamine/xylazine was used to anaesthetize lactating mice. 100 µl of a solution containing 106 CFU of a mix of S. aureus bap::tet and V329 strains or a mix of bap::tet and ΔrepBap strains were used to inoculate L4 (on the left) and R4 (on the right) mammary glands. After 18 h post-infection L4 and R4 mammary glands were inoculated with 200 µl of a solution of PBS containing 100 µg ml−1 of gentamicin (SIGMA). After 3 h,L4 and R4 mammary glands were aseptically removed and homogenized. Viable counts were performed on these homogenates by plating the samples on TSA and TSA containing tetracycline. To evaluate the invasion capacity of each strain we used the competition index method. A competition index equal to 1 indicates similar invasiveness for both strains. A competition index significantly greater than 1 indicates a higher invasion capacity of the Bap mutant and a competition index lower than 1 indicates the opposite. Prior to the start of the coinfection assays, a competition experiment was performed with all strains tested to confirm that coincubation of the strains did not affect their growth capacity.
Data corresponding to adhesion and invasion were compared using the Mann-Whitney tests. Competition indexes of wild type-Δbap and Δbap-ΔrepBap were calculated using t-test and statistical differences were determined with the t-test.
|
10.1371/journal.pcbi.1005331 | Scalable Parameter Estimation for Genome-Scale Biochemical Reaction Networks | Mechanistic mathematical modeling of biochemical reaction networks using ordinary differential equation (ODE) models has improved our understanding of small- and medium-scale biological processes. While the same should in principle hold for large- and genome-scale processes, the computational methods for the analysis of ODE models which describe hundreds or thousands of biochemical species and reactions are missing so far. While individual simulations are feasible, the inference of the model parameters from experimental data is computationally too intensive. In this manuscript, we evaluate adjoint sensitivity analysis for parameter estimation in large scale biochemical reaction networks. We present the approach for time-discrete measurement and compare it to state-of-the-art methods used in systems and computational biology. Our comparison reveals a significantly improved computational efficiency and a superior scalability of adjoint sensitivity analysis. The computational complexity is effectively independent of the number of parameters, enabling the analysis of large- and genome-scale models. Our study of a comprehensive kinetic model of ErbB signaling shows that parameter estimation using adjoint sensitivity analysis requires a fraction of the computation time of established methods. The proposed method will facilitate mechanistic modeling of genome-scale cellular processes, as required in the age of omics.
| In this manuscript, we introduce a scalable method for parameter estimation for genome-scale biochemical reaction networks. Mechanistic models for genome-scale biochemical reaction networks describe the behavior of thousands of chemical species using thousands of parameters. Standard methods for parameter estimation are usually computationally intractable at these scales. Adjoint sensitivity based approaches have been suggested to have superior scalability but any rigorous evaluation is lacking. We implement a toolbox for adjoint sensitivity analysis for biochemical reaction network which also supports the import of SBML models. We show by means of a set of benchmark models that adjoint sensitivity based approaches unequivocally outperform standard approaches for large-scale models and that the achieved speedup increases with respect to both the number of parameters and the number of chemical species in the model. This demonstrates the applicability of adjoint sensitivity based approaches to parameter estimation for genome-scale mechanistic model. The MATLAB toolbox implementing the developed methods is available from http://ICB-DCM.github.io/AMICI/.
| In the life sciences, the abundance of experimental data is rapidly increasing due to the advent of novel measurement devices. Genome and transcriptome sequencing, proteomics and metabolomics provide large datasets [1] at a steadily decreasing cost. While these genome-scale datasets allow for a variety of novel insights [2, 3], a mechanistic understanding on the genome scale is limited by the scalability of currently available computational methods.
For small- and medium-scale biochemical reaction networks mechanistic modeling contributed greatly to the comprehension of biological systems [4]. Ordinary differential equation (ODE) models are nowadays widely used and a variety of software tools are available for model development, simulation and statistical inference [5–7]. Despite great advances during the last decade, mechanistic modeling of biological systems using ODEs is still limited to processes with a few dozens biochemical species and a few hundred parameters. For larger models rigorous parameter inference is intractable. Hence, new algorithms are required for massive and complex genomic datasets and the corresponding genome-scale models.
Mechanistic modeling of a genome-scale biochemical reaction network requires the formulation of a mathematical model and the inference of its parameters, e.g. reaction rates, from experimental data. The construction of genome-scale models is mostly based on prior knowledge collected in databases such as KEGG [8], REACTOME [9] and STRING [10]. Based on these databases a series of semi-automatic methods have been developed for the assembly of the reaction graph [11–13] and the derivation of rate laws [14, 15]. As model construction is challenging and as the information available in databases is limited, in general, a collection of candidate models can be constructed to compensate flaws in individual models [16]. For all these model candidates the parameters have to be estimated from experimental data, a challenging and usually ill-posed problem [17].
To determine maximum likelihood (ML) and maximum a posteriori (MAP) estimates for model parameters, high-dimensional nonlinear and non-convex optimization problems have to be solved. The non-convexity of the optimization problem poses challenges, such as local minima, which have to be addressed by the selection of optimization methods. Commonly used global optimization methods are multi-start local optimization [18], evolutionary and genetic algorithms [19], particle swarm optimizers [20], simulated annealing [21] and hybrid optimizers [22, 23] (see [18, 24–26] for a comprehensive survey). For ODE models with a few hundred parameters and state variables multi-start local optimization methods [18] and related hybrid methods [27] have proven to be successful. These optimization methods use the gradient of the objective function to establish fast local convergence. While the convergence of gradient based optimizers can be significantly improved by providing exact gradients (see e.g. [18, 28, 29]), the gradient calculation is often the computationally most demanding step.
The gradient of the objective function is usually approximated by finite differences. As this method is neither numerically robust nor computationally efficient, several parameter estimation toolboxes employ forward sensitivity analysis. This decreases the numerical error and computation time [18]. However, the dimension of the forward sensitivity equations increases linearly with both the number of state variables and parameters, rendering its application for genome-scale models problematic. In other research fields such as mathematics and engineering, adjoint sensitivity analysis is used for parameter estimation in ordinary and partial differential equation models. Adjoint sensitivity analysis is known to be superior to the forward sensitivity analysis when the number of parameters is large [30]. Adjoint sensitivity analysis has been used for inference of biochemical reaction networks [31–33]. However, the methods were never picked up by the systems and computational biology community, supposedly due to the theoretical complexity of adjoint methods, a missing evaluation on a set of benchmark models, and an absence of an easy-to-use toolbox.
In this manuscript, we provide an intuitive description of adjoint sensitivity analysis for parameter estimation in genome-scale biochemical reaction networks. We describe the end value problem for the adjoint state in the case of discrete-time measurement and provide an user-friendly implementation to compute it numerically. The method is evaluated on seven medium- to large-scale models. By using adjoint sensitivity analysis, the computation time for calculating the objective function gradient becomes effectively independent of the number of parameters with respect to which the gradient is evaluated. Furthermore, for large-scale models adjoint sensitivity analysis can be multiple orders of magnitude faster than other gradient calculation methods used in systems biology. The reduction of the time for gradient evaluation is reflected in the computation time of the optimization. This renders parameter estimation for large-scale models feasible on standard computers, as we illustrate for a comprehensive kinetic model of ErbB signaling.
In this section we introduce the model class and the corresponding estimation problem. Subsequently, gradient calculation using finite differences, forward sensitivity analysis and adjoint sensitivity analysis is described and the theoretical complexity as well as some aspects of the numerical implementation are discussed.
We consider ODE models for biochemical reaction networks,
x ˙ = f ( x , θ ) , x ( t 0 ) = x 0 ( θ ) , (1)
in which x ( t , θ ) ∈ R n x is the concentration vector at time t and θ ∈ R n θ denotes the parameter vector. Parameters are usually kinetic constants, such as binding affinities as well as synthesis, degradation and dimerization rates. The vector field f : R n x × R n θ ↦ R n x describes the temporal evolution of the concentration of the biochemical species. The mapping x 0 : R n θ ↦ R n x provides the parameter dependent initial condition at time t0.
As available experimental techniques usually do not provide measurements of the concentration of all biochemical species, we consider the output map h : R n x × R n θ ↦ R n y. This map models the measurement process, i.e. the dependence of the output (or observables) y ( t , θ ) ∈ R n y at time point t on the state variables and the parameters,
y ( t , θ ) = h ( x ( t , θ ) , θ ) . (2)
The i-th observable yi can be the concentration of a particular biochemical species (e.g. yi = xl) as well as a function of several concentrations and parameters (e.g. yi = θm(xl1 + xl2)).
We consider discrete-time, noise corrupted measurements
y ¯ i j = y i ( t j , θ ) + ϵ i j , ϵ i j ∼ N ( 0 , σ i j 2 ) , (3)
yielding the experimental data D = { ( ( y ¯ i j ) i = 1 n y , t j ) } j = 1 N. The number of time points at which measurements have been collected is denoted by N.
Remark: For simplicity of notation we assume throughout the manuscript that the noise variances, σ i j 2, are known and that there are no missing values. However, the methods we will present in the following as well as the respective implementations also work when this is not the case. For details we refer to the S1 Supporting Information.
We estimate the unknown parameter θ from the experimental data D using ML estimation. Parameters are estimated by minimizing the negative log-likelihood, an objective function indicating the difference between experiment and simulation. In the case of independent, normally distributed measurement noise with known variances the objective function is given by
J ( θ ) = 1 2 ∑ i = 1 n y ∑ j = 1 N y ¯ i j - y i ( t j , θ ) σ i j 2 , (4)
where yi(tj, θ) is the value of the output computed from Eqs (1) and (2) for parameter value θ. The minimization,
θ * = arg min θ ∈ Θ J ( θ ) , (5)
of this weighted least squares J yields the ML estimate of the parameters.
The optimization problem Eq (5) is in general nonlinear and non-convex. Thus, the objective function can possess multiple local minima and global optimization strategies need to be used. For ODE models multi-start local optimization has been shown to perform well [18]. In multi-start local optimization, independent local optimization runs are initialized at randomly sampled initial points in parameter space. The individual local optimizations are run until the stopping criteria are met and the results are collected. The collected results are visualized by sorting them according to the final objective function value. This visualization reveals local optima and the size of their basin of attraction. For details we refer to the survey by Raue et al. [18]. In this study, initial points are generated using latin hypercube sampling and local optimization is performed using the interior point and the trust-region-reflective algorithm implemented in the MATLAB function fmincon.m. Gradients are computed using finite differences, forward sensitivity analysis or adjoint sensitivity analysis.
A näive approximation to the gradient of the objective function with respect to θk is obtained by finite differences,
∂ J ∂ θ k ≈ J ( θ + a e k ) - J ( θ - b e k ) a + b , (6)
with a, b ≥ 0 and the kth unit vector ek. In practice forward differences (a = ϵ, b = 0), backward differences (a = 0, b = ϵ) and central differences (a = ϵ, b = ϵ) are widely used. For the computation of forward finite differences, this yields a procedure with three steps:
In theory, forward and backward differences provide approximations of order ϵ while central differences provide more accurate approximations of order ϵ2, provided that J is sufficiently smooth. In practice the optimal choice of a and b depends on the accuracy of the numerical integration [18]. If the integration accuracy is high, an accurate approximation of the gradient can be achieved using a, b ≪ 1. For lower integration accuracies, larger values of a and b usually yield better approximations. A good choice of a and b is typically not clear a priori (cf. [34] and the references therein).
The computational complexity of evaluating gradients using finite differences is affine linear in the number of parameters. Forward and backward differences require in total nθ + 1 function evaluations. Central differences require in total 2nθ function evaluations. As already a single simulation of a large-scale model is time-consuming, the gradient calculation using finite differences can be limiting.
State-of-the-art systems biology toolboxes, such as the MATLAB toolbox Data2Dynamics [7], use forward sensitivity analysis for gradient evaluation. The gradient of the objective function is
∂ J ∂ θ k = ∑ i = 1 n y ∑ j = 1 N y ¯ i j - y i ( t j , θ ) σ i j 2 s i , k y ( t j ) , (7)
with s i , k y ( t ) : [ t 0 , t N ] ↦ R denoting the sensitivity of output yi at time point t with respect to parameter θk. Governing equations for the sensitivities are obtained by differentiating Eqs (1) and (2) with respect to θk and reordering the derivatives. This yields
s ˙ k x = ∂ f ∂ x s k x + ∂ f ∂ θ k , s k x ( t 0 ) = ∂ x 0 ∂ θ k s i , k y = ∂ h i ∂ x s k x + ∂ h i ∂ θ k (8)
with s k x ( t ) : [ t 0 , t N ] ↦ R n x denoting the sensitivity of the state x with respect to θk. Note that here and in the following, the dependencies of f, h, x0 and their (partial) derivatives on t, x and θ are not stated explicitly but have the to be assumed. For a more detailed presentation we refer to the S1 Supporting Information Section 1.
Forward sensitivity analysis consists of three steps:
Step 1 and 2 are often combined, which enables simultaneous error control and the reuse of the Jacobian [30]. The simultaneous error control allows for the calculation of accurate and reliable gradients. The reuse of the Jacobian improves the computational efficiency.
The number of state and output sensitivities increases linearly with the number of parameters. While this is unproblematic for small- and medium-sized models, solving forward sensitivity equations for systems with several thousand state variable bears technical challenges. Code compilation can take multiple hours and require more memory than what is available on standard machines. Furthermore, while forward sensitivity analysis is usually faster than finite differences, in practice the complexity still increases roughly linearly with the number of parameters.
In the numerics community, adjoint sensitivity analysis is frequently used to compute the gradients of a functional with respect to the parameters if the function depends on the solution of a differential equation [35]. In contrast to forward sensitivity analysis, adjoint sensitivity analysis does not rely on the state sensitivities s k x ( t ) but on the adjoint state p(t).
The calculation of the objective function gradient using adjoint sensitivity analysis consists of three steps:
Step 1 and 2, which are usually the computationally intensive steps, are independent of the parameter dimension. The complexity of Step 3 increases linearly with the number of parameters, yet the computation time required for this step is typically negligible.
The calculation of state and output trajectories (Step 1) is standard and does not require special methods. The non-trivial element in adjoint sensitivity analysis is the calculation of the adjoint state p ( t ) ∈ R n x (Step 2). For discrete-time measurements—the usual case in systems and computational biology—the adjoint state is piece-wise continuous in time and defined by a sequence of backward differential equations. For t > tN, the adjoint state is zero, p(t) = 0. Starting from this end value the trajectory of the adjoint state is calculated backwards in time, from the last measurement t = tN to the initial time t = t0. At the time points at which measurements have been collected, tN, …, t1, the adjoint state is reinitialised as
p ( t j ) = lim t → t j + p ( t ) + ∑ i = 1 n y ∂ h i ∂ x T y ¯ i j - y i ( t j ) σ i j 2 , (9)
which usually results in a discontinuity of p(t) at tj. Starting from the end value p(tj) as defined in Eq (9) the adjoint state evolves backwards in time until the next measurement point tj−1 or the initial time t0 is reached. This evolution is governed by the time-dependent linear ODE
p ˙ = - ∂ f ∂ x T p . (10)
The repeated evaluation of Eqs (9) and (10) until t = t0 yields the trajectory of the adjoint state. Given this trajectory, the gradient of the objective function with respect to the individual parameters is
∂ J ∂ θ k = - ∫ t 0 t N p T ∂ f ∂ θ k d t - ∑ i , j ∂ h i ∂ θ k y ¯ i j - y i ( t j ) σ i j 2 - p ( t 0 ) T ∂ x 0 ∂ θ k . (11)
Accordingly, the availability of the adjoint state simplifies the calculation of the objective function to nθ one-dimensional integration problems over short time intervals whose union is the total time interval [t0, tN].
Algorithm 1: Gradient evaluation using adjoint sensitivity analysis
% State and output
Step 1 Compute state and output trajectories using Eqs (1) and (2).
% Adjoint state
Step 2.1 Set end value for adjoint state, ∀t > tN: p(t) = 0.
for j = N to 1 do
Step 2.2 Compute end value for adjoint state according to the jth measurement using Eq (9).
Step 2.3 Compute trajectory of adjoint state on time interval t = (tj−1, tj] by solving Eq (10).
end
% Objective function gradient
for k = 1 to nθ do
Step 3 Evaluation of the sensitivity ∂J/∂θk using Eq (11).
end
Pseudo-code for the calculation of the adjoint state and the objective function gradient is provided in Algorithm 1. We note that in order to use standard ODE solvers the end value problem Eq (10) can be transformed in an initial value problem by applying the time transformation τ = tN − t. The derivation of the adjoint sensitivities for discrete-time measurements is provided in the S1 Supporting Information Section 1.
The key difference of the adjoint compared to the forward sensitivity analysis is that the derivatives of the state and the output trajectory with respect to the parameters are not explicitly calculated. Instead, the sensitivity of the objective function is directly computed. This results in practice in a computation time of the gradient which is almost independent of the number of parameters. A visual summary of the different sensitivity analysis methods is provided in Fig 1. Besides the procedures also the computational complexity is indicated.
The implementation of adjoint sensitivity analysis is non-trivial and error-prone. To render this method available to the systems and computational biology community, we implemented the Advanced Matlab Interface for CVODES and IDAS (AMICI). This toolbox allows for a simple symbolic definition of ODE models (1) and (2) as well as the automatic generation of native C code for efficient numerical simulation. The compiled binaries can be executed from MATLAB for the numerical evaluation of the model and the objective function gradient. Internally, the SUNDIALS solvers suite is employed [30], which offers a broad spectrum of state-of-the-art numerical integration of differential equations. In addition to the standard functionality of SUNDIALS, our implementation allows for parameter and state dependent discontinuities. The toolbox and a detailed documentation can be downloaded from http://ICB-DCM.github.io/AMICI/.
In the following, we will illustrate the properties of adjoint sensitivity analysis for biochemical reaction networks. For this purpose, we study several models provided in the BioPreDyn benchmark suite [27] and from the curated branch of the Biomodels Database [37]. We compare adjoint sensitivity analysis with forward sensitivity analysis and finite differences regarding accuracy, computational efficiency and scalability for a set of medium- to large-scale models.
For the comparison of different gradient calculation methods, we consider a set of standard models from the Biomodels Database [37] and the BioPreDyn benchmark suite [27]. From the biomodels database we considered models for the regulation of insulin signaling by oxidative stress (BM1) [38], the sea urchin endomesoderm network (BM2) [39], and the ErbB sigaling pathway (BM3) [40]. From BioPreDyn benchmark suite we considered models for central carbon metabolism in E. coli (B2) [41], enzymatic and transcriptional regulation of carbon metabolism in E. coli (B3) [42], metabolism of CHO cells (B4) [43], and signaling downstream of EGF and TNF (B5) [44]. Genome-wide kinetic metabolic models of S. cerevisiae and E.coli (B1) [45] contained in the BioPreDyn benchmark suite and the Biomodels Database [15, 45] were disregarded due to previously reported numerical problems [27, 45]. The considered models possess 18-500 state variable and 86-1801 parameters. A comprehensive summary regarding the investigated models is provided in Table 1.
To obtain realistic simulation times for adjoint sensitivities realistic experimental data is necessary (see S1 Supporting Information Section 3). For the BioPreDyn models we used the data provided in the suite, for the ErbB signaling pathway we used the experimental data provided in the original publication and for the remaining models we generated synthetic data using the nominal parameter provided in the SBML definition.
In the following, we will compare the performance of forward and adjoint sensitivities for these models. As the model of ErbB signaling has the largest number of state variables and is of high practical interest in the context of cancer research, we will analyze the scalability of finite differences and forward and adjoint sensitivity analysis for this model in greater detail. Moreover, we will compare the computational efficiency of forward and adjoint sensitivity analysis for parameter estimation for the model of ErbB signaling.
The evaluation of the objective function gradient is the computationally demanding step in deterministic local optimization. For this reason, we compared the computation time for finite differences, forward sensitivity analysis and adjoint sensitivity analysis and studied the scalability of these approaches at the nominal parameter θ0 which was provided in the SBML definitions of the investigated models.
For the comprehensive model of ErbB signaling we found that the computation times for finite differences and forward sensitivity analysis behave similarly (Fig 2a). As predicted by the theory, for both methods the computation time increased linearly with the number of parameters. Still, forward sensitivities are computationally more efficient than finite differences, as reported in previous studies [18].
Adjoint sensitivity analysis requires the solution to the adjoint problem, independent of the number of parameters. For the considered model, solving the adjoint problem a single time takes roughly 2-3-times longer than solving the forward problem. Accordingly, adjoint sensitivity analysis with respect to a small number of parameter is disadvantageous. However, adjoint sensitivity analysis scales better than forward sensitivity analysis and finite differences. Indeed, the computation time for adjoint sensitivity analysis is almost independent of the number of parameters. While computing the sensitivity with respect to a single parameter takes on average 10.09 seconds, computing the sensitivity with respect to all 219 parameters takes merely 14.32 seconds. We observe an average increase of 1.9 ⋅ 10−2 seconds per additional parameter for adjoint sensitivity analysis which is significantly lower than the expected 3.24 seconds for forward sensitivity analysis and 4.72 seconds for finite differences. If the sensitivities with respect to more than 4 parameters are required, adjoint sensitivity analysis outperforms both forward sensitivity analysis and finite differences. For 219 parameters, adjoint sensitivity analysis is 48-times faster than forward sensitivities and 72-times faster than finite differences.
To ensure that the observed speedup is not unique to the model of ErbB signaling (BM3) we also evaluated the speedup of adjoint sensitivity analysis over forward sensitivity analysis on models B2-5 and BM1-2. The results are presented in Fig 2b and 2c. We find that for all models, but model B3, gradient calculation using adjoint sensitivity is computationally more efficient than gradient calculation using forward sensitivities (speedup > 1). For model B3 the backwards integration required a much higher number of integration steps (4 ⋅ 106) than the forward integration (6 ⋅ 103), which results to a poor performance of the adjoint method. One reason for this poor performance could be that, in contrast to other models, the right hand side of the differential equation of model B3 consists almost exclusively of non-linear, non-mass-action terms.
Excluding model B3 we find an polynomial increase in the speedup with respect to the number of parameters nθ (Fig 2b), as predicted by theory. Moreover, we find that the product nθ ⋅ nx, which corresponds to the size of the system of forward sensitivity equations, is an even better predictor (R2 = 0.99) than nθ alone (R2 = 0.83). This suggest that adjoint sensitivity analysis is not only beneficial for systems with a large number of parameters, but can also be beneficial for systems with a large number of state variables. As we are not aware of any similar observations in the mathematics or engineering community, this could be due to the structure of biological reaction networks.
Our results suggest that adjoint sensitivity analysis is an excellent candidate for parameter estimation in large-scale models as it provides good scaling with respect to both, the number of parameters and the number of state variables.
Efficient local optimization requires accurate and robust gradient evaluation [18]. To assess the accuracy of the gradient computed using adjoint sensitivity analysis, we compared this gradient to the gradients computed via finite differences and forward sensitivity analysis. Fig 3 visualizes the results for the model of ErbB signaling (BM3) at the nominal parameter θ0 which was provided in the SBML definition. The results are similar for other starting points.
The comparison of the gradients obtained using finite differences and adjoint sensitivity analysis revealed small discrepancies (Fig 3a). The median relative difference (as defined in S1 Supporting Information Section 2) between finite differences and adjoint sensitivity analysis is 1.5 ⋅ 10−3. For parameters θk to which the objective function J was relatively insensitive, ∂J/∂θk < 10−2, there are much higher discrepancies, up to a relative error of 2.9 ⋅ 103.
Forward and adjoint sensitivity analysis yielded almost identical gradient elements over several orders of magnitude (Fig 3b). This was expected as both forward and adjoint sensitivity analysis exploit error-controlled numerical integration for the sensitivities. To assess numerical robustness of adjoint sensitivity analysis, we also compared the results obtained for high and low integration accuracies (Fig 3c). For both comparisons we found the similar median relative and maximum relative error, namely 2.6 ⋅ 10−6 and 9.3 ⋅ 10−4. This underlines the robustness of the sensitivitity based methods and ensures that differences observed in Fig 3a indeed originate from the inaccuracy of finite differences.
Our results demonstrate that adjoint sensitivity analysis provides objective function gradients which are as accurate and robust as those obtained using forward sensitivity analysis.
As adjoint sensitivity analysis provides accurate gradients for a significantly reduced computational cost, this can boost the performance of a variety of optimization methods. Yet, in contrast to forward sensitivity analysis, adjoint sensitivities do not yield sensitivities of observables and it is thus not possible to approximate the Hessian of the objective function via the Fisher Information Matrix [46]. This prohibits the use of possibly more efficient Newton-type algorithms which exploit second order information. Therefore, adjoint sensitivities are limited to quasi-Newton type optimization algorithms, e.g. the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm [47, 48], for which the Hessian is iteratively approximated from the gradient during optimization. In principle, the exact calculation of the Hessian and Hessian-Vector products is possible via second order forward and adjoint sensitivity analysis [49, 50], which possess similar scaling properties as the first order methods. However, both forward and adjoint approaches come at an additional cost and are thus not considered in this study.
To assess whether the use of adjoint sensitivities for optimization is still viable, we compared the performance of the interior point algorithm using adjoint sensitivity analysis with the BFGS approximation of the Hessian to the performance of the trust-region reflective algorithm using forward sensitivity analysis with Fisher Information Matrix as approximation of the Hessian. For both algorithms we used the MATLAB implementation in fmincon.m. The employed setup of the trust-region algorithm is equivalent to the use of lsqnonlin.m which is the default optimization algorithm in the MATLAB toolbox Data2Dynamics [7], which was employed to win several DREAM challenges. For the considered model the computation time of forward sensitivities is comparable in Data2Dynamics and AMICI. Therefore, we expect that Data2Dynamics would perform similar to the trust-region reflective algorithm coupled to forward sensitivity analysis.
We evaluated the performance for the model of ErbB signaling based on 100 multi-starts which were initialized at the same initial points for both optimization methods. For 41 out of 100 initial points the gradient could not be evaluated due numerical problems. These optimization runs are omitted in all further analysis. To limit the expected computation to a bearable amount we allowed a maximum of 10 iterations for the forward sensitivity approach and 500 iterations for the adjoint sensitivity approach. As the previously observed speedup in gradient computation was roughly 48 fold, we expected this setup should yield similar computation times for both approaches.
We found that for the considered number of iterations, both approaches perform similar in terms of objective function value compared across iterations (Fig 4a and 4b). However, the computational cost of one iteration was much cheaper for the optimizer using adjoint sensitivity analysis. Accordingly, given a fixed computation time the interior-point method using adjoint sensitivities outperforms the trust-region method employing forward sensitivities and the FIM (Fig 4c and 4d). In the allowed computation time, the interior point algorithm using adjoint sensitivities could reduce the objective function by up to two orders of magnitude (Fig 4c). This was possible although many model parameters seem to be non-identifiable (see S1 Supporting Information Section 4), which can cause problems.
To quantify the speedup of the optimization using adjoint sensitivity analysis over the optimization using forward sensitivity analysis, we performed a pairwise comparison of the minimal time required by the adjoint sensitivity approach to reach the final objective function value of the forward sensitivity approach for the individual points (Fig 4e). The median speedup achieved across all multi-starts was 54 (Fig 4f), which was similar to the 48 fold speedup achieved in the gradient computation. The availability of the Fisher Information Matrix for forward sensitivities did not compensate for the significantly reduced computation time achieved using adjoint sensitivity analysis. This could be due to the fact that adjoint sensitivity based approach, being able to carry out many iterations in a short time-frame, can build a reasonable approximation of the Hessian approximation relatively fast.
In summary, this application demonstrates the applicability of adjoint sensitivity analysis for parameter estimation in large-scale biochemical reaction networks. Possessing similar accuracy as forward sensitivities, the scalability is improved which results in an increased optimizer efficiency. For the model of ErbB signaling, optimization using adjoint sensitivity analysis outperformed optimization using forward sensitivity analysis.
Mechanistic mathematical modeling at the genome scale is an important step towards a holistic understanding of biological processes. To enable modeling at this scale, scalable computational methods are required which are applicable to networks with thousands of compounds. In this manuscript, we present a gradient computation method which meets this requirement and which renders parameter estimation for large-scale models significantly more efficient. Adjoint sensitivity analysis, which is extensively used in other research fields, is a powerful tool for estimating parameters of large-scale ODE models of biochemical reaction networks.
Our study of several benchmark models with up to 500 state variables and up to 1801 parameters demonstrated that adjoint sensitivity analysis provides accurate gradients in a computation time which is much lower than for established methods and effectively independent of the number of parameters. To achieve this, the adjoint state is computed using a piece-wise continuous backward differential equation. This backward differential equation has the same dimension as the original model, yet the computation time required to solve it usually is slightly larger. As a result, finite differences and forward sensitivity analysis might be more efficient if the sensitivities with respect to a few parameters are required. The same holds for alternatives like complex-step derivative approximation techniques [51] and forward-mode automatic differentiation [28, 52]. For systems with many parameters, adjoint sensitivity analysis is advantageous. A scalable alternative might be reverse-mode automatic differentiation [28, 53], which remains to be evaluated for the considered class of problems.
For the model of ErbB signaling we could show that adjoint sensitivity based optimization outperforms forward sensitivity based optimization, which is the standard in most systems biology toolboxes. With the availability of the MATLAB toolbox AMICI the adjoint sensitivity based approach becomes accessible for other researchers. AMICI allows for the fully automated generation of executables for adjoint or forward sensitivity analysis from symbolic model definitions. This way, the toolbox is easy-to-use and can easily be integrated with existing toolboxes. Also other MATLAB toolboxes for computational modeling, e.g. AMIGO [6], Data2Dynamics [7], MEIGO [54] and SBtoolbox2 [55] could be extended to exploit adjoint sensitivity analysis. In addition to adjoint sensitivity analysis, these MATLAB toolboxes could exploit forward sensitivity analysis available via AMICI, as AMICI yields computation times comparable to those of tailored numerical methods such as odeSD [56] (S1 Supporting Information Section 5) or Data2Dynamics [7]. Moreover AMICI comes with detailed documentation and is already now used by several research labs.
Our study of the model of ErbB signaling suggests that for the available data, a large number of parameters remains non-identifiable. While novel technologies provide rich dataset, we expect that non-identifiability will remain a problem. In particular if merely relative measurements are available, as the case for many measurement techniques, additional unknown scaling factors need to be introduced. These scaling factors are, in combination with initial conditions and total abundances, often the source of practical and structural non-identifiabilites [18]. Fortunately, for a broad range of biological questions, these information are not necessary and also state-of-the-art methods optimization seem to work reasonably well in the presence of non-identifiabilities. For the considered model of EreB signaling, we were able to achieve a significant decrease in the objective function value, despite the non-identifiability of parameters. This demonstrates that gradient based optimization is still feasible for large-scale problems. Yet, we believe that convergence of the optimizer could be improved by regularizing the objective function by integrating prior knowledge, possibly in a Bayesian framework [57], from databases such as SABIO-RK [58] or BRENDA [59].
Beyond the use in optimization, gradients computed using adjoint sensitivity analysis will also facilitate the development of more efficient uncertainty analysis methods. Riemann manifold Langevin and Hamiltonian Monte Carlo methods [60, 61] exploit the first and second order local structure of the posterior distribution and profit from more efficient gradient evaluation. The same holds for novel emulator-based sampling procedures [62] and approaches for posterior approximation [63]. By exploiting the proposed approach, rigorous Bayesian parameter estimation for models with hundreds of parameters could become a standard tool instead of an exception [64, 65].
In conclusion, adjoint sensitivity analysis will facilitate the development of large- and genome-scale mechanistic models for cellular processes as well as other (multi-scale) biological processes [66]. This will complement available statistical analysis methods for omics data [67] by providing mechanistic insights and render a holistic understanding feasible.
|
10.1371/journal.pgen.1007789 | Arabidopsis class I formins control membrane-originated actin polymerization at pollen tube tips | A population of dynamic apical actin filaments is required for rapid polarized pollen tube growth. However, the cellular mechanisms driving their assembly remain incompletely understood. It was postulated that formin is a major player in nucleating apical actin assembly, but direct genetic and cytological evidence remains to be firmly established. Here we found that both Arabidopsis formin 3 (AtFH3) and formin 5 (AtFH5) are involved in the regulation of apical actin polymerization and actin array construction in pollen tubes, with AtFH3 playing a more dominant role. We found that both formins have plasma membrane (PM) localization signals but exhibit distinct PM localization patterns in the pollen tube, and loss of their function reduces the amount of apical actin filaments. Live-cell imaging revealed that the reduction in filamentous actin is very likely due to the decrease in filament elongation. Furthermore, we found that the rate of tip-directed vesicle transport is reduced and the pattern of apical vesicle accumulation is altered in formin loss-of-function mutant pollen tubes, which explains to some extent the reduction in pollen tube elongation. Thus, we provide direct genetic and cytological evidence showing that formin is an important player in nucleating actin assembly from the PM at pollen tube tips.
| Actin polymerization has been implicated in the regulation of rapid polarized pollen tube growth. The important role of actin polymerization is well appreciated, but the mechanisms that regulate rapid actin polymerization in pollen tubes remain incompletely understood. It was postulated that one of the major actin polymerization pathways in pollen tubes involves formin/profilin modules. However, direct genetic and cytological evidence is still required to support the role of formin in this framework. Using state-of-the-art live-cell imaging in combination with reverse genetic approaches, we demonstrate here that two class I formins, Arabidopsis formin 3 (AtFH3) and formin 5 (AtFH5), are involved in the regulation of apical actin polymerization and actin array construction in pollen tubes. In support of the role of AtFH3 and AtFH5 in regulating membrane-originated apical actin polymerization, we found that both of them are localized to the plasma membrane (PM) at pollen tube tips. Live-cell imaging revealed that the reduction in filamentous actin is very likely due to the decrease in elongation of actin filaments originating from the apical membrane. We also found that AtFH3 and AtFH5 exhibit distinct PM localization patterns in the pollen tube, suggesting that they might have distinct roles in regulating actin polymerization in pollen tubes. Our study provides direct genetic and cytological evidence that formins act as important players in regulating apical actin assembly in pollen tubes.
| The pollen tube is the passage for two non-motile sperm cells and its proper growth is essential for successful reproduction in flowering plants [1–3]. Pollen tube growth is tightly regulated, and this raises many fascinating questions. Numerous studies suggest that the actin cytoskeleton is the core of the regulatory network of pollen tube growth, presumably by coordinating with various cellular events, such as the trafficking, docking and fusion of vesicles and the construction of the cell wall [4–7]. Actin filaments are arranged into distinct structures within different regions of growing pollen tubes, and these structures carry out distinct cellular functions [8–13]. Highly dynamic actin filaments within the apical region were demonstrated to be directly associated with the growth and turning of pollen tubes [14–16]. To date, however, we still have an incomplete understanding of how those apical actin filaments are constantly generated in pollen tubes.
An essential step of actin polymerization is actin nucleation, which is controlled by various actin nucleation factors in cells. Among the actin nucleation factors identified in the literature [17], the formins and the Arp2/3 complex are found in plants [18, 19]. The formin proteins are characterized by the presence of two formin homology (FH) domains, FH1 and FH2, which are capable of nucleating actin assembly from actin or actin bound to profilin [18, 20–22]. Plant formins are categorized into two classes, designated as class I and class II. Class I formins have a transmembrane (TM) domain at their N-terminus followed by the C-terminal FH1 and FH2 domains, whereas class II formins do not have an N-terminal TM domain but carry an N-terminal phosphatase and tensin-related (PTEN)-like domain besides the conserved FH1 and FH2 domains [18, 19, 22–24]. In vitro biochemical analyses showed that most plant formins have the characteristic formin-mediated actin nucleating and barbed end capping and elongating activities [25–31]. Some plant formins were shown to have actin filament bundling [27–29, 32, 33] and microtubule interacting activities [29, 32–34], though the details of the mechanisms underlying these properties may vary between the different proteins.
As important regulators of actin dynamics, the plant formins have been implicated in numerous physiological cellular processes, such as epidermal pavement cell morphogenesis [35], cell division [32], cytokinesis [25], cell-to-cell trafficking [36], and interaction with pathogens [37], as well as the response to auxin signaling [38]. In particular, the formins have been implicated in polarized root hair growth [39–42] and pollen tube growth [4, 31, 43]. Specifically, after characterizing the cellular functions of the Arabidopsis formin gene AtFH5, Cheung et al. [4] proposed that formin nucleates actin assembly from the membrane for the construction of the subapical actin structure. In line with this finding, a recent report showed that the pollen-specific Lilium longiflorum Formin 1 (LiFH1) is involved in constructing the actin fringe structure [43]. However, considering that LIFH1 also has actin filament-bundling activity and was proposed to nucleate actin filaments from the surface of LiFH1-localized vesicles, more work is needed to understand how the formins fit into the apical actin polymerization pathway in general. Nonetheless, it was proposed that actin assembly mediated by formin-profilin modules may be a major pathway for actin polymerization from the apical membrane in the pollen tube [44]. The notion of formin acting as the major player in nucleating actin assembly is actually consistent with the scenario in which actin monomers are predicted to be buffered by an equal amount of profilin in pollen [45–47]. Considering that multiple formin isovariants exist in pollen, it is important to carefully document their precise intracellular localization and dynamics, their mechanism of action and their functional coordination.
Here, we showed that two class I formins, AtFH3 and AtFH5, localize to endomembrane systems and the plasma membrane (PM). However, they have distinct PM distribution patterns: AtFH3 is localized evenly throughout the entire pollen tube while AtFH5 is concentrated at pollen tube tips. We demonstrated that both AtFH3 and AtFH5 are involved in the regulation of membrane-originated actin polymerization within the growth domain of the pollen tube and they have overlapping function in this aspect. Loss of function of AtFH3 and AtFH5 reduces the velocity of tip-directed vesicle transport and alters the apical vesicle accumulation pattern in the pollen tube, further supporting the active role of apical actin filaments in regulating vesicle traffic. Thus, we provide strong evidence that class I formins control membrane-originated actin polymerization to enable the construction of the apical actin structure in the pollen tube.
We previously showed that RNAi-mediated downregulation of AtFH3 impairs the formation of shank-oriented longitudinal actin cables in the pollen tube [31]. However, more work is required to determine whether and how AtFH3 may be involved in the regulation of actin polymerization within the apical region of the pollen tube. To better understand the mechanism of action of AtFH3 in regulating actin polymerization in pollen cells, we sought to analyze stable T-DNA insertion mutants of AtFH3. In addition, considering that AtFH5 was previously shown to nucleate actin assembly from the plasma membrane for the construction of the subapical actin structure within the apical dome of the pollen tube [4], we also sought to determine whether there is functional coordination of AtFH3 with AtFH5 in regulating apical actin polymerization in pollen tubes. To this end, we analyzed T-DNA insertion mutants for AtFH3 (fh3-1 and fh3-2) and AtFH5 (fh5-2 and fh5-3) as well as the double mutants (fh3-1 fh5-2; fh3-2 fh5-3) (Fig 1A). The results showed that fh3-1, fh3-2, fh5-2 and fh5-3 are knockout alleles (Fig 1B–1E). We found that the pollen germination percentage was slightly but significantly reduced in pollen derived from fh3-1, fh3-2, fh5-3, fh3-1 fh5-2 and fh3-2 fh5-3 mutant plants (Fig 1F and 1G). In addition, we found that formin loss-of-function mutant pollen tubes grew significantly more slowly than WT pollen tubes (Fig 1H). Furthermore, we found that the width of pollen tubes was increased slightly but significantly in fh3-1 fh5-2 and fh3-2 fh5-3 double mutants compared to WT (Fig 1I). Interestingly, we found that growing fh3-1 fh5-2 pollen tubes were more curved than WT, fh3-1 or fh5-2 pollen tubes (Fig 1J). This was supported by measurements showing that the ratio of the actual length of the pollen tube to the linear length was significantly increased in fh3-1 fh5-2 pollen tubes (Fig 1K). Thus, these data suggest that AtFH3 and AtFH5 are required for normal polarized pollen tube growth, and AtFH3 has a more dominant role in this process.
We next examined the organization of the actin cytoskeleton in WT and formin mutant pollen tubes using fluorescently labeled phalloidin. We found that the fluorescent signal from apical actin filaments was weaker in fh3, fh5 and fh3 fh5 mutant pollen tubes than in WT (Figs 2A and 2B and S1A). This is consistent with previously published work showing that both AtFH3 and AtFH5 are bona fide actin nucleation-promoting factors [25, 31]. The reduction in the level of filamentous actin was also noted in formin loss-of-function mutant pollen grains when they were compared to WT pollen grains (S2 Fig). It was previously proposed that within the pollen tube, actin filaments are arranged into two distinct arrays based on their origin [15]. Actin filaments originating from the apical membrane are organized into a unique apical actin structure which, in Arabidopsis pollen tubes, has its base 4–5 μm away from the tip [15]. Our results showed that the reduction in the level of actin filaments is very prominent within the region of the pollen tube that corresponds to the apical actin structure (Figs 2A and 2B and S1A). We evaluated the quantitative contribution of AtFH3 and/or AtFH5 to apical actin polymerization by measuring the amount of actin filaments within the region that is occupied by the apical actin structure. The results showed that loss of function of AtFH3 causes a comparatively more severe reduction in the actin filament level than loss of function of AtFH5 (Figs 2C and S1B). In addition, loss of function of both AtFH3 and AtFH5 caused an even more severe reduction in the actin filament level (Figs 2C and S1B), suggesting that AtFH3 and AtFH5 have overlapping function in regulating apical actin polymerization. We found that, although loss of function of AtFH3 or AtFH5 also causes reduction in the amount of actin filaments in the shank region, loss of function of both AtFH3 and AtFH5 does not have overt effect on the amount of actin filaments in the shank region of pollen tubes (Figs 2D and S1C). The reduction in the level of apical actin filaments was directly visualized by generating 3D plots of the 2D distribution of actin filament staining (Fig 2E). We found that, although the level of apical actin filaments in pollen tubes is not significantly different between single AtFH3 loss-of-function mutants and AtFH3 and AtFH5 loss-of-function double mutants (Figs 2C and S1B), actin filaments appear more fragmented and disorganized within the apical region of pollen tubes from double mutants than from fh3 single mutants (Figs 2A and S1A). To quantify the degree of actin filament disorganization, we measured the angles formed between the apical actin filaments and the growth axis of the pollen tube. We noticed that the angles were significantly higher in fh3-1 and fh3-1 fh5-2 mutant pollen tubes (Fig 2F), indicating that the apical actin filaments were relatively disorganized in these mutants. The increase in the angle is greater in fh3-1 fh5-2 pollen tubes than in fh3-1 pollen tubes (Fig 2F). The increase in the angle was also noticed for actin filaments within the shank region of formin loss-of-function mutant pollen tubes (Figs 2G and S1D). Thus, the data showed that both AtFH3 and AtFH5 are involved in the regulation of the polymerization and organization of apical actin filaments in the pollen tube, and AtFH3 has a more dominant role in this respect.
To reveal how loss of function of AtFH3 and/or AtFH5 leads to the reduction in pollen tube growth, we examined the distribution of vesicles in pollen tubes. Transport vesicles were decorated with YFP-RabA4b as described previously [48]. We found that YFP-RabA4b-decorated transport vesicles accumulated in an inverted “V” cone shape in the WT pollen tube (Fig 3A). By contrast, the inverted “V” cone shape created by accumulation of apical vesicles is not very obvious in formin loss-of-function mutant pollen tubes (Fig 3A). Considering that the amount of apical actin filaments is reduced in formin loss-of-function mutant pollen tubes (Figs 2A–2C and 2E and S1A and S1B), this result is consistent with the notion that apical actin filaments spatially restrict vesicles within the apical region of the pollen tube [6, 15]. Surprisingly, we found that the region of vesicle accumulation was enlarged in fh3-1 fh5-2 pollen tubes compared to WT pollen tubes (Fig 3A and 3B), which is presumably because the abnormally organized apical actin filaments cannot physically restrict the vesicles within the apical region. In addition, the fluorescence of vesicles is obviously brighter in fh3-1 fh5-2 pollen tubes than in WT pollen tubes (Fig 3A and 3C), which is very likely because the backward movement of vesicles from the tip is severely reduced. To examine the dynamics of YFP-RabA4b-decorated vesicles, we used the technique of fluorescence recovery after photobleaching (FRAP). After bleaching the apical and subapical regions, we found that the recovery rate of vesicles is reduced in fh3-1 and fh3-1 fh5-2 mutant pollen tubes (Fig 3D and 3E and S1, S2 and S4 Movies). By comparison, the recovery rate of vesicles in fh5-2 pollen tubes is only slightly slower than WT pollen tubes (Fig 3D and 3E and S1 and S3 Movies). The extent of the alteration in vesicle recovery rate within the apical and subapical regions of the pollen tube correlates well with the extent of the reduction in apical actin filaments (Figs 2A–2C and 2E and S1A and S1B). Thus, the results showed that loss of function of AtFH3 and/or AtFH5 alters the apical vesicle accumulation pattern and reduces the rate of vesicle turnover in the pollen tube.
To determine the precise intracellular localization of AtFH3 and AtFH5, we generated green fluorescent protein (GFP) fusion constructs of AtFH3 and AtFH5 driven by their own promoters, AtFH3pro:AtFH3-eGFP and AtFH5pro:AtFH5-eGFP, and transformed them into fh3-1 and fh5-2 to generate the transgenic plants AtFH3pro:AtFH3-eGFP;fh3-1 and AtFH5pro:AtFH5-eGFP;fh5-2, respectively. We found that transformation of those constructs rescued the defects in the amount and organization of apical and subapical actin filaments (S3 Fig), suggesting that the GFP fusion constructs are functional. Confocal microscopy revealed that both AtFH3-eGFP and AtFH5-eGFP form punctate structures in the cytoplasm of pollen grains and pollen tubes (Fig 4A, 4B, 4D and 4E). This suggests that the fusion proteins are associated with the endomembrane system, which is consistent with previous characterization of AtFH5 [4]. In addition, both AtFH3 and AtFH5 are able to localize to the PM but exhibit distinct patterns: AtFH3 is localized quite evenly on the PM along the entire pollen tube (Fig 4A and 4B and S5 Movie) while AtFH5 is concentrated on the PM within the apical dome of the tube (Fig 4D and 4E and S6 Movie). After plasmolysis, AtFH5-eGFP retained its association with apical membranes while AtFH3-eGFP was retained on the PM along the pollen tube (S4 Fig). AtFH3 and AtFH5 also had distinct localization patterns in ungerminated pollen grains, with AtFH3 exhibiting obvious PM localization (Fig 4A) whereas AtFH5 does not exhibit obvious PM localization (Fig 4D). The endomembrane and PM localization of AtFH3 and AtFH5 in pollen tubes were further confirmed by staining with FM4-64 dyes, which showed that AtFH3-eGFP and AtFH5-eGFP overlapped with FM4-64 dyes on the cell membrane and punctate structures within the cytoplasm (Fig 4C and 4F). The subcellular localization data are consistent with the presence of a transmembrane (TM) domain in AtFH3 and AtFH5 [4, 25, 31]. We found that depolymerization of the actin cytoskeleton (S5A Fig) does not prevent the PM and endomembrane targeting of AtFH3 and AtFH5 (S5B and S5C Fig). This suggests that their targeting to the PM and endomembranes in pollen tubes does not require their interaction with the actin cytoskeleton. Furthermore, we found that the N-terminus of AtFH3, which contains the signal peptide (SP) and TM, is sufficient for targeting of AtFH3 to the PM and endomembranes (S6A Fig). This is consistent with previous observations that the membrane localization of LiFH1 is determined by its N-terminus, which also contains the SP-TM domain [43]. Strikingly, we found that replacement of the TM domain of AtFH3 with that of AtFH5 endows AtFH3 with a PM localization pattern similar to that of AtFH5 (S6B Fig). Thus, our study suggests that both AtFH3 and AtFH5 are able to localize to the PM and endomembrane system, and the membrane localization pattern is dictated by their TM domains.
To understand the defective actin filament organization in formin loss-of-function mutant pollen tubes, we traced the dynamics of individual actin filaments decorated with Lifeact-eGFP as described previously [16, 49]. We found that actin filaments are continuously polymerized from the apical membrane during WT pollen tube growth, and consequently form a bright apical actin structure (Fig 5A) [15]. However, we found that apical actin polymerization was impaired in formin loss-of-function mutant pollen tubes, and this affected the formation of the apical actin structure (Figs 5A and S7 and S7–S10 Movies). We next traced the dynamics of individual actin filaments and determined the parameters associated with them. Given that apical actin polymerization is severely impaired in fh3-1 fh5-2 mutant pollen tubes and it is hard to select individual apical membrane-originated actin filaments for measurement, we only traced the dynamics of individual actin filaments in WT, fh3-1 and fh5-2 pollen tubes and carefully compared their dynamic parameters. We found that the elongation rate of actin filaments originating from the apical membrane is reduced significantly in fh3-1 and fh5-2 pollen tubes when compared to WT pollen tubes (Fig 5B). This explains to some extent the impairment in the formation of the apical actin structure. In addition, we found that although there is no overt difference in the severing frequency of actin filaments between fh3-1 and fh5-2 pollen tubes and WT pollen tubes (Fig 5D), the depolymerization rate of actin filaments is reduced significantly in fh3-1 and fh5-2 pollen tubes when compared to WT pollen tubes (Fig 5C). Furthermore, no overt difference was detected in the maximal filament lifetime of apical actin filaments in fh3-1 and fh5-2 pollen tubes (Fig 5F), but the maximal filament length of apical actin filaments is reduced significantly in fh3-1 pollen tubes (Fig 5E), which is very likely due to the reduction in the filament elongation rate of apical actin filaments in fh3-1 pollen tubes. Given that AtFH3 and AtFH5 are bona fide actin nucleation-promoting factors [25, 31], the reduction in the filament depolymerization rate in formin mutant pollen tubes is very likely indirect, and is presumably related to the reduction in the actin filament level in formin mutant pollen tubes. Together, these data suggest that the impairment in the formation of the apical actin structure is mainly caused by the defects in formin-mediated actin filament elongation.
Here we provide direct genetic and cytological evidence showing that two class I formins, AtFH3 and AtFH5, are involved in the regulation of actin polymerization and construction of the apical actin structure in the pollen tube. This finding provides another piece of evidence to show that class I formins are major actin nucleation factors that control apical actin polymerization [4], and further supports the notion that actin assembly mediated by formin/profilin modules is one of the major actin polymerization pathways in the pollen tube [4, 44]. Our study thus substantially enhances our understanding of the molecular mechanisms underpinning rapid actin assembly at pollen tube tips and provides significant insights into actin-mediated regulation of pollen tube growth.
We found that AtFH3 and AtFH5 are involved in the regulation of actin polymerization in pollen cells (Figs 2 and 5 and S1, S2 and S7). Within pollen tubes, the reduction in the level of filamentous actin resulting from loss of function of AtFH3 and AtFH5 is comparatively more severe at the tip (Figs 2 and 5 and S1 and S7), and this is very likely because the activity of formin is strictly required at the pollen tube tip where active actin polymerization occurs [15]. Correspondingly, we found that formin is relatively concentrated on the PM at the tip of the pollen tube (Fig 4). In support of the role of formin in regulating apical actin polymerization, a recent study showed that a class I formin, LiFH1, is involved in the construction of the actin fringe in the pollen tube [43]. However, LiFH1 is biochemically distinct from AtFH3 and AtFH5 since it has actin filament-bundling activity [43]. If there is a formin that behaves like LIFH1 in Arabidopsis pollen tubes, it will be interesting to explore if and how it coordinates with AtFH3 and AtFH5 to regulate apical actin polymerization. In terms of the effect of loss of function of AtFH3 on the organization of actin filaments in the shank of pollen tubes, our results differ slightly from a previous report that RNAi-mediated downregulation of AtFH3 causes severe defects including disorganized shank-localized actin cables and depolarized pollen tube growth [31], as we noticed the disorganization of actin filaments in the shank of fh3 pollen tubes and inhibition of pollen tube growth but not that severe (Figs 1, 2 and 5). We do not currently know the reason for this, but it could be due to off-target effects derived from the RNAi-mediated downregulation approach. For instance, this RNAi construct might target to other pollen-expressed formins. If this is indeed the case, it explains why loss of function of AtFH3 and/or AtFH5 causes weak phenotype in term of pollen tube growth. Certainly, we also cannot rule out the possibility that Arp2/3 complex might take a partial role in nucleating actin assembly to compensate for the loss of AtFH3 and/or AtFH5 in pollen tubes, although Arp2/3 complex and formins nucleate actin assembly using different biochemical mechanisms.
Nonetheless, we convincingly demonstrate that AtFH3 and AtFH5 contribute to actin polymerization at the pollen tube tip and AtFH3 plays a more dominant role than AtFH5 in this process (Figs 2 and 5 and S1 and S7). This is actually consistent with the fact that the expression of AtFH3 is more abundant than AtFH5 in pollen (https://www.genevestigator.com/gv/index.jsp). Considering these results along with observations showing that actin filaments are mainly generated from the apical membrane (Figs 5A and S7) [15, 16, 44, 50, 51], it is fair for us to propose that the membrane-anchored class I formins, AtFH3 and AtFH5, drive actin polymerization by utilizing profilin-actin complexes in the cytoplasm within the apical region of the pollen tube (Fig 5G). We showed that AtFH3 and AtFH5 are abundant within the cytoplasm of the growth domain and are presumably localized on vesicles (Fig 4B and 4E). These observations suggest that the activity of formins on the surface of vesicles is maintained at a very low level since no obvious actin polymerization was detected from the surface of vesicles in pollen tubes. Furthermore, it was reported that apical actin polymerization occurs concurrently with and is required for pollen tube growth [15]. The mechanism that activates formins on the PM during pollen tube growth is of great interest. Compared to non-plant formins, plant formins lack the GTPase-binding domain (GBD) and the diaphanous autoregulatory domain (DAD) that are crucial for the regulation of their actin nucleation activity [18, 19, 24]. The molecular mechanisms that tightly regulate the activity of plant formins remain a mystery.
We found that AtFH3 and AtFH5 both localize to the PM, and exhibit distinct PM localization patterns (Fig 4). Considering that the intracellular localization of AtFH3 and AtFH5 is determined by their TM domains (S6 Fig), the distinct PM localization pattern of AtFH3 and AtFH5 suggests that their TM domains have distinct functions. In support of this notion, we found that substitution of the TM domain of AtFH3 with that of AtFH5 enables AtFH3 to exhibit a PM localization pattern similar to that of AtFH5 (S6B Fig). Our data suggest that, although AtFH3 and AtFH5 belong to the same subclass, functional divergence of their TM domains has endowed them with distinct PM localization patterns. The two proteins might consequently perform distinct roles in regulating membrane-originated actin polymerization within the pollen tube. The function of formins, as regulators of actin dynamics, is achieved through their FH1FH2 domain [18]. It remains to be determined whether the C-terminal FH1FH2 domain of AtFH3 and AtFH5 might have distinct actin regulatory functions. Previous studies revealed that both AtFH3 and AtFH5 are bona fide actin nucleation factors [25, 31], but no side-by-side comparison has been performed. Given that actin monomers were predicted to be buffered by an equal amount of profilin in pollen [9], AtFH3 and AtFH5 might differ in their ability to utilize profilin-actin complexes in the pollen tube. There are at least five actin isovariants and two profilin isovariants in Arabidopsis pollen, and they were reported to be distributed uniformly in the cytoplasm of the pollen tube overall [44, 52, 53]. It is possible that they may form different profilin-actin complexes within the cytoplasm. In this regard, AtFH3 and AtFH5 might have preference for certain profilin-actin complexes in the pollen tube. Consequently, the combination of different actin, profilin and formin isovariants may fine-tune the actin polymerization machinery to meet the demands of rapid pollen tube growth. In support of this speculation, a previous report showed that AtFH4 interacts specifically with profilin 2 (PFN2) rather than PFN3 [41].
The Arabidopsis plants were cultured at 22°C under a 16-h light/8-h dark cycle. The T-DNA insertion mutants, fh3-1 (Salk_150350), fh5-2 (Salk_044464), and fh5-3 (Salk_152090) were obtained from Nottingham Arabidopsis Stock Center on the Columbia-0 ecotype (Col-0) background. They were backcrossed with Col-0 three times before the subsequent phenotypic analyses. fh3-2 (CSHL_GT24923) was obtained from Cold Spring Harbor Laboratory and backcrossed with Col-0 three times before the phenotypic characterization. The genotyping of fh3-1, fh5-2 and fh5-3 was performed using primer pairs fh3-1 salk_150350-LP/fh3-1 salk_150350-RP and fh5-2 salk_044464-LP/fh5-2 salk_044464-RP, and fh5-3 salk_152090-LP/fh5-3 salk_152090-RP (S1 Table), respectively, in combination with Salk_LB 1.3 (S1 Table). The genotyping of fh3-2 was performed using primer pair fh3-2 CSHL_GT24923-LP/fh3-2 CSHL_GT24923-RP along with Ds3-1 (see S1 Table). The T-DNA insertion mutant fh5-2 has been characterized previously [25]. To determine the functional coordination between AtFH3 and AtFH5, fh3-1 fh5-2 and fh3-2 fh5-3 double mutants were generated by crossing fh3-1 with fh5-2 or fh3-2 with fh5-3.
To complement fh3-1 and fh5-2 and indicate the intracellular localization of AtFH3 and AtFH5, GFP fusion constructs of AtFH3 and AtFH5 driven by their own promoters were generated. To generate the AtFH3-eGFP fusion construct, the nucleotide sequence containing the promoter and genomic region of AtFH3 were amplified from Arabidopsis genomic DNA with the primer pair AtFH3pg-PstI-F/AtFH3pg-KpnI-R (see S1 Table) and eGFP was amplified from pCAMBIA1301 carrying eGFP with eGFP-SacI-F/eGFP-EcoRI-R (see S1 Table). The PCR products were subsequently moved into pCAMBIA1301 to generate pCAMBIA1301-gFormin3-eGFP. To generate the AtFH5-eGFP fusion construct, the promoter sequence and the genomic sequence of AtFH5 were amplified with the primer pairs AtFH5pro-F/AtFH5pro-R and AtFH5genomic-F/AtFH5genomic-R (see S1 Table), respectively. Given that no suitable restriction enzyme sites were available, the AtFH5 genomic sequence was mutated to disrupt an internal SacI restriction site so that SacI could then be used for the subsequent cloning. The AtFH5 genomic sequence was amplified with primers gAtFH5-Mut-F/gAtFH5-Mut-R using the AtFH5 genomic sequence as the template. The product was subsequently moved into pCAMBIA1301 to generate the final pCAMBIA1301-gFormin5-eGFP construct. The constructs gFormin3-eGFP-pCAMBIA1301 and gFormin5-eGFP-pCAMBIA1301 were transformed into fh3-1 and fh5-2 to generate the transgenic plants, gFormin3-eGFP-pCAMBIA1301;fh3-1 and gFormin5-eGFP-pCAMBIA1301;fh5-2, respectively, using the agro bacteria-mediated flower-dipping method [54]. The transgenic plants at T3 were used for the subsequent analysis. To determine whether the intracellular localization pattern of AtFH3 is determined by its N-terminus, which contains signal peptide (SP) and transmembrane (TM) domain, we amplified the sequence containing both SP and TM of AtFH3 (AtFH3-SP-TM) using primer pair AtFH3-SPTM-F/AtFH3-SPTM-R (S1 Table). The PCR product of AtFH3-SP-TM, along with the Lat52 promoter amplified with pair Lat52-F/Lat52-R (S1 Table), was moved into pCAMBIA1301 to generate pCAMBIA1301-Lat52pro-AtFH3-SP-TM-eGFP. The construct was subsequently transformed into WT Arabidopsis plants. Pollen derived from the transgenic plants was germinated on solid GM for 2 h, then observed under an Olympus FV1000MPE multiphoton laser scanning confocal microscope equipped with a 100× objective (numerical aperture of 1.4). Samples were excited under a 488-nm argon laser with the emission wavelength set at 505–605 nm.
To replace the TM domain of AtFH3 with that of AtFH5, overlap PCR was performed to amplify the sequence of the promoter of AtFH3 (AtFH3pro) and AtFH5-SP-TM with primer pairs 3+5TM F2/3+5TM R2 and 3+5TM F3/3+5TM R3 (S1 Table) using AtFH3pro and AtFH5-SP-TM as the template, respectively. Subsequently, the overlap products were amplified specifically with primer pair 3+5TM F1-XbaI/3+5TM R1-KpnI (S1 Table). Given that no suitable restriction sites were available, the AtFH3pro-AtFH5-SP-TM genomic sequence was subsequently mutated using PCR with primer pair 3+5TM-Mut-F/3+5TM-Mut-R (S1 Table) to disrupt an internal PstI restriction site in order to facilitate subsequent cloning. The sequences of AtFH3pro-AtFH5-SP-TM and AtFH3 FH1FH2 were then amplified with primer pairs 3+5TM F2-PstI/3+5TM R2-XbaI and AtFH3 FH1FH2-F/AtFH3 FH1FH2-R (S1 Table), respectively. The error-free PCR products were subsequently moved into pCAMBIA1301 to generate pCAMBIA1301-AtFH3pro-AtFH5-SP-TM-AtFH3-FH1FH2-eGFP. The construct pCAMBIA1301-AtFH3pro-AtFH5-SP-TM-AtFH3-FH1FH2-eGFP was finally transformed into fh3-1 to generate the transgenic plants, pCAMBIA1301-AtFH3pro-AtFH5-SP-TM-AtFH3-FH1FH2-eGFP;fh3-1. The transgenic Arabidopsis plants at T3 were used for the subsequent analysis.
qRT-PCR was performed to determine the transcript levels of AtFH3 and/or AtFH5 in the formin T-DNA insertion mutants. Total RNA was isolated from pollen derived from WT (wild-type), fh3-1, fh3-2, fh5-2, fh5-3, fh3-1 fh5-2 and fh3-2 fh5-3 plants using TRIzol reagent (Invitrogen) according to the manufacturer’s instructions, and cDNA was subsequently synthesized using MMLV reverse transcriptase (Promega) with oligo-d(T)18. To determine the AtFH3 transcript levels, partial coding region sequences of AtFH3 were amplified with primer pairs AtFH3-F1/AtFH3-R1 and AtFH3-F2/AtFH3-R2 (see S1 Table). To determine the AtFH5 transcript levels, the partial coding region of AFH5 was amplified with the primer pair AtFH5-F/AtFH5-R. To determine the AtFH3 and AtFH5 transcript levels in the complementation plants, the primer pairs AtFH3-F2/AtFH3-R2 and AtFH5-F/AtFH5-R (see S1 Table) were used, respectively. The internal control was eIF4A, which was amplified with the primer pair q-eIF4A-F/q-eIF4A-R (see S1 Table). The real-time PCR data were analyzed with the method of Livak (2-ΔΔCt) [55].
In vitro Arabidopsis pollen germination was performed according to previously described methods [56]. Briefly, pollen was isolated from newly opened flowers and placed on pollen germination medium [GM: 1 mM CaCl2, 1 mM Ca(NO3)2, 1 mM MgSO4, 0.01% (w/v) H3BO3, and 18% (w/v) sucrose solidified with 0.8% (w/v) agar, pH 6.9~7.0]. The plates were cultured at 28°C under moist conditions. After 2 h of culture, the pollen germination rate was quantified by observing pollen grains and pollen tubes under an IX71 microscope (Olympus) equipped with a 10× objective. Images were collected with a Retiga EXi Fast 1394 CCD (charge-coupled device) camera using Image-Pro Express 6.3 software. To calculate the pollen germination percentage, a minimum of 300 pollen grains was counted in each experiment. At least three experiments were performed.
To accurately calculate the pollen tube growth rate, we developed a new method based on calculating the slope of a kymograph of a single growing pollen tube. Briefly, after the pollen tube grew to an average length of approximately 200–300 μm, the solid pollen germination medium containing the germinating pollen was moved to a circular plate (Cat# D35-20-1-N, In Vitro Scientific) under an IX71 microscope (Olympus) equipped with a 4× objective. A microscope field containing at least 15~20 pollen tubes was identified, and the growth of individual pollen tubes was monitored by collecting time-lapse images (about 15–20 images in total) at time intervals of 1 min. A kymograph was created along the growth direction at the center of the growing pollen tube and the growth rate of the pollen tube was calculated from the slope of the kymograph. The experiments were repeated at least three times.
Pollen tubes were stained with the lipophilic dye FM4-64 (Invitrogen). The loading of pollen tubes with FM4-64 was achieved by direct addition of FM4-64 dye (5 μM in liquid pollen germination medium) on the surface of solid pollen germination medium. After incubation with FM4-64 solution for 15 min, images were captured with an Olympus FV1000MPE multiphoton laser scanning confocal microscope as described above. FM4-64 dye was excited with an argon laser at 546 nm, and the emission wavelength was set in a range of 600–650 nm.
To reveal the organization of the actin cytoskeleton in pollen grains and pollen tubes, pollen grains were germinated for 2 h on solid GM, then subjected to fixation and staining with Alexa-488/568 phalloidin as described previously [52, 57]. Actin filaments were observed with an Olympus FV1000MPE multiphoton laser scanning confocal microscope equipped with a 100× objective (numerical aperture of 1.4). The fluorescent phalloidin was excited with an argon laser at 488 nm and 560 nm, and the emission wavelength was set in the range of 505–605 nm and 650–700 nm, respectively. The relative amount of actin filaments in pollen grains and pollen tubes was quantified by measuring the fluorescence pixel intensity using ImageJ software (http://rsbweb.nih.gov/ij/; version 1.46). At least three experiments were performed. The organization of apical actin filaments or bundles was quantified by determining the angles formed between each apical actin filament or bundle and the growth axis of pollen tubes, which was performed with ImageJ roughly as described previously for the quantification of the angles formed between longitudinal actin cables and the growth axis of the pollen tube in the shank region [56]. To ensure that each apical actin filament or bundle was analyzed only once, three to four optical sections were excluded for analysis in each pollen tube. Since we do not know the polarity of apical actin filaments or bundles, we only selected the small angles. More than 200 apical actin filaments or bundles from 10 pollen tubes were measured for each genotype.
In order to visualize the dynamics of actin filaments in pollen tubes, the actin marker Lifeact-eGFP was introduced into formin loss-of-function mutants (fh3-1, fh3-2, fh5-2, fh5-3, fh3-1 fh5-2 and fh3-2 fh5-3) by crossing the mutants with transgenic WT plants harboring Lat52:Lifeact-eGFP [16]. Time-lapse Z-series images were collected every 2 s using MetaMorph software with the step size set at 0.7 μm. The dynamics of individual actin filaments were quantified by measuring dynamic parameters, including the elongation rate, depolymerization rate, severing frequency, maximum filament length and maximum filament lifetime as described previously [16, 57]. At least ten pollen tubes for each genotype were analyzed. A kymograph taken along the growth direction at the center of the pollen tube was created to analyze the F-actin intensity along the growing pollen tube as described previously [16].
To observe the tip-directed vesicle transport in pollen tubes, YFP-RabA4b was introduced into the formin loss-of-function mutants by crossing the mutants with transgenic WT plants expressing Lat52:YFP-RabA4b [48, 58]. Pollen from the resulting plants was germinated on solid GM at 28°C, and when the pollen tubes reached about 150 μm, they were imaged under an Olympus FV1000MPE multiphoton laser scanning confocal microscope equipped with a 100× objective (numerical aperture of 1.4). Samples were excited under a 488-nm laser with the emission wavelength set at 505–605 nm. Optical sections were scanned with the step size set at 0.7 μm. For the fluorescence recovery after photobleaching (FRAP) experiments, apical regions were bleached for 3 s using a 488-nm laser at 100% power and a 405-nm laser at 45% power. Fluorescence recovery was recorded at 2 s intervals for 200s with a 488-nm laser at 10% power. To determine the recovery rate, the mean gray value of the apical region (0–5 μm away from the tip) was measured using ImageJ software and plotted against the elapsed time as described previously [52, 58]. Experiments were repeated at last 20 times and the values of YFP-RabA4b fluorescence were averaged and used for subsequent exponential curve fitting as described previously [52].
|
10.1371/journal.pcbi.1000339 | The Homeostasis of Plasmodium falciparum-Infected Red Blood Cells | The asexual reproduction cycle of Plasmodium falciparum, the parasite responsible for severe malaria, occurs within red blood cells. A merozoite invades a red cell in the circulation, develops and multiplies, and after about 48 hours ruptures the host cell, releasing 15–32 merozoites ready to invade new red blood cells. During this cycle, the parasite increases the host cell permeability so much that when similar permeabilization was simulated on uninfected red cells, lysis occurred before ∼48 h. So how could infected cells, with a growing parasite inside, prevent lysis before the parasite has completed its developmental cycle? A mathematical model of the homeostasis of infected red cells suggested that it is the wasteful consumption of host cell hemoglobin that prevents early lysis by the progressive reduction in the colloid-osmotic pressure within the host (the colloid-osmotic hypothesis). However, two critical model predictions, that infected cells would swell to near prelytic sphericity and that the hemoglobin concentration would become progressively reduced, remained controversial. In this paper, we are able for the first time to correlate model predictions with recent experimental data in the literature and explore the fine details of the homeostasis of infected red blood cells during five model-defined periods of parasite development. The conclusions suggest that infected red cells do reach proximity to lytic rupture regardless of their actual volume, thus requiring a progressive reduction in their hemoglobin concentration to prevent premature lysis.
| The parasite Plasmodium falciparum is responsible for severe malaria in humans. The 48 hour asexual reproduction cycle of the parasite within red blood cells is responsible for the symptoms in this disease. Within this period, the parasite causes massive changes in the host red cell, increasing some metabolic activities hundredfold, making it leaky to many nutrients and waste products, and consuming most of the cell's hemoglobin, far more than it needs for its own metabolism. The challenge that we faced was to explain how the infected cell maintained its integrity throughout such a violent cycle. Seeking clues, we developed a mathematical model of an infected cell in which we encoded our current knowledge and understanding of the complex processes that control cell homeostasis. We present here for the first time a detailed description of the model and a critical analysis of its predictions in relation to the available experimental evidence. The results support the view that host-cell integrity is maintained by the progressive reduction in the hemoglobin concentration within the host cell, resulting in a reduced rate and extent of swelling.
| Plasmodium falciparum, Pf, is responsible for the most severe form of malaria in humans, representing a major cause of morbidity and mortality, especially among children. The pathology of malaria is caused by the intraerythrocytic stage of the parasite cycle. Invasion of a red blood cell (RBC) by a Pf merozoite converts a metabolically languid, hemoglobin-filled cell lacking intracellular organelles and structures, into a complex double cell, with a eukaryotic organism growing and multiplying inside, protected from immune attack. After a relatively quiescent period of about 15–20 h post-invasion, infected RBCs exhibit large increases in metabolic activity and solute traffic [1]–[7] across their membrane. The elevated metabolic rate persists until late stages of development and relaxes only during the latest hours of the parasite's 48 h asexual reproduction cycle. Staines et al. [8] showed that if uninfected human RBCs were permeabilized to the same extent the uninfected cells would hemolyze by the unbalanced net gain of NaCl and osmotic water over a shorter time-course than that needed for parasite maturation and exit. How, then, is the integrity of parasitized cells preserved for the duration of the intraerythrocytic cycle, considering that they have a parasite growing to a substantial volume inside?
This puzzle prompted an investigation on how premature lysis is prevented in falciparum-infected RBCs. A mathematical model of the homeostasis of parasitized RBCs was formulated to attempt an understanding of the processes involved [9],[10]. The model encoded all known kinetic parameters relevant to the control of host red cell volume, i.e., pH, membrane potential, ion content, ion transport across the RBC membrane and parasite growth. The initial simulations with the model produced a result which led to the formulation of a “colloid-osmotic hypothesis” to explain how infected RBCs (IRBCs) resist premature lysis. The hypothesis linked lysis resistance to hemoglobin consumption, a link hitherto never suspected.
It had been well established that during the process of growth and maturation within RBCs malaria parasites ingest and digest hemoglobin (Hb) to levels far above those required by parasite protein synthesis [11]. Moreover, the amino acids produced in vast excess by hemoglobin proteolysis are rapidly released to the medium across the host RBC membrane through the so called “new permeation pathways” (NPPs) [7], [11]–[14], without apparent generation of any osmotic stress. Hb ingestion and digestion and heme detoxification, required to prevent damage to both parasite and host cell, are high energy-consuming processes [15]–[17]. It was therefore puzzling why Hb was consumed in such vast excess. The original model simulations [9] suggested that excess Hb digestion was necessary to reduce the colloid-osmotic pressure within the host cell, thus preventing its premature swelling to the critical hemolytic volume (CHV).
The simulations led to two critical predictions: (i) that excess Hb ingestion and digestion would cause not only a dramatic fall in the Hb content of the host cell, as had been established already from experimental evidence, but also a progressive and large decline in its Hb concentration, and (ii) that parasite volume growth together with host cell swelling late in the cell cycle would bring IRBC volumes very near their CHV. Allen and Kirk [18] argued that parasite volume growth was overestimated in the model because of its assumption that the parasite retains all of the volume taken up via the endocytotic feeding process, leading to exaggerated IRBC volume estimates. They stressed that most of the host volume ingested may be lost by different processes (e.g, Hb breakdown, Na+ extrusion from the parasite compartment) thus freeing up space for the parasite to expand into the host cell and limit the extent of swelling undergone by the infected cell as the parasite volume increases and cations enter the cell through colloid-osmosis. If IRBC volumes do not approach the presumed CHVs, then both the role attributed by the model to excess Hb consumption and the need for freeing space appear irrelevant. It became clear from these considerations that a thorough re-evaluation of the model assumptions and predictions in the light of past and recent experimental evidence became necessary.
In this paper we present a new, detailed analysis of the homeostasis of P. falciparum-infected RBCs. A full description of the model equations is given (see Text S1), and a comprehensive analysis of model assumptions and predictions over the full range of parameter values supported by experimental evidence is provided. We conclude with a re-evaluation of the colloid-osmotic hypothesis. The original model simulations [9] were generated using a restricted set of parameter values and were reported using a minimal subset of model variables, leaving out much potentially useful information on the homeostasis of IRBCs. The present account overcomes these shortcomings and defines our current understanding of the homeostasis of P. falciparum-infected RBCs.
The presentation of the results below, and their analysis, was guided by two specific aims: to provide a critical overview of the experimental evidence on which the information encoded in the model is based, and to allow a rational exploration of the model predictions over a wide range of possible parameter values. This exercise is necessary to enable predictions to be confronted with available results in the literature, to outline the open questions in the field and to direct future research.
Figure 1A illustrates the time courses of NPP development and of Hb consumption, as encoded in the original version of the model, based on the experimental results from Staines et al. [8] and from Krugliak et al. [12], respectively. Figure 1B shows the time-dependent increase in Na+, K+ and anion permeabilities through the NPP pathway [7]. Note the difference in ordinate scales between anion and cation permeabilities, corresponding to the well-documented anion selectivity of the NPP pathway [7],[19]. We analyse first the rationale behind the two curves in Figure 1A and then the significance of the time shifts between the onset of the NPP-mediated permeabilization and of Hb consumption.
The stage-dependent changes in NPP-mediated permeabilities were measured in samples from synchronized Pf cultures [8]. They were encoded in the model as represented in Figure 1A and 1B. The curves may be interpreted in either of two ways: as a gradual simultaneous increase in NPP-mediated permeability in all the parasitized cells (graded response), or as the net population variation in onset time of sudden permeability changes in individual cells (all-or-none response). Can the available experimental evidence help discriminate between graded or all-or-none alternatives? Isotonic solutions of NPP-permeant solutes such as sorbitol have been extensively used to selectively hemolyse IRBC with developed NPPs [20]. Analysis of the lysis kinetics of IRBCs renders results compatible with both types of responses [21]. Patch-clamp studies have not yet documented intermediate conductance stages in NPP activated IRBCs [22]. Therefore, the all or none response remains a distinct possibility, deserving investigation here by analysing the predicted effects of a sudden increase in NPP-mediated permeability.
The stage-dependent changes in Hb consumption were defined within wide error margins [12],[23]. Hb consumption of up to 80% of the host cell Hb is known to proceed gradually (see, e.g., [23], and there is no “all or none” alternative to gradual Hb consumption. The most important and well supported feature of the two curves in Figure 1A is that NPP development precedes Hb consumption. Sorbitol, alanine and other solutes whose permeability through uninfected RBC membranes is negligible have been extensively used to probe for NPP permeabilization [7]. In isotonic solutions of sorbitol or alanine, IRBCs with developed NPPs rapidly lyse; only those with young ring-stage parasites remain intact. The importance of delayed Hb digestion relative to NPP development is that by the time the large excess of amino acids produced by globin proteolysis reaches the host cell membrane, the permeability path available for their rapid downgradient exit is available, thus preventing osmotic stress from accumulated amino acids within the host. However, the experimental errors in the observed half-times of NPP development and Hb ingestion curves are relatively large and the possible effects of interval variations between the curves in Figure 1 on IRBC homeostasis deserve exploration.
Figures 2 and 3 show the results of a typical simulation with parameter values chosen for convenient illustration of the five homeostatic periods defined by inflexions in the curves reporting net fluid movements in the host RBC (Phases 1 to 5 in Figure 2A). The model predictions here allow a detailed analysis of the homeostatic processes at work during the different stages of parasite development.
During the stage of initial quiescence (Phase 1), from invasion to about 20 h post-invasion, all IRBC variables remain essentially unchanged from their initial levels. Phase 2, K+-driven net fluid loss, is triggered by NPP activation. The immediate effect of the increase in Na+, K+ and anion (A−) permeabilities (Figure 1B) is to induce the dissipation of the steep initial Na+ and K+ gradients (Figure 2B and 2C), unrestricted by co-anion movements [24],[25].
Initially the opposite driving forces for Na+ and K+ gradient dissipation have similar magnitude, as represented in Figure 2D and 2E by the respective electrochemical driving gradients ΔENa and ΔEK. The PK/PNa permeability ratio for cation selectivities is however set at 2.3 [8] and thus determines that the loss of KCl transiently exceeds NaCl gain. This causes a transient net fall in RBC cytosolic anion content and concentration (Figure 2B and 2C, respectively) and net loss of water (Figure 2A, cell water and Figure 2F). Figure 3A shows that the K+ efflux, which initially exceeds Na+ influx, rapidly returns to near-zero baseline levels as the K+ gradient is dissipated. The transient dehydration of the IRBC during this second period generates secondary transient changes in other homeostatic variables: increase in Hb concentration (Figure 3B), reduced anion content (Figure 2B, A−) and anion concentration (Figure 2C, [A−]), and cell acidification (Figure 3C). The transient acidification results from the brief increase in [Cl−]o/[Cl−]i ratio due to net KCl loss; the combined operation of the CO2 shunt and anion exchanger rapidly readjusts the proton ratio to restore the equilibrium condition [H+]i/[H+]o = [Cl−]o/[Cl−]i, with consequent cell acidification [26]–[28].
In Phase 3, Na+-driven fluid gain, the direction of net fluid movement is reversed following the reversal of the gradients driving net salt flows. This reversal also affects the direction of change in all associated variables (Figures 2 and 3). Figure 3A shows that the net fluxes of Na+ and anions into the cell persist long after the net K+ flux has returned to baseline levels, and Figure 2D and 2E shows the time-dependent changes in driving gradients which determine the direction of net ion and fluid fluxes at all times according to the model (see Text S1).
In Phase 4, fluid loss, the rate of Hb consumption is maximal. This rate determines the volume of cytosol that the parasite needs to ingest in order to incorporate the amount of Hb prescribed by the Hb consumption function (Figure 1). When this volume exceeds the concomitant Na+-driven fluid gain, host cell water contents and host cell volume are transiently reduced (Figure 2A). Phase 4 is characterized by the steepest rates of Hb fall (Figure 3B, Hb) and parasite growth (Figure 2A, open triangles), and by a decline in cell water (Figure 2A, solid triangles). Transient reductions in Na+ and anion contents (Figure 2B, Na+,A−) result from the transfer of RBC cytosol to the parasite as part of the Hb ingestion process. Additional reduction in host cell volume results from the removal of the space occupied by Hb molecules. Hb has a specific volume of about 0.74 ml/g [29] and contributes with about 25% to the total volume of a normal RBC. Therefore, a loss of 70–80% of Hb from a cell containing in average 34 pg of Hb is equivalent to a volume loss of between 15–20 fl by the end of the asexual cycle.
Phase 5, sustained swelling, is characterized by continuous NaCl and water gains by the host cell (Figures 2A (cell water), 2B (Na+ and A−), and 2F) driven by the inward Na+ gradient. The rate of fluid gain is reduced relative to that in phase 3 (Figure 2F) because of the marked reduction in driving force for net NaCl gain (Figure 2D and 2E) and in colloid-osmotic pressure due to the fall in Hb concentration (Figure 3B). Parasite and IRBC volumes also increase at slower rates (Figure 2A (Parasite, IRBC)) following the reduced Hb consumption and fluid gains relative to Phase 4. As the anion concentration increases (Figure 2C), the membrane potential becomes progressively more depolarized and the equilibrium potentials of all ions approach the membrane potential Em (Figure 2D and 2E), with consequent cell alkalinisation (Figure 3C). The Na pump, initially stimulated by the increased intracellular Na+ concentration, shows late inhibition (Figure 3D). This inhibition results from a predicted reduction in Mg2+ concentration in the host cell cytosol following global cytosolic transfers to the parasite during Hb ingestion. Late swelling further reduces the Mg2+ concentration. The Mg2+/ATP ratio is an important regulator of Na pump activity and departure from its normal value near unity is inhibitory to the pump [30],[31]. Atamna and Ginsburg [32] measured the Mg2+ content of host and parasite compartments in IRBCs with mature trophozoite stage parasites and found that the Mg2+ content of the host cell compartment was over 60% lower than that of uninfected RBCs. They suggested that such a reduction may partially inhibit active transport by the sodium and calcium pumps. It remains to be elucidated whether the actual mechanism of Mg2+ deprivation in the host cell is the one implied by the model.
Figure 4 illustrates a condition in which NPPs are switched on almost instantly to analyse the homeostatic effects of all-or-none NPP activation (Figure 4A). The rest of parameter values were the same as for Figure 2. It can be seen that the main effect of all-or-none NPP activation is to compress the time-course of the events described for phase 2 (Figure 2), with relatively minor long-term quantitative changes in volumes (Figure 4B) and in Hb concentration (Figure 4C). The relative duration and magnitude of the effects described for each period in the examples chosen for Figures 2 and 4 will vary with the choice of parameter values. These effects are analysed below. The important point to note is that the underlying homeostatic processes described for each period remain essentially the same.
The time course of volume growth of P. falciparum parasites throughout their asexual reproduction cycle in human RBCs has not been characterized. Parasite volume increases throughout the cycle but it is unknown whether this growth is uniform or variable. The minimal final parasite volume in a cell with a single parasite has to equal the sum of the volumes of all the merozoites produced plus the volume occupied by the residual body. Without relevant information available, it was difficult to design a rational strategy to model parasite volume growth. Because the time-course of parasite volume growth could be roughly associated with that of Hb ingestion, linking these two variables was considered an acceptable modelling strategy. In the initial formulation of the model [9], parasite volume at each instant of time was defined by the cumulative volume of ingested host cell cytosol up to that time. This volume, in turn, was determined by all the complex homeostatic factors that influenced the volume of host cytosol in which the prescribed amount of Hb to be digested at each time was contained. For maximal Hb consumption around 70–80%, this strategy predicted terminal parasite volumes of about 70 to 90 fl (Figure 2A, Parasite), values near the mean volume of uninfected RBCs. However, previous results suggested that this approach overestimated parasite volume.
Saliba et al. [33] measured the water content of parasites at the mature trophozoite stage to be less than 30 fl. Recent results by Elliott et al. [23] suggest that single parasite volumes seldom exceed 50 fl at any developmental stage. Therefore, to explore the effect of more realistic estimates of parasite volumes a coupling factor was introduced. It defines the global volume-growth of the parasite in each iteration of the numerical computation as a fraction of the volume of cytosol incorporated during that iteration (see Text S1). For coupling factor values of less than 0.7, this approach implicitly corrects for parasite volume losses due to Hb breakdown, because, although hemozoin is retained, the volume occupied by the globin molecules largely vanishes in the process of exporting the resulting amino acids to the external medium. As explained above, this volume may account for up to 20 fl. The results of simulations using the same set of parameter values applied in the example of Figure 2, varying only the value of the coupling factor, are shown in Figure 5 which reports predicted parasite volumes as a function of time post-invasion. From these results, only coupling factor values in the range 0.3 to 0.7 appear to cover the observed range of terminal parasite volumes for single infections. This range then will be tested in the global simulations attempted below, in comparison with the original value of 1 [9],[10].
Figure 6 shows the model predictions for five selected variables, plotted as a function of time post-invasion. The chosen range of variation for each parameter was based on experimental results when available or on outcomes of simulations consistent with observation. For instance, although single parasite volumes may remain within the 30 to 50 fl range, IRBCs are often seen with two viable parasites reaching segmentor stages (L. Bannister, personal communication), or with additional volume occupied by developmentally arrested parasites [34]. From the perspective relevant to the homeostasis of the host cell, it is the combined parasite volume that counts, hence the choice of coupling ratios spanning values from 0.3 to 1. The range of half-time values for NPP development and Hb ingestion is shown within ±1 standard deviation of the experimentally-reported means [8],[12]. Figure 6 shows only simulations with parameter values in which IRBC volumes remain below the spherical volume cells can attain within a maximally-stretched membrane, at which point they would lyse. This maximal volume is usually described as the critical hemolytic volume. Following Ponder [35], the nominal CHV was set at a mean value of 1.7 times the original volume of the modelled cell.
The immediate conclusion from gross comparisons between Figure 6A–C and Figure 6D and 6E is that whereas host cell water (Figure 6A), IRBC volume (relative to uninfected RBC volume, Figure 6B), and parasite volumes (Figure 6C) can vary over a very wide range and with large oscillations within the five homeostatic periods described for Figure 2 (Figure 6A and 6B), the predicted decline pattern in host cell Hb concentration remains remarkably uniform (Figure 6E). Therefore, the single novel and invariant prediction of the colloid-osmotic hypothesis is that the Hb concentration within the host cell has to become progressively reduced, regardless of parasite and IRBC volumes (Figure 6E).
The analysis of the homeostasis of Pf infected RBCs (Figures 2, 3, 4, and 6) provides a number of novel insights:
We consider next how experimental results in the literature compare with our model predictions. In Figure 7, experimental measurements of the five selected variables reported in Figure 6 are shown as rectangles over the grey silhouettes of the variable ranges in Figure 6. Despite the large variability in the experimental results, it is clear that they fall within the low range of values for host cell water (Figure 7A), IRBC volume (Figure 7B), and parasite volume (Figure 7C) for single infections [33],[38],[39]. The hemoglobin content measurements show a declining pattern covering the full range of the values encoded in the model (Figure 7D). Most significantly, the recent measurements of the stage-dependent changes in host Hb concentration by Park et al. [40] and Esposito et al. [41], obtained with two independent techniques confirm the declining pattern predicted by the model (Figure 7E), thus lending support to its most relevant prediction.
Do the large variations in parasite and host cell volumes explored with the model and also apparent in the experimental measurements (Figure 7) reflect true IRBC polymorphisms or merely error margins? IRBC polymorphisms are evident in relation to a number of characteristics which can be easily observed and recorded in live cultures: single or multiple invasion, developmental and viability differences among parasites in multiple invasions, IRBC shapes and volumes, parasite sizes and shapes, hemozoin particle content, aggregation state of hemozoin crystals, number of merozoites contained and released, etc. Many of these variations are observed in highly synchronized populations and cannot be attributed simply to differences in developmental stage of the parasite, or to viability state of the IRBCs. It is therefore plausible that the domain of stable homeostatic solutions predicted by the model (Figure 6) does indeed reflect, at least in part, true homeostatic polymorphisms of IRBCs. If so, parasite, host and IRBC volumes may vary within wide margins from cell to cell without necessarily compromising the osmotic stability of the IRBC. This, however, questions the fundamental tenet of the colloid-osmotic hypothesis, that excess Hb consumption is necessary to prevent premature IRBC lysis.
The problem can be clearly illustrated with an example. Let us consider an IRBC whose volume remains near a relative cell volume of 1 throughout the asexual reproduction cycle, as in many of the curves shown in Figure 6B, and as documented experimentally [23],[33],[40]. If the critical hemolytic volume remains set at 1.7 times the initial RBC volume, as originally assumed based on data from uninfected RBCs [35], model simulations indicate that ∼20% Hb consumption would be enough to prevent the premature lysis of a cell with relative volume around 1. So, for such a cell, excess Hb consumption would appear irrelevant for lysis prevention. But the available evidence overwhelmingly supports the view that Hb is consumed in large excess in all viable IRBCs. It follows that reduced colloid-osmosis may not be the main reason for excess Hb consumption, at least not in all instances. However, as discussed next, the conundrum here rests with the rigid attribution of a CHV of 1.7 in the model simulations, not with the basic understanding of IRBC homeostasis provided by the model.
Previous results from osmotic fragility studies in IRBCs showed that the osmotic fragility of RBCs infected with mature trophozoite- and schizont-stage parasites is substantially increased relative to IRBCs with ring-stage parasites or to uninfected cohorts [9],[10]. If the actual volume of at least some IRBCs harbouring mature parasites remains low, then for the osmotic fragility to be increased their CHV has to be somehow reduced.
How can this occur? The CHV of each RBC depends critically on membrane area [35],[42],[43]. If membrane area is reduced by infection, the CHV will also be reduced. The relation between volume (V) and area (A) in a sphere is given by V = A3/2/(6π1/2). Therefore, since the ratio of maximal volumes (V1, V2) of two cells with different surface areas (A1, A2) is V1/V2 = (A1/ A2)1.5, a fractional decrease in area will propagate to a fractional decrease in volume to the power of 1.5, stressing the magnified effect of effective membrane area reductions on CHV. Early results in the literature report opposing claims in relation to membrane area changes in infected IRBCs: population estimates report substantial reductions [39] whereas single cell measurements suggested no change [44]. Recent movies by Glushakova et al. [45] show infected cells about to rupture whose near spherical diameter is less than 80% that of surrounding uninfected discocytes indicating reduced membrane area. Considering, in addition, that the membrane geometry and fluid properties of Pf-infected RBCs become progressively altered by knobs and increased rigidity [46]–[51], the increased osmotic fragility may be compounded by an increased lytic vulnerability to volume expansion, resulting from reductions in membrane area, in the capacity to effect a normal expansion of the full membrane area before lytic rupture, or both.
The increase in the osmotic fragility of IRBCs reflects a progressive hemolytic vulnerability of Pf-infected RBCs to volume expansion by fluid gains. This shows that IRBCs become progressively closer to their effective CHV as the parasite matures, regardless of their actual volume levels. Although the contributions of membrane area loss and other factors to this hemolytic vulnerability remain to be elucidated, excess Hb consumption retains its credential as a general protection mechanism for IRBCs of all volumes, by preventing excessive rates of fluid gain. This mechanism depends critically on the prediction illustrated in Figures 3B and 6E that the Hb concentration must become progressively reduced in all IRBCs, regardless of the specific volume evolution of each IRBC.
In conclusion: the original formulation of the colloid-osmotic hypothesis, using a coupling coefficient of one, predicted that IRBCs would swell close to a CHV level taken as the mean value for uninfected RBCs, premature rupture being prevented by the reduced Hb concentration. Simulations with coupling coefficient values below 0.7 deliver more realistic IRBC volume estimates (Figures 5B and 7B), but experimental results indicate that the progressive proximity of IRBCs to a reduced CHV is retained. Thus, whatever the reason for CHV proximity, reduction in Hb concentration remains essential for preventing rapid fluid gains leading to premature IRBC lysis.
The mathematical-computational model of the homeostasis of Plasmodium falciparum-infected red blood cells (IRCM) was derived as an extension of the original Lew-Bookchin red cell model (RCM) [26]. Both models are available as free-standing executable files from http://www.pdn.cam.ac.uk/staff/lew/index.html. The IRCM first computes a “Reference” steady-state (RS) meant to represent the initial condition of a human red blood cell just invaded by a falciparum merozoite generating a ring-stage internalized parasite occupying 4% of the red cell volume. In the formulation of the RS the programme offers a large variety of options for the user to change constitutive properties of the IRBC such as the value of all the parameters tested in the simulations reported in this paper. For the simulations reported here the medium was assumed to be an infinite reservoir (vanishingly low hematocrit condition). With the RS defined, the programme is set to follow the dynamic evolution of the IRBC system (Dynamic state, DS) for 48 hours with a data output frequency chosen by the user. To enable realistic comparisons with experimental results, experimental conditions can be simulated to explore the modified dynamic behaviour of the system.
|
10.1371/journal.pbio.1001249 | Stochastic Expression of the Interferon-β Gene | Virus infection of mammalian cells induces the production of high levels of type I interferons (IFNα and β), cytokines that orchestrate antiviral innate and adaptive immunity. Previous studies have shown that only a fraction of the infected cells produce IFN. However, the mechanisms responsible for this stochastic expression are poorly understood. Here we report an in depth analysis of IFN-expressing and non-expressing mouse cells infected with Sendai virus. Mouse embryonic fibroblasts in which an internal ribosome entry site/yellow fluorescent protein gene was inserted downstream from the endogenous IFNβ gene were used to distinguish between the two cell types, and they were isolated from each other using fluorescence-activated cell sorting methods. Analysis of the separated cells revealed that stochastic IFNβ expression is a consequence of cell-to-cell variability in the levels and/or activities of limiting components at every level of the virus induction process, ranging from viral replication and expression, to the sensing of viral RNA by host factors, to activation of the signaling pathway, to the levels of activated transcription factors. We propose that this highly complex stochastic IFNβ gene expression evolved to optimize both the level and distribution of type I IFNs in response to virus infection.
| Eukaryotic cells can respond to extracellular signals by triggering the activation of specific genes. Viral infection of mammalian cells, for example, induces a high level of expression of type I interferons (IFNα and β), proteins required for antiviral immunity that protects cells from the infection. Previous studies have shown that the expression of the IFNβ gene is stochastic, and under optimal conditions only a fraction of the infected cells express the IFNβ gene. At present neither the mechanisms nor functions of this interesting phenomenon are well understood. We have addressed this question by analyzing IFN-expressing and non-expressing mouse cells that were infected with the highly transmissible Sendai virus. We show that stochastic IFNβ gene expression is a consequence of cell-to-cell differences in limiting levels and/or activities of virus components at every level of the virus induction process, from viral replication to expression. These differences include the sensing of viral RNA by host factors, the activation of the signaling pathway, and the levels of activated transcription factors. Our findings reveal the complexity of the regulatory mechanisms controlling stochastic IFNβ gene expression. We propose that the stochastic expression of IFN allows for an even distribution of IFN, thus avoiding over-expression of IFN in infected cells.
| Eukaryotic cells respond to extracellular signals and environmental stresses by coordinately activating specific sets of genes. Signals from the cell surface or cytoplasm trigger signaling pathways that culminate in the binding of distinct combinations of coordinately activated transcription factors to promoter and enhancer elements that regulate gene expression. A well-characterized example of this is the activation of type I interferon (IFN) gene expression in response to virus infection or double-stranded RNA (dsRNA) treatment [1],[2]. After infection, viral RNA is detected in the cytoplasm by one of two RNA helicases, retinoic acid-inducible gene I (RIG-I) or melanoma differentiation-associated gene 5 (MDA5), which respond to different types of viruses [3]. RIG-I recognizes short dsRNA or panhandle RNA bearing a 5′ triphosphate group [3], and its activity is positively regulated by the ubiquitin E3 ligase tripartite motif 25 (Trim25) [4]. When RIG-I or MDA5 bind to RNA, they form heterodimers, undergo a conformational change, and expose a critical N-terminal caspase-recruiting domain (CARD) [5],[6]. This domain interacts with the CARD domain of the downstream adaptor protein mitochondrial antiviral signaling (MAVS) (also known as IPS-1/Cardif/VISA) on the mitochondrial membrane [7]. The association of RIG-I with MAVS initiates the recruitment of adaptor proteins and leads to the activation of the transcription factors IFN regulatory factors 3 and 7 (IRF3 and IRF7) and NF-κB by the TANK-binding kinase 1 (TBK1) [8]–[10] and IKKα and IKKβ, respectively [7],[11]. Activated IRF3/IRF7 and NF-κB translocate into the nucleus and, along with the transcription factors ATF2/cJun, bind the IFN-β gene enhancer and recruit additional transcription components to form an enhanceosome [12]. This complex signaling and promoter recognition mechanism functions to coordinately activate a specific set of transcription factors that recognize the unique enhancer sequence of the IFNβ gene and thus specifically activate IFN gene expression.
Early in situ hybridization (ISH) studies revealed that induction of IFNβ expression by virus infection or dsRNA treatment in both human and mouse cells is stochastic [13],[14]. That is, only a fraction of the infected cells express IFNβ. This “noisy” expression is not due to genetic variation within the cell population, as multiple subclones of individual cells display the same low percentage of cells expressing IFNβ [14]. In addition, different mouse and human cell lines display different percentages of expressed cells, and the levels of IFNβ gene expression can be increased in low expressing cell lines by fusing them with high expressing lines, or by treating low expressing lines with IFNβ [13],[14]. These studies suggest that stochastic expression of the IFNβ gene is a consequence of cell-to-cell differences in limiting cellular components required for IFN induction, and that one or more of the limiting factors are inducible by IFNβ [13].
Stochastic expression has been observed with a number of other cytokine genes, including IL-2 [15], IL-4 [16],[17], IL-10 [18], IL-5, and IL-13 [19]. In many of these cases, expression is both stochastic and monoallelic. Recent studies of IFNβ gene expression revealed that stochastic expression in human cells is initially monoallelic, and becomes biallelic later in the induction [20],[21]. In one study the stochastic expression of the IFNβ gene was proposed to be a consequence of intrinsic noise due to stochastic enhanceosome assembly [21]. Subsequently, an analysis of human HeLa cells identified a specific set of Alu-repetitive DNA sequences bearing NF-κB binding sites that associate with the IFNβ gene through interchromosomal interactions, and in so doing are thought to increase the local concentration of NF-κB. Initially, only one of the two chromosomes associates with the specialized NF-κB binding sequence, resulting in early monoallelic expression. Secretion of IFN leads to an increased expression of limiting factors (most likely IRF7, which is inducible by IFN), obviating the need for interchromosomal interactions, and leading to the activation of the second IFNβ allele [20]. More recently, heterogeneity in the infecting viruses, rather than cell cycle differences, has been proposed to be the primary source of IFN stochastic expression [22]. Many functions have been proposed for biological noise, ranging from cell fate decisions during development to survival in fluctuating environments [23]. In the case of the IFN genes, neither the mechanisms nor functions of biological noise are well understood.
Here we report a detailed analysis of stochastic IFNβ gene expression in mouse cells. We make use of an IFN-IRES-YFP reporter mouse [24] to perform a detailed analysis of differences between virus-infected cells that either express or do not express IFNβ. Our results reveal a complex picture of stochastic expression of the IFNβ gene, in which the levels of components required for virtually every step in the virus induction pathway are limiting. This includes components required for viral replication and expression, for sensing the presence of viral RNA by the host, and for the virus induction signaling pathway, and the transcription factors required of IFNβ gene expression. Remarkably, in spite of this complexity the percentage of expressing cells remains constant through recloning and cell division, indicating that the stochasm of clonal cells is genetically programmed.
Sendai virus (SeV) infection of either mouse or human cells leads to the expression of IFNβ mRNA in only a fraction of the infected cells (Figures 1A, 1B, and S1A), and the percentage of expressing cells differs between different cell lines. The time course of mouse IFNβ expression determined by ISH (Figure 1B) is consistent with that from the quantitative PCR (qPCR) analysis (Figure S1B and S1C). Remarkably, the percentage of cells expressing IFN did not exceed 20%, even at the latest time point (Figure 1B). The absence of IFNβ signal in the majority of cells is not an artifact of hybridization, as β-actin mRNA was detected in all cells (Figure S1D). IFNβ mRNA is specifically detected with an antisense IFNβ RNA probe, while no signal is detected with a sense RNA probe (Figure S1E). In addition, similar percentages of IFNβ-expressing cells were detected by immunofluorescent staining using an IFNβ antibody (Figure S1F), strongly supporting the reproducibility and specificity of the IFNβ ISH.
As mentioned above, enhanceosome assembly and limiting amounts of NF-κB have been proposed to be the primary limiting steps in stochastic expression of the human IFNβ gene [20],[21]. To determine whether this stochastic expression is unique to the IFNβ gene because of the complexity of the IFNβ enhanceosome, or is more general, we examined the expression of the IFNα genes, which are coinduced with IFNβ, but have simple enhancer/promoters, and do not require NF-κB [25],[26]. Using either a mouse IFNα4 or human IFNα8 probe, we found that IFNα genes are also stochastically expressed in both mouse and human cells, respectively (Figures 1C and S1G). Although NF-κB has been shown to be a limiting factor in the activation of the human IFNβ gene [20], it is not required for IFNβ expression in mouse cells [27]. Thus, in spite of this difference both the mouse and human IFNβ genes are stochastically expressed. We also examined other virus-inducible genes, and found that they too are stochastically expressed (see below). Each of these virus-inducible genes requires different levels and combinations of transcription factors, yet they are all stochastic. In all of these cases (mouse and human IFNβ and IFNα and the other virus-inducible genes), the common requirement is the RIG-I virus-inducible signaling pathway. We therefore carried out experiments to determine whether limiting components in this pathway contribute to the observed stochastic expression.
To investigate the mechanism of stochastic IFNβ gene expression, we made use of an IFNβ reporter-knock-in mouse, in which YFP expression allows tracking of IFNβ expression at a single-cell level [24]. Using IFNβ/YFP homozygous mouse embryonic fibroblasts (MEFs) and fluorescence-activated cell sorting (FACS), we obtained pure populations of IFNβ-producing and IFNβ-negative cells upon SeV infection. As expected, IFNβ mRNA is high in the YFP-positive cells, and very low in the YFP-negative cells (Figure S2A). As expected, the IFNα2 and IFNα4 genes are also highly expressed in the YFP-positive cells, and not in the YFP-negative cells (Figure S2A). These observations indicate that replication of the infecting virus and/or components in the RIG-I pathway are the limiting steps in the uninduced cells, rather than intrinsic differences in the IFNβ and α promoters.
We also detected the relative mRNA abundance of other virus-inducible genes in IFNβ-expressing and non-expressing cells. As shown in Figure 1D, transcription levels of all tested inflammatory cytokine or chemokine genes are much higher in IFNβ-producing cells compared to nonproducers. Considering the fact that IFNβ-producing cells account for only 10% of the total cell population, we conclude that expression of all these virus-inducible genes is also stochastic and that these genes are coordinately activated with the type I IFN genes. Activation of these virus-inducible genes is known to require the RIG-I signaling pathway [28]–[31]. Thus, our results indicate that stochastic gene expression is due primarily to limiting components in the signaling pathway and not to gene-to-gene variation in the mechanism of gene activation.
In the case of human cells, stochastic expression of the IFNβ gene is randomly monoallelic early and biallelic late in infection, and the activation of the second IFNβ allele is inducible by IFN [20],[21]. However, the nature of allelic expression of the IFNβ gene has not been addressed in mouse cells. By using IFNβ/YFP heterozygous MEFs, we showed that early after infection (<8 h post-infection [h.p.i.]), IFNβ gene expression was primarily monoallelic, while late in infection (8–16 h.p.i.), the majority of IFNβ-expressing cells were both IFNβ and YFP double-positive cells indicating that, as with human cells, a switch to biallelic expression also occurs in mouse cells (Figure S2B).
Previous studies have shown that the levels of IFNβ gene expression can be increased by priming the cells with IFNβ [13]. Using both mouse and human primary fibroblasts, we showed that IFNβ pretreatment also increases the percentages of IFNβ-expressing cells (Figure S3), indicating that the limiting factor(s) contributing to stochastic IFNβ gene expression are, indeed, inducible by IFNβ. One of these IFN-inducible factors is IRF7 ([20] and see below).
To examine the role of the infecting virus in stochastic IFNβ gene expression, we infected primary MEFs with SeV followed by immunofluorescent staining using a SeV antibody. As shown in Figure S4A, most, if not all, of the cells are uniformly infected by SeV, far more than could explain the small percentage of cells expressing IFNβ gene. When we used increasing multiplicities of SeV (as defined by hemagglutination units [HAU]) to infect primary MEFs, we found that the percentage of IFNβ-producing cells increased as the HAU was increased, reaching a maximum of approximately 18% at the peak (Figure S4B). However, as more virus was added (>200 HAU), the percentage of IFNβ-producing cells decreased. Thus, the viral titer is not a limiting factor in the observed stochastic IFNβ gene expression. Next, we determined viral transcript levels in both IFNβ-producing and nonproducing cells. We found that the nucleoprotein (NP), matrix protein, and L polymerase protein mRNA transcripts were present at significantly higher levels in IFNβ-producing cells compared to the nonproducers (Figures 2A and S4C). In addition, higher levels of SeV NP protein were detected in IFNβ-producing cells (Figure 2B).
The RNA helicase RIG-I detects viral genomic RNA and defective interfering (DI) genomes [32],[33]. We therefore examined the levels of viral and DI genomes in both IFNβ-producing and nonproducing cells. As shown in Figure 2C (upper panel), more SeV DI genomes were detected in IFNβ-producing cells compared to IFNβ-nonproducing cells at 8 and 12 h.p.i. Using a primer pair that specifically detects viral genomic RNA, we also detected more viral genomes in IFNβ-producing MEFs 8 and 12 h.p.i. (Figure 2C, lower panel). These results are consistent with the observed viral NP mRNA levels (Figure S4C), and indicate that viral replication is more efficient in the IFN-producing cells. We also investigated the induction activities of total RNA extracted from both IFNβ-producing and nonproducing cells. As shown in Figure S4D, total RNA from IFNβ-producing cells infected for 8 or 12 h induced more IFNβ expression compared to total RNA from IFNβ-nonproducers at the same time points. We conclude that viral mRNA, DI genomes, and viral genomes are present at higher levels in IFNβ-producing cells than in nonproducers. Thus, differences in the efficiency of viral replication/transcription contribute to the stochastic expression of the IFNβ gene.
Previous studies led to the conclusion that the stochastic expression of the IFNβ gene is a feature of the infecting virus, and not of the host cell [22]. To address this possibility, we determined the number of cells that have high levels of viral RNA and produce IFNβ at 8 h.p.i. As shown in Figure 2D, after 8 h of virus infection, approximately 38% SeV-high cells (upper left and upper right) were detected, and about 9% YFP-positive cells (upper right and lower right). Although a higher percentage of IFNβ-expressing cells was observed within the SeV-high cell population (6.56% versus 2.42%), only 17% (6.56% out of 38%) of SeV-high cells produce IFN. Thus, although cell-to-cell differences in viral replication contribute to the stochastic expression of IFN, these differences are not sufficient to explain the extent of stochastic IFN gene expression.
To further investigate the mechanism of stochastic IFNβ gene expression, we determined the localization of various components of the signaling pathway required for IFN production using nuclear and cytoplasmic fractions separated from both expressing and non-expressing cells. Consistent with the limiting component hypothesis, we detected phosphorylation and translocation of IRF3 in the YFP-positive cells, but not in the YFP-negative cells (Figure 3A). Previous studies have shown that IRF3, like IRF7, is phosphorylated by the TBK1 kinase, and translocates from the cytoplasm to the nucleus. As both IRF3 and IRF7 are activated via the RIG-I pathway, our results suggest that one or more components of the RIG-I signaling pathway are limiting in the cells that fail to express IFN. A similar result was obtained with sorted cells at 12 h.p.i. (Figure 3B).
In human cells both NF-κB and IRF3/IRF7 are required for virus induction of the IFNβ gene [12],[34]. The human and mouse IFNβ enhancers differ in only two nucleotides out of 45 bases. However, in mouse cells NF-κB is required only for early antiviral activity, when the level of active IRF3 is low, but is not required for maximum levels of IFNβ expression late in induction [27],[35]. Consistent with this finding, we show that only a small fraction of the p65 subunit of NF-κB translocates to the nucleus 8 h.p.i., and little difference is observed in NF-κB localization between the YFP-positive and YFP-negative cells (Figure 3A).
The observation that IRF3 activation and translocation occurs in only a fraction of virus-infected cells suggests that upstream components in the RIG-I signaling pathway differ in IFNβ-producing and nonproducing cells. Western blotting results (Figure 3C) showed that IFNβ-producing cells have higher levels of both RIG-I and MDA5 than the nonproducing population. Trim25, an E3 ligase required for RIG-I activation [4], is also present at a higher level in the IFNβ-producing cells (Figure 3C). The increase in protein levels appears to be a consequence of differential transcription of the tested genes, as mRNA levels of all three genes are higher in IFNβ-producing cells (Figure 3D). We conclude that the IFNβ-producing cells have higher levels of essential RIG-I signaling pathway components than the IFNβ-nonproducing cells. Thus, at least part of the observed stochastic expression is due to limiting RIG-I pathway components in the cells that do not express IFN.
By contrast to the RNA detectors, the protein levels for both MAVS and TBK1, two essential components of the RIG-I signaling pathway [7],[9], were lower in the IFNβ-producing cells (Figure 3C). However, this is likely due to the degradation and/or cleavage of the MAVS protein in infected cells [36]–[38]. The data of Figure 3C suggest that TBK1 is also targeted for degradation during virus infection, consistent with the observation that TBK1 is subject to proteasome-dependent degradation [39]. Thus the turnover of both MAVS and TBK1 may be required for the post-induction turn-off of IFNβ gene expression [38].
We have shown that the RIG-I signaling pathway is selectively activated in IFNβ-expressing cells, and this is due only in part to the cell-to-cell differences in virus infection/replication. Our results also suggest that IFNβ-producing cells have a more potent signaling pathway than IFNβ-non-expressing cells. To further explore this possibility, we established a series of L929 stable cell lines that express RIG-I, MDA5, or Trim25 under the control of a tetracycline-inducible promoter (Figure S5A). As shown in Figure S5B and S5C, high levels of exogenous RIG-I only slightly increased the percentage of IFNβ-producing cells. A larger increase was observed with MDA5 and Trim25, but the final percentage in both cases was still under 30%. Thus, these upstream components appear to be among several limiting factors in the cell population.
Additional components in the RIG-I signaling pathway were tested using the same approach, and high percentages of IFNβ-producing cells were observed (Figure 4A and 4B). While a large difference between tetracycline-negative and tetracycline-positive cells was observed with the TBK1 line, only a small difference was observed between the corresponding MAVS lines. However, a large difference was observed between the non-transformed and transformed MAVS lines, suggesting that a low level of leaky transcription in the MAVS line is sufficient to dramatically increase the number of IFNβ-expressing cells. These data clearly indicate that both MAVS and TBK1 are limiting components in the RIG-I pathway and therefore contribute significantly to stochastic IFNβ expression.
We have shown that over-expression of RIG-I or Trim25 alone only slightly increases the percentage of IFNβ-producing cells, but it is possible that both must be expressed to achieve maximum levels of IFNβ production. We therefore transfected RIG-I stable transfectants with a Trim25 expression plasmid, and the other way around. The cells were then induced with tetracycline, infected with SeV, and examined for IFNβ mRNA expression. Control experiments using a GFP reporter indicated that under our experimental conditions approximately 70% of cells can be transfected with the second plasmid (Figure S5D). As shown in Figure 4C and 4D, a dramatic increase was observed only 6 h.p.i. when either the RIG-I or Trim25 lines were transfected with Trim25 or RIG-I, respectively. This observation was confirmed by carrying out intracellular staining and flow cytometry experiments using IFNβ/YFP homozygous MEFs (Figure S6). We conclude that the combination of RIG-I and Trim25 is limiting in the RIG-I pathway.
We note that the increase of IFNβ-expressing cells was not observed in uninfected cells, with the only exception being MAVS. Thus, over-expression of these signaling components did not bypass the requirement for signaling pathway activation.
Expression of the IFNβ gene requires an active RIG-I signaling pathway and assembly of the enhanceosome complex on the IFNβ promoter. To investigate whether individual enhanceosome components are limiting factors, we established a series of tetracycline-inducible L929 stable lines that express IRF3, IRF7, or p65 genes. Figure 5A and 5B show that, without tetracycline induction, only 10%–15% of the cells produce detectable levels of IFNβ mRNA in response to virus infection. Remarkably, the percentage of IFNβ-producing cells upon SeV infection increased to 85% when IRF7 expression was induced by tetracycline in every cell (Figure S7A). A smaller increase (55%) was observed when IRF3 was over-expressed, whereas increasing the concentration of NF-κB had little effect, consistent with the data in Figure 3A, and previously published studies [27]. Interestingly, IRF7 over-expression also significantly increased the percentage of IFNα-producing cells after virus infection (Figure S7B and S7C). It is known that IRF7 is required for maximum induction of type I IFN genes [25], and its basal protein level is very low in most cell types except for plasmacytoid dendritic cells [26],[40]. We conclude that IRF7 is a critical limiting factor that is a major contributor to stochastic expression of mouse IFNα and β genes. This conclusion is also supported by our ISH results from 4E-BP1/4E-BP2 double-knockout MEFs (Figure 5C and 5D). Previous studies have indentified 4E-BPs as negative regulators of type I IFN production via translational repression of IRF7 mRNA [41]. As shown in Figure 5C and 5D, we observed a 4-fold increase of the percentage of IFNβ-expressing cells in 4E-BP1/4E-BP2 double-knockout MEFs compared to wild-type MEFs, consistent with the conclusion that a limiting amount of IRF7 is a major contributor to the stochastic expression of IFNβ.
We also found that type I IFN induction was exceptionally high, with much faster kinetics in cells expressing exogenous IRF7 than in control cells (Figure S7D). In the absence of tetracycline induction, low levels of IFNβ, IFNα4, and IFNα2 mRNA were first detected 6 h, 9 h, and 12 h.p.i., respectively. When the cells were treated with tetracycline, the kinetics of IFN gene transcription changed significantly. IFNβ, IFNα4, and IFNα2 transcripts could be detected as early as 4 h after virus infection. Even at 24 h.p.i., steady and robust transcription of these genes could still be detected. These observations are consistent with a model in which IRF3 is normally activated early for IFN gene induction. Later, higher levels of IRF7 are produced by IFN and are required for both IFNβ and IFNα gene expression, but IRF7 is rapidly turned over, leading to the cessation of both IFNβ and IFNα gene expression [1],[25],[26]. By contrast, in the presence of excess IRF7 in the tetracyline-activated cells, both IFNβ and IFNα are activated earlier, and continue to be expressed because of the continuous presence of IRF7.
We have shown that over-expression of IRF7 or both RIG-I and Trim25 almost completely eliminates stochastic IFNβ expression (Figures 4C, 4D, and 5). To investigate the connection between these observations, we carried out microarray analysis to compare genome-wide expression profiles of L929-IRF7 stable transfectants treated with or without tetracycline. Interestingly, upon IRF7 over-expression, only two up-regulated signaling pathways were identified from the KEGG Pathway Database, and the RIG-I-like receptor signaling pathway is the most up-regulated (p = 3.6E-06) (Figure S8A and S8B) [42]. We did not identify signaling pathways that were similarly enriched among the down-regulated genes. Using qPCR, we confirmed that the mRNA levels of both RIG-I and Trim25 were higher in IRF7 over-expressing cells (Figure S8C). Considering the low basal expression level of IRF7, we conclude that a high level of IRF7 protein increases the percentage of IFNβ-expressing cells not only by increasing its own abundance, but also by up-regulating the RIG-I signaling pathway to increase the potency of activation of the IFNβ gene.
IFNβ gene expression can also be induced by transfection of the synthetic dsRNA polyriboinosinic polyribocytidylic acid (poly I∶C), and this induction occurs mainly through the MDA5 signaling pathway [43]. Early studies revealed that induction of IFNβ expression by dsRNA treatment is also stochastic [13],[14]. We therefore asked whether stochastic IFNβ gene expression induced by dsRNA is due to cell-to-cell variation in the levels of MDA5 and IRF7. Using FACS analysis, we found that poly I∶C–induced IFNβ expression is also stochastic (Figure 6A). When IFNβ/YFP homozygous MEFs were electroporated with Cy5-labeled poly I∶C, only 9% of the cells produced IFNβ as detected by the presence of YFP. However, the electroporation efficiency was over 99% (Figure 6A, left panel). Interestingly, based on the Cy5 intensity, there were two populations of cells, which contained different amounts of poly I∶C. When we gated these two populations out as “poly I∶C–high” and “poly I∶C–low”, we observed that the “poly I∶C–high” population included more cells producing IFNβ (Figure 6A, right panel), indicating that the amount of inducer does affect the extent of stochastic IFNβ expression. However, only a small percentage of “poly I∶C–high” cells expressed the IFNβ gene, clearly indicating that other limiting factor(s) dominate the stochastic IFNβ expression induced by poly I∶C transfection. We therefore carried out experiments to identify these limiting components.
L929-MDA5 and L929-RIG-I stable transfectants were transfected with poly I∶C followed by ISH to detect IFNβ expression. As shown in Figure 6B and 6C, over-expression of RIG-I only slightly increased the percentage of IFNβ-producing cells. By contrast over-expression of MDA5, the major cytoplasmic receptor for poly I∶C, led to a substantial increase in the percentage of IFNβ-producing cells (from 15% to 65%). Considering that the transfection efficiency is approximately 75% (data not shown), over-expression of MDA5 basically eliminates stochastic expression of the IFNβ gene in response to poly I∶C transfection. Furthermore, the results of the flow cytometry experiment also supported this conclusion. As shown in Figure 6D, after 8 h of poly I∶C stimulation, we observed approximately 2.6% YFP-positive cells. Within this population, about 70% of the YFP-positive cells had higher levels of MDA5 protein (1.86% out of 2.67%). We note that the percentage of YFP-positive cells is much lower than that observed with virus infection (Figures 2D and S6).
Over-expression of IRF3 or IRF7 also increased the percentage of IFNβ-producing cells in response to poly I∶C (Figure S9A and S9B). As shown in Figure S8, over-expression of the IRF7 gene up-regulates MDA5 gene expression. Considering its low basal expression level, IRF7 is also an important limiting factor in stochastic IFNβ expression induced by poly I∶C transfection. Taken together, these data show that poly I∶C–induced stochastic IFNβ expression depends on the abundance of both poly I∶C and signaling pathway protein MDA5 as well as IRF3/IRF7, which is similar to what was found in the case of virus infection.
We also asked whether the concentrations of proteins regulating IFNβ expression are sufficiently different from cell to cell to account for the stochastic IFNβ expression. Using flow cytometry, we measured the distributions of six components in the RIG-I signaling pathway for which specific antibodies are available. As shown in Figure 7A and 7B, all six proteins were log-normally distributed across the population. Quantitative immunofluorescence data for individual components show similar distributions of each factor at the single-cell level (Figure S10). Combined with our previous data, these observations suggest that naturally occurring differences in the protein levels of signaling pathway components are the primary cause of cell-to-cell variability in IFNβ expression upon virus infection.
When IFN is secreted from virus-infected cells in vivo, it binds to type I IFN receptors on surrounding cells and activates a large set of genes encoding antiviral proteins (interferon-stimulated genes [ISGs]) via the Jak/STAT signal transduction pathway. We therefore carried out experiments to determine whether the induction of antiviral ISGs is also stochastic. As shown in Figure 7C, ISG15 is expressed in all cells upon treatment with IFNβ. Thus, when IFN is secreted, all of the surrounding cells produce antiviral proteins. This result is also consistent with previous observations showing that the antiviral response induced by IFN is a robust feature common to all cells, and is independent of the stochastic expression of IFN receptor IFNAR [44].
Regulation of type I IFN production is essential for the innate immune response to viral infections [45],[46]. However, high levels of IFNβ can be toxic [47],[48]. Thus, IFNβ production must be tightly regulated. This regulation appears to be both temporal and stochastic. Type I IFN genes are tightly repressed prior to virus infection, activated upon infection, and then rapidly turned off several hours later (Figure S1B and S1C). Previous studies of several cytokine genes suggest that this stochastic gene expression provides an additional mechanism of regulation whereby optimal levels of cytokine production are determined by the frequency of expressing cells rather than by protein levels per cell [18],[19],[49]. Thus, it is possible that stochastic expression is a primary mechanism for controlling the optimal level of IFNβ production in vivo. In particular, we have shown that while IFN production is stochastic, the activation of the antiviral gene program by secreted IFN is not. Thus, stochastic expression of IFN would allow the regional distribution of the cytokine and activation of the surrounding cells, without producing toxic levels of IFN.
Previous studies have implicated as limiting steps enhanceosome assembly [20],[21] and the assembly of an interchromosomal transcriptional hub formed through interactions between Alu elements bearing NF-κB sites [20]. More recently, the infecting virus, rather than intrinsic properties of the infected cell, has been implicated in this stochasm [22]. The data presented here reveal a far more complex mechanism in which cell-to-cell variations in limiting components required to support viral replication, to detect and signal the presence of viral RNA, and to activate transcription factors all contribute to the observed stochastic expression (Figure 7D). It seems likely that the key limiting factor varies between cell types, cell lines, and organisms.
The earliest step in the virus induction signaling pathway is entry of virus or dsRNA into the cell. We have shown that both inducers elicit stochastic expression, but in neither case is this due to limiting inducer (Figures S4B and 6A). We showed that both IFNβ-producing and nonproducing cells were infected by SeV (Figure 2B). However, the IFNβ-producing cells contained significantly higher levels of the products of viral replication and transcription. Thus, it appears that there are cell-to-cell differences in the ability to support efficient viral replication, and these differences influence the probability of IFNβ gene expression. Presumably, high levels of RNA inducer in the IFN-producing cells overcome limiting amounts of RIG-I or MDA5. However, differences in viral replication alone cannot explain the observed stochasm in IFNβ production. A previous study, using a cell line transfected with an IFNβ-GFP reporter, concluded that stochastic IFNβ expression is due entirely to heterogeneity in the infecting virus [22]. However, in that study the IFNβ-GFP cell line was preselected to minimize stochastic expression of the reporter. In addition, that study involved a stably transfected gene, while the present study made use of the endogenous gene. The results presented here strongly indicate that heterogeneity of both the virus and host cells together are responsible for the stochastic expression of IFNβ.
We have identified multiple limiting steps in the activation of IFNβ gene expression, ranging from initial steps in virus infection and replication, to the signaling pathway, to the activation and binding of transcriptional activator proteins to the IFNβ promoter. For example, over-expression of individual components in the RIG-I signaling pathway increases the percentage of IFN-expressing cells. The largest increase was observed with IRF7, which lies at the endpoint of the RIG-I pathway, and also positively controls the expression of components in the RIG-I signaling pathway. Taken together, these data are consistent with a model in which the probability of expression of the IFNβ gene in individual cells depends primarily on the activation of the RIG-I signaling pathway and the presence of sufficient numbers of IRF7 molecules to activate transcription (Figure 7D). This conclusion is consistent with the observation that both IFNβ and IFNα are stochastically expressed in response to virus infection (Figure 1A and 1C). The expression of both genes requires activation of the RIG-I pathway and active IRF7 [50].
We find that limiting amounts of other RIG-I pathway components also contribute to stochastic expression of the IFNβ gene, as we observed higher levels of RIG-I/Trim25 and MDA5 mRNA and protein levels in the IFNβ-producing cells than in the nonproducers (Figure 3). In addition, over-expression of RIG-I and Trim25 together leads to a dramatic increase in the percentage of cells that express IFNβ (Figure 4C and 4D). Similar results were obtained with high levels of expression of the RIG-I signaling components MAVS and TBK1 and the transcription factors IRF3 and IRF7 (Figures 4A, 4B, and 5). Thus, it appears that many, if not all, of the components in the RIG-I signaling pathway, from the sensors of viral RNA to the essential transcription factors, can be limiting components in the virus induction pathway.
The largest increase in the percentage of IFN-producing cells was observed when IRF7 was over-expressed. IRF7 is the master regulator of type I IFN gene expression [25], and is present at low levels in all cell types except plasmacytoid dendritic cells, where it is constitutively abundant [26],[40]. Our over-expression experiments show that high levels of IRF7 promote the transcription of type I IFN genes (Figure S7D), and essentially eliminate the stochastic expression of both the IFNβ and α genes (Figures 5 and S7). In a previous study in human cells, both NF-κB and IRF7 over-expression was shown to partially suppress stochastic IFNβ expression [20]. Our results are consistent with this observation. However, there are two differences. First, based, at least in part, on the lack of requirement of NF-κB in murine cells, we observed a relatively small effect of increasing NF-κB expression. Second, we saw a greater effect of IRF7 expression in murine cells than was observed in human cells. Over-expression of IRF7 in L929 cells almost completely eliminated stochastic expression of both IFNβ and α genes, while in human HeLa cells high levels of IRF7 increase the percentage of IFNβ-producing cells to almost 55% [20]. Deleting the IRF7 translational repressors, 4E-BPs, also increased the IFNβ-expressing MEFs by 4-fold (Figure 5C and 5D). We also showed that the RIG-I signaling pathway, and in particular RIG-I and Trim25, are up-regulated in IRF7 over-expressing cells (Figure S8). We conclude that limiting amounts of active IRF7 appear to be overcome by two mechanisms: positive auto-regulation of IRF7 expression, and IRF7-dependent up-regulation of the RIG-I signaling pathway.
We note that in addition to IFNβ, several other virus-inducible genes, including TNFα, IL-6, CCL4, and CCL5, are highly expressed in the IFNβ-producing cells compared to nonproducers, suggesting that many, if not all, of the virus-inducible genes are stochastically expressed. The common feature of the activation of all of these genes is that they all require the RIG-I signaling pathway [28]–[31]. Thus, we conclude that stochastic gene expression is primarily due to limiting components in the signaling pathway but not gene-to-gene variation in the mechanism of gene activation.
We showed that although the IFNβ gene is stochastically expressed upon virus infection, the antiviral ISGs, e.g., ISG15, were equally induced in all cells (Figure 7C). However, we note that RIG-I, Trim25, and MDA5, which are also antiviral ISGs, are highly expressed in IFNβ-producing cells compared to nonproducing cells (Figure 3C and 3D). We believe that the differences we observed here reflect naturally occurring cell-to-cell variability in the levels of expression of these genes prior to virus infection, and that this variability is the primary source of stochastic IFNβ gene expression. However, at later times after virus infection, we expect that the differences in the mRNA or protein levels of these genes between the YFP-positive and YFP-negative populations will be much smaller compared to those at earlier stages (8 h.p.i.). As shown in Figure S11A and S11B, our qPCR data and Western blot data support this expectation. The IFNβ gene is also stochastically expressed in IFNAR-deficient MEFs, which suggests that the IFNAR levels or an IFNβ feedback loop are not major factors responsible for stochastic IFNβ gene expression (Figure S11C). We further measured the distributions of six components in the RIG-I signaling pathway. As shown in Figures 7A, 7B, and S10, all six proteins were log-normally distributed across the cell population, an observation that is consistent with data on other proteins [51],[52]. Thus, naturally occurring differences in the protein levels and activities of individual signaling pathway components and transcription factors account for stochastic IFNβ expression induced by both poly I∶C induction and virus infection.
Previous studies have shown that naturally occurring differences in the levels of proteins in the apoptotic signaling pathway are the primary reasons for cell-to-cell variability in the probability of cell death [52]. Thus, the results presented here not only reveal the complexity of the regulatory mechanisms controlling stochastic IFNβ gene expression, but also suggest a general mechanism used in different biological processes to establish and control stochastic gene expression. A remarkable feature of stochastic expression is that it appears to be an intrinsic property of different clonal populations of cells. For example, if a particular cell line displays a certain percentage of activated cells, that percentage differs from other cell lines, and is retained when the cells are recloned [14]. Thus, the extent of stochasm appears to be a genetic and epigenetic feature of clonal cell populations.
All cell lines, including L929, RAW 264.7, MG63, and 293T, were from the American Type Culture Center; primary MEFs were isolated using standard protocols from IFNβ/YFP mice [24]. Primary human foreskin fibroblast cells were purchased from PromoCell. All cells were cultured in DMEM (Gibco) supplemented with 10% FBS (Gibco) in a 5% CO2 incubator. Cycloheximide was purchased from Sigma-Aldrich. Human and mouse recombinant IFN proteins were purchased from PBL Interferonsource. Brefeldin A solution was purchased from eBioscience. Poly I∶C was purchased from InvivoGen. Cy5-labeled poly I∶C was generated using Label IT Nucleic Acid Labeling Kit (Mirus). The different expression constructs were generated by cloning the coding sequences of each gene by PCR and inserting them into the vector pt-REX-DEST30, which has the tetracycline-inducible promoter (Invitrogen).
Concentrated SeV stock (Cantell strain, Charles River Lab) was added to cultured cells at a concentration of 200 HAU/ml and incubated for the times indicated. Poly I∶C transfection was carried out using either lipofectamine2000 (Invitrogen) or electroporation using Amaxa MEF2 Nucleofector Kit (Lonza).
Total RNA was extracted with Trizol reagent (Invitrogen). Real-time quantitative reverse transcription PCR (qRT-PCR) was conducted according to standard protocols.
Antibody against YFP was from Chemicon (Millipore) or Abcam. RIG-I, MAVS, and GAPDH antibodies were from Cell Signaling. Antibodies against p65, HDAC1, and Trim25 were from Santa Cruz Biotechnology. MDA5 and TBK1 antibodies were from Abcam and Imgenex, respectively. IFNβ antibody used for FACS was from Millipore. SeV antibodies were kindly provided by Dr. Atsushi Kato (National Institute of Infectious Diseases, Japan). Nuclear/cytosol fractionation was performed using Nuclear/Cytosol Fractionation Kit (BioVision). Western blots were carried out using standard protocols.
Antisense RNA probes recognizing mouse IFNβ or β-actin were synthesized using T7 or SP6 polymerase and digoxigenin-labeled nucleotides (Roche Applied Science). Cells were cultured on poly-D-lysine-coated 24-well plates (Fisher) and either mock- or virus-infected for the times indicated. Cells were then washed twice with PBS and fixed with 4% paraformaldehyde. Hybridization, washes, and staining were carried out as previously described [53].
MEF cells were fixed with IC Fixation Buffer and permeabilized with Permeabilization Buffer (both from eBioscience). After incubation with appropriate antibodies, flow cytometry was done with a FACSCalibur, and data were analyzed with CellQuest software (both from Becton Dickinson).
Total RNA from untreated and tetracycline-induced L929-IRF7 cells were prepared using Trizol reagent (Invitrogen) followed by purification using MEGAclear (Ambion). Biotinylated RNA probes were synthesized by two rounds of amplification using the MessageAmp II aRNA Amplification kit (Ambion). The probes were hybridized with Affymetrix Mouse Genome 430A_2.0 array chips. Affymetrix DAT files were processed using the Affymetrix Gene Chip Operating System to create CEL files. Normalized expression values were analyzed with the Bioconductor Limma package, an approach for implementing empirical Bayes linear modeling [42]. For all comparison tests, genes with an absolute fold change in transcript level exceeding 1.5 and p<0.05 were selected for further analyses. The likelihood of overrepresentation of KEGG signaling pathways in the up- or down-regulated gene list relative to a background of all array genes was calculated by Fisher's exact test for statistical analysis.
|
10.1371/journal.pntd.0001615 | Modeling the Control of Trypanosomiasis Using Trypanocides or Insecticide-Treated Livestock | In Uganda, Rhodesian sleeping sickness, caused by Trypanosoma brucei rhodesiense, and animal trypanosomiasis caused by T. vivax and T. congolense, are being controlled by treating cattle with trypanocides and/or insecticides. We used a mathematical model to identify treatment coverages required to break transmission when host populations consisted of various proportions of wild and domestic mammals, and reptiles.
An Ro model for trypanosomiasis was generalized to allow tsetse to feed off multiple host species. Assuming populations of cattle and humans only, pre-intervention Ro values for T. vivax, T. congolense, and T. brucei were 388, 64 and 3, respectively. Treating cattle with trypanocides reduced R0 for T. brucei to <1 if >65% of cattle were treated, vs 100% coverage necessary for T. vivax and T. congolense. The presence of wild mammalian hosts increased the coverage required and made control of T. vivax and T. congolense impossible. When tsetse fed only on cattle or humans, R0 for T. brucei was <1 if 20% of cattle were treated with insecticide, compared to 55% for T. congolense. If wild mammalian hosts were also present, control of the two species was impossible if proportions of non-human bloodmeals from cattle were <40% or <70%, respectively. R0 was <1 for T. vivax only when insecticide treatment led to reductions in the tsetse population. Under such circumstances R0<1 for T. brucei and T. congolense if cattle make up 30% and 55%, respectively of the non-human tsetse bloodmeals, as long as all cattle are treated with insecticide.
In settled areas of Uganda with few wild hosts, control of Rhodesian sleeping sickness is likely to be much more effectively controlled by treating cattle with insecticide than with trypanocides.
| In Uganda, cattle are an important reservoir for Trypanosoma brucei rhodesiense, the causative agent of Rhodesian sleeping sickness (human African trypanosomiasis), transmitted by tsetse flies Glossina fuscipes fuscipes, which feed on cattle, humans, and wild vertebrates, particularly monitor lizards. Trypanosomiasis can be controlled by treating livestock with trypanocides or insecticide – killing parasites or vectors, respectively. Mathematical modeling of trypanosomiasis was used to compare the impact of drug- and insecticide-based interventions on R0 with varying densities of cattle, humans and wild hosts. Intervention impact changes with the number of cattle treated and the proportion of bloodmeals tsetse take from cattle. R0 was always reduced more by treating cattle with insecticide rather than trypanocides. In the absence of wild hosts, the model suggests that control of sleeping sickness (R0<1) could be achieved by treating ∼65% of cattle with trypanocides or ∼20% with insecticide. Required coverage increases as wild mammals provide increasing proportion of tsetse bloodmeals: if 60% of non-human bloodmeals are from wild hosts then all cattle have to be treated with insecticide. Conversely, it is reduced if lizards, which do not harbor trypanosomes, are important hosts and/or if insecticides are used at a scale where tsetse numbers decline.
| Across sub-Saharan Africa, a variety of Trypanosoma spp transmitted by tsetse flies (Glossina spp) cause human and animal trypanosomiases. There are >10,000 cases/year of Human African Trypanosomiasis (HAT) [1] with an estimated burden of ∼1.3 million Disability Adjusted Life Years (DALYs) [2] and economic losses in excess of $1 billion due to human and animal trypanosomiasis [3]. While interventions can be directed against the vector or the parasite, emphasis has usually been on the use of drugs to treat the disease both in humans and in livestock.
While the importance of treating cases, especially human ones, cannot be overstated, several advances in our understanding of tsetse biology and ecology, and improvements in the cost-effectiveness of tsetse control [4], [5], have revived interest in that approach to disease management. First, the use of satellite navigation as an aid to nocturnal aerial spraying, spraying much larger areas than previously, and protecting the sprayed areas with odor-baited targets, has provided impressive results, such as the eradication of G. m. centralis from Botswana [6]. Second, the demonstration of the importance of odor for host location in some species of tsetse provided a means of attracting them to insecticide-treated targets and, by killing the flies, provided control of cattle and human trypanosomiasis [7]–[10]. Third, the particularly low reproductive rate in tsetse made it possible to use as few as four such targets per square kilometer to eliminate isolated populations of G. pallidipes Austen and two sub-species of G. morsitans [9], [11]. The method is cheaper than aerial spraying and more environmentally friendly than insecticidal ground spraying, game destruction or habitat clearance [11]. Issues of cost, logistics, government commitment, and theft of materials have meant, however, that the approach has not been used in large-scale control programs except in Zimbabwe and in the Western Province of Zambia [11], [12].
Part of the reason for this limited use stems from the fact that, simultaneously with the development of insecticide-treated target technology, it was realized that tsetse control could be achieved equally effectively by applying insecticide to the very livestock - generally cattle - off which the tsetse were feeding. This approach has been used very successfully in areas where tsetse feed predominantly on cattle [13], [14], though it would be less effective in areas where – as in large parts of Zimbabwe and Tanzania – the predominant food source for the tsetse are wild mammals.
Whereas insecticide-treated cattle (ITC) can be used in operations aimed at eliminating tsetse populations, animal trypanosomiasis can also be reduced to low levels even where tsetse populations persist [15]. It is, of course, relief from cattle disease – rather than issues of tsetse fly control versus eradication – which most interests stockholders in tsetse areas and which can be used to interest the stockholder in becoming actively involved in tsetse and trypanosomiasis control [13]. Recent advances in our understanding of the feeding behavior of tsetse on cattle have led to even cheaper methods of tsetse control where the insecticide is applied to the body regions and/or individual animals on which most tsetse feed [16], [17]. This restricted application of pyrethroids is comparable in its cost and simplicity to the widespread use of trypanocides by farmers to prevent or cure trypanosomiasis in their livestock [16].
There are several possible reasons why these advances in affordable, low-technology tsetse control have not, as yet, played a significant role in efforts against HAT. First, there is an imperative to find and treat infected humans and livestock and this approach is thus the foundation of all efforts against the disease. Second, the odor-baited devices used so effectively in efforts against animal trypanosomiasis [10] are less effective against the important vectors of HAT [18], [19]. This poor efficacy is probably related, in part, to the distinctions between the host relationships of the various tsetse species. The important vectors of animal trypanosomiasis, i.e., the Morsitans-group tsetse, feed almost exclusively on mammals (e.g. warthog, kudu, buffalo and cattle) which they locate largely by odor, whereas the Palpalis-group species, which are the main vectors of HAT, are less responsive to odors and include reptiles and birds in their diet. For instance, between 50 and 90% of meals taken by Glossina fuscipes fuscipes are from monitor lizard [20] which themselves do not support all the trypanosome species infective to mammals [21].
In this paper, we investigate the theoretical effects of two different approaches to trypanosomiasis control, both of which have already been shown to be of interest to small-scale stockholders in resource-limited settings [22]. First we consider the effect of treating animals with trypanocides, which prevent the disease without having any insecticidal effect. Second, we consider the use of the ITC method, which has no direct trypanocidal effect but which increases mortality in the vectors. We limit our study to the situation typical of eastern and southern Africa, where Trypanosoma vivax, T. congolense and T. brucei rhodesiense occur in livestock and wildlife - and where the last-named parasite also causes “Rhodesian” sleeping sickness in humans [23], [24].
We generalize the Rogers [25] two-host model for trypanosomiasis to one where a single species of tsetse can feed off any finite number (n) of vertebrate hosts. The formal proof that Rogers' model can be generalized in this way is given in the Supporting Information (Text S1). The overall basic reproductive rate (R0) of a trypanosome species is given by:(1)where D = 1 for T. vivax and T. congolense andfor T. brucei, and where the following definitions apply: R0 = overall basic reproductive rate; formally, in a completely susceptible population, the number of trypanosome-infected tsetse arising from each infected fly; c = P(infected blood meal gives mature infection in fly); u = Daily mortality rate of the flies; T = Incubation period in tsetse (all time units are days); ai = pi/d, where pi = Proportion of tsetse bloodmeals from species i, d = Duration of feeding cycle in flies; bi = P(infected fly bite produces infection in species i); mi = V/Ni, where V = Number of tsetse, Ni = Number of animals of species i, 1/ri = Duration of infection in species i. The parameter D differs between T. brucei and the other species of trypanosomiasis because it is assumed that tsetse can only be infected with T. brucei when they take their first bloodmeal. It is assumed that the probability of infection for the other species is independent of a fly's feeding history: to distinguish this situation Rogers also replaced c with c′ for T. brucei [23]. The default values for the parameters of his two-host model for Rhodesian sleeping sickness [23] are copied here for convenience, in Tables 1 and 2.
We extend the model to consider cases where, in addition to humans and domestic stock (cattle), the following vertebrate species are present: (1) wild mammals; (2) monitor lizards; (3) wild mammals and monitor lizards. The interventions to be considered involve the treatment of cattle with: (1) prophylactic trypanocides that kill trypanosomes but have no effect on tsetse mortality; (2) ITC, i.e., topical application to hosts of insecticides that kill tsetse but have no direct effect on trypanosome mortality.
The use of ITC can reduce R0 in two ways. First, in common with all insecticidal techniques, it reduces the average life expectancy of tsetse, so decreasing the abundance of the flies and the proportion of the population that is old enough to harbor mature, transmissible infections. Second, and in contrast with other insecticidal techniques such as traps or insecticide-treated targets, ITC kills specifically those tsetse that become infected from the reservoir of disease in cattle.
Since the Rogers model assumes that the abundance and age structure of the tsetse population is constant, it is particularly suitable for highlighting the second type of effect, and so for comparing ITC and trypanocides as means of reducing the probability that a fly will become infected. In the present paper we first use the Rogers model to address this matter under circumstances in which various levels of the use of trypanocides or insecticide treatment are applied to cattle that represent different proportions of the overall cattle population, and with host populations composed of various species. We then identify the extra benefit that ITC produces via reductions in the abundance and mean age of the tsetse population, and predict the relative merits of using ITC and trypanocides, as assessed via the model.
As a preliminary check we inserted the published default parameter values (see Tables 1 and 2, [25]) into Equation (1) for the scenario where only (untreated) cattle and humans provided the source of tsetse bloodmeals, and obtained the published values for R0: 388.2 for T. vivax, 64.4 for T. congolense, and 2.65 for T. brucei. The last value is made up the sum of two components, 2.54 from the cattle and 0.11 from humans, implying that T. brucei would not survive in the absence of the cattle reservoir [25]. To control, and eventually eliminate, T. brucei the goal therefore must be to reduce the combined R0, for human and non-human hosts, to a value less than unity.
We now turn to the use of the insecticide-treated cattle (ITC) method of control – where the vectors, rather than the trypanosome, are targeted. In the previous sections we have assumed a fixed daily rate for adult tsetse mortality (Table 1). When considering the use of ITC, however, we need to decompose this factor into the mortality occurring at the time of feeding and that occurring between feeds. The former has generally been considered the dominant component [28], [29] even where the host is not treated with insecticide.
If the probability of surviving a feed is qf and the probability of surviving a non-feeding day is qn then a fly survives a complete feeding cycle of d days with probability qf qnd. With qf = 0.96, qn = 0.98, and with the assumed four-day feeding interval [25], the probability of surviving from one feeding cycle would then be approximately 0.96×0.984 = 0.885 and the daily mortality rate is calculated as −ln(0.885)/4≈0.03, as originally assumed [25].
Where some hosts are treated with insecticide we assume that flies always die if they feed off a treated animal; the probability of a given fly surviving a feed is thus the product of the probabilities that it feeds off an un-treated host and survives that meal. We assume further that flies feed off all cattle at random, particularly with respect to the animal's treatment status. If the proportion of cattle treated is pi then the probability of a fly surviving a feeding cycle is now (1−pi) qf qnd. For example, with the above values for qf, qn and d, and if 10% of the cattle are treated, the survival probability will be 0.9×0.885 = 0.797 and the daily mortality is now approximately 0.057. As a first approximation we ignore any extra mortality arising from a fly feeding off a human, rather than cattle or wildlife.
Figures 1, 2, 3, and 4 provide estimates of the control of trypanosomiasis, by way either of the use of trypanocidal drugs or ITC, in the situation where there is sufficient birth, of uninfected flies, to ensure that the tsetse population stays at a constant level [25]. This should be a reasonable assumption in the case where trypanocidal treatment is used to control trypanosomiasis and there is no imposed mortality on the tsetse population.
When ITC is used, the population could only be kept constant if the increase in mortality is balanced by an increase in birth and/or immigration. If birth is the predominant source of replacements then Figures 3 and 4 reflect the control situation. If, however, the population is kept constant due to immigration then the replacement flies will be predominantly older flies, with above-average probability of being infected with trypanosomes, so that Figures 3 and 4 over-estimate the efficacy of ITC.
However, where ITC is used, either against closed populations of tsetse or on a sufficiently large scale that immigration is limited at sites far from the boundary, the expectation is that the fly population will decrease. Inspection of Equation (1) shows that, other things being equal, R0 changes linearly with the tsetse population so that, where the use of ITC produces a decline in population levels the effect on R0 will be larger than indicated in Figure 3. We follow Smith & McKenzie [31] in estimating that, if mortality was increased from some value u to u′, the initial vector population (V) would decrease to Vu/u′.
Taking this factor into account changes the threshold value for the required percentage of cattle among non-human hosts. Thus, under the assumption of a constant tsetse population, it was impossible to force R0<1 for T. vivax (Figures 3A, 4A, 5). However, if tsetse populations are reduced as a consequence of ITC, R0<1 for T. vivax as long as cattle make up >90% of the non-human hosts (Figure 5). The proportions of cattle among non-human hosts, required to force R0<1, declines from roughly 70% to 55% for T. congolense and 40% to 30% for T. brucei (Figure 5).
For purposes of comparing our results with previous work we have, initially, adhered closely to the design, and the parameterization, of the Rogers model – which provides a useful tool for investigating the dynamics of trypanosomiasis. It is recognized, however, that some fundamental details of the model can be improved. For example, the model makes no distinction between male and female tsetse, which are known to differ with respect to longevity, mobility, infectivity and responses to baits [32], and does not allow that mortality changes as a function of age [33].
Moreover, advances in our knowledge over the past 23 years allow the selection of parameter values that better reflect the field situation. Thus, the feeding interval is certainly shorter than four days and where tsetse make more than one visit to a host per feeding cycle [30], [34] this will impact on both the probability that they transmit a trypanosome, and the probability that they are killed when they alight on an animal that has been treated with insecticide.
Most seriously, the model assumes that the abundance and age structure of the tsetse population is constant. This can be a reasonable assumption where no tsetse control efforts are in place, or when trypanosomiasis control consists simply of treating livestock with trypanocides that have no insecticidal effect. If cattle provide a substantial proportion of tsetse bloodmeals and if a significant proportion of these cattle are treated with insecticide, however, it may be expected that both the size of the population in the area under treatment, and its mean age, will tend to decline. On the other hand the model also ignores the problem of invasion from adjacent infected areas and this further complicates the estimation of the effect of ITC.
Finally, we have not modified Rogers' implicit assumption that tsetse feed at random off the individuals of a given host species. This is known not to be the case and this consideration will complicate the modeling [17]. Nonetheless, in the limit, where some individuals provide no bloodmeals at all for tsetse, they effectively do not exist from the modeling point of view. One could thus simplify the problem by considering the “effective” number of individuals in a herd – being the numbers that do provide bloodmeals. In the same way, baboons and impala – which provide almost no bloodmeals for tsetse – do not need to be considered when modeling the dynamics of trypanosomiasis.
It would not be easy to incorporate all of these details into the present model and still maintain the simplicity that allowed the model to be generalized to apply to the variety of situations considered here. The more general model can, however, be investigated using simulation models; the results of such an exercise will be reported in a separate paper.
Despite the above limitations, the theoretical development presented here suggests that the use of ITC should provide a potent tool for controlling, or even eliminating, trypanosomiasis in situations where cattle provide the majority of bloodmeals for tsetse. The dynamics of transmission ensure that the requisite proportion favoring the use of ITC depends on the species of trypanosome involved; for T. vivax there is little hope of eliminating the disease unless at least 90% of the tsetse bloodmeals are from cattle – and then only if insecticide treatment is such that all tsetse feeding off cattle are killed, and if the situation is such that the increased tsetse mortality results in a decline in the fly numbers. For T. brucei the situation is very much more favorable; even if 70% of bloodmeals are being taken from wildlife, treatment with insecticide of the cattle providing the remaining meals from non-humans allows R0 to be reduced to unity. The situation for T. congolense is intermediate between these extremes. By contrast, the use of trypanocides will never allow T. vivax and T. congolense to be eliminated, even where tsetse feed only on cattle – unless all animals are kept permanently on a perfect trypanocide. T. brucei could be controlled – but only in the absence of wildlife hosts.
The classical Rhodesian sleeping sickness foci are often associated with protected areas [35], the vectors are Morsitans-group tsetse and the hosts are wild mammals such as warthog, buffalo and bushbuck. Tackling these foci is very difficult: block treatment of wild hosts with trypanocides is impossible and hence vector control is the only option. Moreover, the flies are highly mobile [36] and widely dispersed across a range of habitats and hence, to be effective, tsetse control must be applied across the entire protected area. This approach is illustrated by the use of aerial spraying and insecticide-treated targets to eliminate tsetse from the Okavango Delta (area≈15,000 km2) of Botswana [6]. Few countries have the resources for such large-scale interventions and hence sleeping sickness persists in parts of east and southern Africa.
By contrast, tackling Rhodesian sleeping sickness transmitted by G. fuscipes might be more tractable for several reasons. First, the underlying R0 of T. b. rhodesiense is likely to be low. Studies of the diet of G. f. fuscipes in Uganda and Kenya have shown that monitor lizards (Varanus nilotica) provide between ∼50% and >90% of bloodmeals [20], [37]–[39] and it seems likely that poikilothermic hosts such as monitor lizards will not be competent hosts for mammalian trypanosomes. The only published study [21] confirms this for T. congolense and the results for T. brucei are equivocal but not compelling: no human-infective trypanosome has been recovered from a lizard, only one wild lizard (N = 46) has been found with T. brucei s.l., and experimental infections of captive lizards – which were not subject to the range of temperatures found in nature – produced, at most, a low and transient parasitaemia. Our results suggest that if lizards are indeed refractory to mammalian trypanosomes and form >80% of the diet of tsetse, then the R0 for T. b. rhodesiense is less than 1. Hence we might expect that Rhodesian sleeping sickness will be associated with areas where lizards are not abundant such as away from the shores of Lake Victoria and/or in densely settled areas where wild hosts are absent. Consistent with this hypothesis, the current foci of Rhodesian sleeping sickness in Uganda are, paradoxically, not near the shores or islands of Lakes Victoria and Kyogu, where tsetse are abundant, but rather at sites further inland [40], [41].
In areas where lizards are not important hosts, then livestock, particularly cattle, are important [20], [38]. In the case of Uganda, the densities of cattle frequently exceed 50 head/km2 [42] and the degraded environment leads to relatively low densities of tsetse [38]. Increasing the host∶vector ratio reduces R0: for densities of 10 host/km2 and 5000 tsetse/km2 our model (with other parameters as in Tables 1 and 2) suggests R0 = 13 for T. brucei; with 50 hosts and 5000 tsetse/km2 the value is 3, and with 50 hosts and 500 tsetse/km2 it is 0.3.
Second, G. f. fuscipes are restricted to riverine habitats and are less mobile than Morsitans species such as G. pallidipes [36] and hence vector control can be applied on a smaller scale, focused on riverine and lacustrine habitats.
Third, the abundance of cattle in settled areas, their importance as a host for tsetse and their need for water – and hence daily presence in the riverine and wetland habitats where G. f. fuscipes is concentrated – means that insecticide-treated cattle should be particularly effective baits. Hence, SE Uganda, the place where Rhodesian sleeping sickness is most serious, accounting for over half (2848/5086) of all cases across Africa [35], is probably the easiest to tackle.
Present evidence for the superior efficacy of ITC assumes greater importance due to indications over the last decade that the economy of this technique can be improved substantially, with no material loss of performance. The application of insecticide can be restricted to the legs and belly of cattle where most tsetse feed, thereby reducing the material costs of treatment by ∼90% [16]. In addition, since most tsetse feed on the larger and older animals within a herd [17], [43], only these animals need be treated, with further savings in cost. As a consequence, the annual material cost of ITC is reduced to <US$2 per beast per year [44] – comparable to the cost of a single dose of diminazene aceturate to cure trypanosomiasis. The restricted application of pyrethroids to older cattle allows young stock to be exposed to ticks and hence develop a natural immunity to tick-borne diseases [45] and reduces impact on dung fauna [46], [47] which play an important role in maintaining soil fertility and, ultimately, productive pasturage. Against these favorable indications for the usefulness of ITC there is the problem that the technique can be used only in districts where cattle occur, although modeling suggests that ITC can be effective even when cattle are distributed patchily, i.e., absent from bands of habitat up to several kilometers wide [48].
Nonetheless, for the densely-settled rural areas of central and southern Uganda where Rhodesian sleeping sickness is most acute, our findings suggest that relatively modest levels of treatment (∼20% even if tsetse numbers are not reduced by the intervention) could lead to the elimination of HAT. Hence there is the exciting prospect that an important public health benefit might arise through the private actions of livestock keepers using cheap, simple and environmentally-benign methods to control vector-borne diseases in their livestock [22].
|
10.1371/journal.ppat.1000224 | snoRNA, a Novel Precursor of microRNA in Giardia lamblia | An Argonaute homolog and a functional Dicer have been identified in the ancient eukaryote Giardia lamblia, which apparently lacks the ability to perform RNA interference (RNAi). The Giardia Argonaute plays an essential role in growth and is capable of binding specifically to the m7G-cap, suggesting a potential involvement in microRNA (miRNA)-mediated translational repression. To test such a possibility, small RNAs were isolated from Giardia trophozoites, cloned, and sequenced. A 26-nucleotide (nt) small RNA (miR2) was identified as a product of Dicer-processed snoRNA GlsR17 and localized to the cytoplasm by fluorescence in situ hybridization, whereas GlsR17 was found primarily in the nucleolus of only one of the two nuclei in Giardia. Three other small RNAs were also identified as products of snoRNAs, suggesting that the latter could be novel precursors of miRNAs in Giardia. Putative miR2 target sites were identified at the 3′-untranslated regions (UTR) of 22 variant surface protein mRNAs using the miRanda program. In vivo expression of Renilla luciferase mRNA containing six identical miR2 target sites in the 3′-UTR was reduced by 40% when co-transfected with synthetic miR2, while the level of luciferase mRNA remained unaffected. Thus, miR2 likely affects translation but not mRNA stability. This repression, however, was not observed when Argonaute was knocked down in Giardia using a ribozyme-antisense RNA. Instead, an enhancement of luciferase expression was observed, suggesting a loss of endogenous miR2-mediated repression when this protein is depleted. Additionally, the level of miR2 was significantly reduced when Dicer was knocked down. In all, the evidence indicates the presence of a snoRNA-derived miRNA-mediated translational repression in Giardia.
| Gene regulation in Giardia lamblia, a primitive parasitic protozoan responsible for the diarrheal disease giardiasis, is poorly understood. There is no consensus promoter sequence. A simple eight–base pair AT-rich region is sufficient to initiate gene transcription in this organism. Thus, the main control of gene expression may occur after the stage of transcription. The presence of Dicer and Argonaute homologs in Giardia suggested that microRNA (miRNA)-mediated translational repression could be one mechanism of gene regulation. In this work, we characterized the presence of the miRNA pathway in Giardia as well as identified the novel use of small nucleolar RNA (snoRNA) as miRNA precursors. Potential target sites for one small RNA (miR2) were identified with the miRanda program. In vivo reporter assays confirmed the specific interaction between the target sites and miR2. A ribozyme-mediated reduction of Dicer and Argonaute in Giardia showed that the former is required for miR2 production whereas the latter functions in mediating the inhibition of reporter expression, which agrees with the roles of these two proteins. This is the first evidence of miRNA-mediated gene regulation in Giardia and the first demonstration of the use of snoRNAs as miRNA precursors.
| The role of small non-coding RNAs in gene regulation has been extensively studied in recent years [1]. MicroRNAs (miRNA) are a major class of small RNAs that are involved in gene regulation via a translational repression mechanism. They play important roles in regulation of genes involved in development [2], cell differentiation [3], and cell maintenance [4]. In higher eukaryotes, maturation of miRNAs from the initial RNA Polymerase II transcripts requires the actions of several proteins. Drosha, a nuclear endoribonuclease III, is known to cleave the primary-miRNAs to produce pre-miRNAs [5]. Exportin5 is responsible for exporting the pre-miRNAs out of the nucleus [6],[7]. Dicer, a cytoplamic endoribonuclease III, cleaves the pre-miRNAs to produce mature miRNAs [8]. Argonaute, which is a major component in the RNA-induced silencing complex (RISC), binds to the mature miRNA [9],[10]. An imperfect complementation between the miRNA incorporated into the RISC complex and its target site located at the 3′-untranslated region (UTR) of mRNA results in translational repression [11],[12]. One possible mechanism of repression involves binding of the Argonaute in the RISC complex to the 7-methylguanosine (m7G) cap of the mRNA resulting in inhibition of translation initiation [13].
Giardia lamblia is a unicellular and binucleated protozoan responsible for giardiasis in humans [14]. Phylogenetic analysis has classified it as one of the earliest branching eukaryotes with many primitive features [14]. In higher eukaryotes, gene expression is highly regulated both transcriptionally and translationally. In Giardia, however, few consensus promoters have been identified and an 8 bp AT-rich region was sufficient to initiate transcription [15]. Additionally, Giardia mRNAs have exceedingly short 3′ and 5′-UTRs, thus greatly reducing the availability of regulatory sites for translational regulation. For instance, ribosomal scanning, an essential mechanism for translation initiation in higher eukaryotes and yeast, is absent from Giardia [16]. Therefore, Giardia represents a unique model for studying the evolution of eukaryotic translational regulation.
No RNA interference (RNAi) has been identified in Giardia in spite of repeated trials by several laboratories in the past (data unpublished). An analysis of the Giardia genome showed no homolog of Drosha or Exportin5. It, however, identified a Dicer (XP 001705536) and an Argonaute homolog (XP 001707926). Giardia Dicer is the only Dicer protein whose three-dimensional structure has been resolved by X-ray crystallography [17]. It was shown to cleave double-stranded RNA (dsRNA) in vitro and support RNAi in a Schizosaccharomyces pombe Dicer deletion mutant [17]. Giardia Dicer is thus likely a bona fide functional Dicer. Though the function of Giardia Argonaute-like protein (GlAgo) remains largely unexplored, antisense-ribozyme-mediated knockdown of this protein inhibited cell growth and purified recombinant GlAgo can bind specifically to m7G-cap-Sepharose (see below). These data raised the possibility that the Argonaute and Dicer in Giardia may be involved in miRNA-mediated translational repression.
Abundant antisense RNAs, up to 20% of the total mRNAs, and small RNAs have been identified in Giardia [18],[19]. About 20 snoRNAs have also been identified in Giardia [20]. Previous efforts at cloning small RNAs from Giardia have resulted in identifying small RNAs homologous to the telomeric retroposons, which were postulated to function in silencing retroposons [19]. In our current study, we isolated, cloned and sequenced small RNAs from Giardia and identified some of the known snoRNA sequences among them. One of the small RNAs, miR2, was identified as a Dicer-digested product from GlsR17, previously identified as a box C/D snoRNA in Giardia [20]. Putative target sites for miR2 were identified at the 3′-UTRs of many variant surface protein (VSP) mRNAs. Expression of a reporter mRNA carrying these putative target sites was specifically inhibited by miR2 without affecting the mRNA level. Subsequent analysis also indicated the dependence of this inhibition on the presence of Argonaute, thus verifying the ability of a snoRNA derived miRNA to function in miRNA-mediated translational repression in Giardia.
To find out if GlAgo plays a critical role in the proliferation of Giardia, mRNA encoding this protein was knocked down by expressing an antisense-ribozyme RNA in giardiavirus-infected Giardia trophozoites [21]. Quantitative RT-PCR indicated that the mRNA was reduced by 50% in the transfected cells (data not shown). Growth of the knockdown cells was inhibited, reaching only 58% of the wild type level after 4 days of cultivation (Figure S1A), suggesting that GlAgo plays an important role in Giardia growth.
Recent studies of human Argonaute have identified the presence of a cap-binding motif in the protein [13], which enables it to compete with eIF4E for binding to the m7G cap in mRNA that may explain a part of the mechanism involving Argonaute in miRNA-mediated translational repression [13]. To test whether this cap-binding motif is also present in GlAgo, m7G-Sepharose was incubated with recombinant GlAgo purified from E. coli. The majority of the GlAgo was found in the flow-through suggesting an excessive loading of the recombinant protein to the beads. After extensive washing with the binding buffer and buffer containing 0.1 mM GTP to remove non-specific binding, the protein was specifically eluted off the Sepharose beads with m7GpppG (Figure S2), suggesting that GlAgo binds to the m7G cap in a highly specific manner, which has been demonstrated to be the cap of Giardia mRNA [22]. Therefore, GlAgo may function in miRNA-mediated translational repression in Giardia by competing with eIF4E for binding to the mRNA m7G-cap [13].
To verify if miRNA-mediated translational repression is functional in Giardia, total RNA was isolated from cultures of Giardia WB trophozoites and size fractionated for small RNAs of <40 nucleotides (nts). Isolated small RNAs were cloned using RNA linkers that require the presence of a 5′-phosphate in the small RNA to be ligated by T4 RNA ligase [23]. Therefore, all the cloned small RNAs should contain a 5′ phosphate, which is a known characteristic of Dicer processing product [23]. A library of 101 clones with unique sequences ranging in sizes between 20 and 34 nts was created with the following distributions: 3% 20 nts, 11% 21 nts, 12% 22 nts, 8% 23 nts, 6% 24 nts, 7% 25 nts, 14% 26 nts, 6% 27 nts, 5% 28 nts, 5% 29 nts, 9% 30 nts, 9% 31 nts, 5% 32 nts, and 1% 34 nts. Interestingly, the size distribution results in a peak at 26 nts, which is the expected size of Giardia Dicer cleaved products based on the crystal structure [17]. Of these sequences, 11 were found to be identical to other clones in the library but with longer nucleotide extensions at the 3′ ends. The shorter RNAs were presumed to be degradation products and were discarded. Potential origins of the small RNAs were identified by BLAST search analysis of the Giardia genome database [24]. The results showed that 81 of the 101 clones were fragments of ribosomal RNA or tRNA and were also discarded. Of the remaining clones, 15 were fragments of open reading frames, 4 were fragments of 3 different box C/D snoRNAs [20], and 1 was a fragment of retrotransposon GilT.
Of the four small RNAs from snoRNAs, one is a 24 nt fragment (miR4) from GlsR1 (85 nts), two are 21 and 26 nt 3′-fragments (miR1 and miR3) of GlsR16 (77 nts) and one is a 26 nt 3′-fragment (miR2) from GlsR17 (144 nts). These non-coding RNAs have been previously described as snoRNA based on the presence of box C/D motifs [20] (see Discussion). GlsR1 is predicted to function in 2′-O-ribose methylation based on its sequence complementary to that of Giardia 16S ribosomal RNA as well as its homology to yeast snR70 [20]. SnoRNAs GlsR17 and GlsR16, however, did not contain antisense sequences complementary to ribosomal RNA and were considered to be “orphan” snoRNAs. Therefore, miR2 (5′ CAG CCU AAU CAC CGC CCC UAU AGU CC 3′) from snoRNA GlsR17 and miR1 (5′ CAA CGC ATC ACC GCT CTG ACC 3′) and miR3 (5′ GCA GAC AAC GCA TCA CCG CTC TGA CC 3′) from snoRNA GlsR16 were of particular interest in terms of their potential function in Giardia. Box C/D snoRNAs typically form a hairpin structure with the 5′ and 3′ ends forming part of the stem. Therefore, it is not surprising that MFOLD analysis of full-length snoRNA GlsR16 and the 64 nt 3′-portion of GlsR17 resulted in the formation of thermodynamically favorable hairpin loop structures with the corresponding miRNAs localized to one of the two arms at the 3′-end (Figure 1) [25]. Interestingly, MFOLD analysis of the full length GlsR17 resulted in a double stem loop structure similar to that observed for box H/ACA snoRNAs with miR2 located in the second hairpin structure (Figure S3). The relatively small sizes and the hairpin loop structures of these snoRNAs qualify them as suitable substrates for Dicer action.
To ascertain that miR2 and miR3 are not random degradation products isolated during the initial cloning of small RNAs, their presence in the RNA from Giardia was monitored. The presence of GlsR16 and GlsR17 in total Giardia RNA was repeatedly demonstrated by Northern analysis as anticipated (data not shown). But miR2 and miR3 remained undetectable on these blots presumably due to their relatively low levels. A splinted ligation analysis was then performed on the size-fractionated small RNA samples (<40 nts) because of its relatively high sensitivity based on direct labeling of the specific small RNA of interest through a 3′-end specific ligation [26]. The results from this analysis confirmed the presence of miR2 in the original small RNA sample (Figure 2). The single labeled band with the anticipated size indicates that miR2 is a product of precise processing of GlsR17 in Giardia and not the result of nonspecific degradation. Splinted ligation analysis of miR3 also showed a single specific band with the anticipated size of a 26 nt RNA but not a 21 nt RNA (data not shown). Thus, miR3 is also a likely natural product from GlsR16 in Giardia whereas miR1 is more likely a nonspecific degradation product. A similar analysis of the small RNA derived from retrotransposon GilT failed to produce a specifically labeled band (data not shown). These data indicate that miR2 and miR3 are naturally processed small RNAs from their respective snoRNAs while the retrotransposon GilT small RNA is probably a degradation product.
Identification of putative miR2 target sites in the Giardia genome was performed using the miRanda program [27]. Since most known miRNA target sites have been localized to the 3′- UTRs of mRNAs [28], we focused our target identification to the 3′-UTRs of Giardia open reading frames (ORF). To limit the number of possible candidate target sites, segments of 250 nts (50 nts upstream and 200 nts downstream from the stop codon) were extracted from each of the 9,649 ORFs in the Giardia genome database [24] and used to identify possible target sites for miR2 using miRanda (score threshold = 120, energy threshold = 20 kcal/mol, and scaling factor = 4). Identification was based on localization of the putative binding site (between 50 nts upstream and 50 nts downstream from the stop codon) and the presence of a seed sequence (5 consecutive nts, without G:U base pairings, within 3 nts of the 5′ end of the miRNA).
Of the 9,649 ORFs in the Giardia database [24], 296 ORFs contained potential target sites for miR2 based on the criteria mentioned above. Among these ORFs, 191 were hypothetical proteins and 105 were annotated proteins. Of the latter, 22 were variant-specific surface proteins (VSP) and 7 were trophozoite cysteine-rich surface antigens. VSPs are known to have a highly conserved 3′ end [29]. An alignment of 10 randomly chosen VSP sequences indicated that the last 100 nts at the 3′-end are highly conserved (data not shown). The predicted miR2 target site is 30 nts long with a 20 nt segment within the coding region and 10 nts in the 3′-UTR with the required perfect complementation between the seed sequence of miR2 and the target site located in the 3′-UTR. This remarkable conservation of a 3′-UTR sequence among the 22 VSP transcripts could mean that their expression is subject to a common mechanism of regulation mediated by miR2.
To test the potential consequence of in vivo interactions between miR2 and the putative target sites identified in the 3′-UTR of VSP transcripts, six copies of the putative binding sites from VSP AS12 were added to the 3′-UTR of Renilla luciferase gene in a plasmid construct (Figure 3A). Capped mRNA of this chimeric gene was transcribed in vitro (RL-TS) and electroporated into Giardia WB trophozoites together with chemically synthesized miR2. After incubation at 37°C for 5 hrs, the transformed Giardia cells were lysed and assayed for luciferase activity. Expression of RL-TS, when introduced into Giardia alone, was set at 100%. The inclusion of 0.5 µg of miR2 reduced the luciferase activity by 23% (Figure 3B), while 1 µg of miR2 decreased luciferase activity by 35% (Figure 3B). Little additional inhibition was observed when the concentration of miR2 was increased to 2 µg (40%) (Figure 3B). Thus, exogenously introduced miR2 does repress the expression of RL-TS in Giardia in a concentration-dependent but saturable manner. Quantitative RT-PCR estimation of the RL-TS mRNA levels in transfected Giardia with or without the presence of exogenously introduced miR2 resulted in two Ct values of 20.4±0.95 and 21.2±0.8, respectively (Figure 3C). This lack of apparent difference in mRNA levels indicates that the miR2 repression of luciferase expression is not due to enhanced mRNA degradation.
To confirm that the reduced luciferase activity is attributed to a specific interaction between miR2 and the putative target sites in the 3′-UTR of RL-TS, capped luciferase mRNA without the target sites (RL) was transfected into Giardia cells with or without 1 µg exogenous miR2. The luciferase activity increased in both cases to ∼115% of the control value (Figure 3D), suggesting that the repressive effect of miR2 requires the presence of target sites in the mRNA. It also suggests that the absence of target sites in the mRNA may also relieve the latter from repression by endogenous miR2 so that the expression of RL goes ∼15% beyond that of RL-TS (Figure 3D).
To further test the specificity of the interaction between miR2 and its target sites, Giardia cells were transfected with a mutant miR2, miR2neg (5′ p-CAG GGA UUA CAC CGC CCC UAU AGU CC 3′), in which the five underlined nucleotides of the “seed” sequence in miR2 were mutated to the complementary sequence. The miR2neg is not expected to interact with the target site. Luciferase activity of the RL-TS mRNA was not affected when it was introduced into Giardia with 1 µg of miR2neg. This suggests that the interaction between miR2 and its target sites is sequence specific and essential for the repression of luciferase activity (Figure 3D).
snoRNAs are essential for the maturation of ribosomal RNA and are localized to the nucleolus. miRNAs, however, are localized to the cytoplasm to regulate gene expression. To verify the localizations of GlsR17 and miR2, fluorescence in situ hybridization (FISH) was performed. A 26 nt RNA probe directed against the 5′ end of GlsR17 was 5′-end labeled with FAM and another 26 nt probe complementary to miR2 at the 3′-end of GlsR17 was 5′ end labeled with Cy3. Hybridization of these probes to Giardia WB trophozoites resulted in an exclusive localization of GlsR17 to the nucleus with a specific focus in the putative nucleolus (Figure 4). This outcome is in good agreement with our previous finding that the trimethyl cap, known to cap the snoRNAs in Giardia [30], was specifically localized to the nucleolus-like organelle in Giardia nucleus [30]. The probe for the 3′-end of GlsR17, where miR2 is positioned, stained the nucleolus like the GlsR17 5′ probe, but the majority of the stain was found in the cytoplasm (Figure 4), indicating that miR2, the presumed product from GlsR17 by Dicer digestion, is localized primarily to the cytoplasm. Thus, in the apparent absence of Drosha and Exportin5, the 144 nt snoRNA GlsR17 could be transported into the cytoplasm and trimmed to the mature 26 nt miR2 by Dicer in Giardia.
A surprising and unexpected result from the FISH experiments was that the majority of nucleolar GlsR17 was found in only one of the two Giardia nuclei (Figure 4). While many previous studies indicate that the two nuclei in Giardia are virtually identical in many aspects [31],[32] (see Discussion), our current serendipitous finding may indicate functional differences related to snoRNAs between the two nuclei.
Dicer is responsible for the final maturation of miRNAs. We reasoned that if miR2 and miR3 were miRNAs involved in translational repression in Giardia, prior Dicer processing of their precursors would be required for their maturation. Therefore, a reduction in the Dicer level in Giardia would decrease the levels of both miR2 and miR3. For miR2, Dicer depletion would also relieve the repression of RL-TS expression by the endogenous miR2. Introduction of a specific antisense-hammerhead ribozyme RNA into giardiavirus-infected Giardia trophozoites was used to knock down Giardia Dicer mRNA [21]. This resulted in a 60% knockdown of Dicer mRNA as measured by semi-quantitative RT-PCR and an ∼47% decrease in the growth of Giardia after 4 days, indicating a role of Dicer in Giardia proliferation (Figure S1C and S1D). Splinted ligation analysis of size fractionated small RNAs (<40 nts) from Dicer knockdown cells showed a drastic decrease in the levels of both miR2 and miR3 (Figure 5A and 5B), indicating that Giardia Dicer is required to produce mature miR2 and miR3. To ascertain that other RNA species are not nonspecifically affected by the Dicer knockdown, size-fractionated RNAs (<200 nts) from the Dicer knockdown and the control cells were compared (Figure 5C). There is little difference among the RNAs with >100 nts between the two cell lines. But a band of ∼85 nts appears enhanced and a smeared band close to the 26 nt region seems diminished in the Dicer knockdown cells, which could represent certain miRNA precursors and total miRNAs, respectively. Dicer knockdown thus does not appear to affect RNA level nonspecifically.
Expression of capped RL-TS mRNA transfected into Dicer knockdown cells by electroporation showed a 15% increase from the wild type control (data not shown). This suggests that the lowered miR2 concentration, resulting from the Dicer knockdown, relieves the previously observed endogenous translational repression of RL-TS (Figure 3D).
Domain analysis of GlAgo identified a PIWI domain but not a PAZ domain [33],[34]. Sequence alignments between GlAgo and the Argonautes of various origins, however, was able to identify both a PIWI and a PAZ domain [35]. Since the PAZ domain is known to interact with the small RNAs [36],[37], its apparent absence from GlAgo domain analysis raised the question whether GlAgo could bind to miR2. Recombinant His-tagged GlAgo was expressed and purified from transformed Escherichia coli. Gel shift analysis, using radiolabeled miR2 and recombinant GlAgo, indicated a slower moving radiolabeled band suggesting the formation of a GlAgo-miR2 complex (Figure S4, lanes 2 and 3). This binding of miR2 to GlAgo could be efficiently competed off by unlabeled miR2 (Figure S4, lane 5). Yeast RNA (Figure S4, lane 4) was also able to compete off miR2, indicating a lack of RNA sequence specificity for the binding. This result agrees with the previous observations on binding properties of Argonautes, which indicated that only the presence of a 5′ phosphate and a 3′ hydroxyl in an RNA molecule are required for binding but not sequence specificity [36]–[38]. Thus, GlAgo is apparently capable of binding small RNAs in the same way as the other Argonautes.
To confirm that the repression of luciferase expression by miR2 in Giardia requires GlAgo, the GlAgo knockdown cells developed previously were transfected with the RL-TS transcript along with or without synthetic miR2. The results indicated that, in the absence of 50% of the GlAgo mRNA, expression of the luciferase activity exceeds that of the wild type control by about 20% (Figure 6). This enhanced activity was not affected by introducing exogenous miR2 into the cells, indicating that, by knocking down GlAgo, the repression of luciferase expression by both endogenous and exogenous miR2 was largely abolished. Thus, GlAgo is clearly playing an essential role in the miR2-mediated repression of luciferase expression in Giardia.
In the present study, we have provided evidence that (1) a miRNA-mediated mechanism of translational repression is present and functional in one of the most ancient eukaryotes G. lamblia; (2) the snoRNAs in Giardia can be precursors of miRNAs. To date, miRNA precursors have been identified among non-coding cellular transcripts [23],[39], 3′-UTRs of mRNAs [40],[41], introns [42], transposable elements [43], and viral transcripts [44]. Our current observation that snoRNAs can also function as a miRNA precursor, albeit in a deeply branched organism, has broadened the potential biological roles of this family of small RNAs and provided interesting speculations on the probable route of their evolution.
In eukaryotes, snoRNAs are generally regarded as being responsible for the maturation and modification of ribosomal RNA. Two families of snoRNAs have been identified based on sequence conservation, box C/D and box H/ACA. Box C/D snoRNAs contain conserved Box C (UGAUGA) and Box D (CUGA) elements near the 5′ and 3′ ends, respectively as well as internal copies of these elements termed Box C′ and Box D′ [45],[46]. An interaction between the 5′ and 3′ termini allows the formation of a stem bringing the Box C and Box D elements together to form a hairpin structure. Box C/D snoRNAs serve as the guide for 2′-O-ribose methylation of ribosomal RNA. Box H/ACA snoRNAs have secondary structures consisting of two hairpins separated by a hinge region and a short tail. The box H (ANANNA) element is located in the hinge region while an ACA element is located in the tail, 3 nucleotides from the 3′ end [45],[46]. Box H/ACA snoRNAs guide the pseudouridylation of ribosomal RNA. Interestingly, snoRNAs containing either the box C/D or box H/ACA motif but without an rRNA antisense region to guide modification of rRNA have been also identified [46]. The role of these “orphan” snoRNAs remains unclear. Dicer could theoretically trim the general hairpin structures of snoRNAs, which range from 60 to 160 nts in length in Giardia, without prior processing to produce potentially functional miRNAs.
In Giardia, small nuclear RNAs (snRNAs) were originally identified with antibodies directed against the trimethyl cap [47]. Subsequently, cloning and bioinformatics analysis identified approximately 20 characterized snoRNAs and 60 putative snoRNAs [20],[48]. GlsR16 and GlsR17 were among the 20 characterized snoRNAs and categorized in the orphan group. The 60 putative snoRNAs have not yet been tested experimentally for their real functions in Giardia, though some of them were postulated to modify rRNA due to the presence of anti-sense sequences that target rRNA. It is possible that some or all of them may function as precursors of miRNAs for specific translational repression. This is not to say that the snoRNAs are the only miRNA precursors in Giardia. Other non-coding RNAs, such as the abundant antisense RNAs [18], could also be a potential source for miRNA. However, in the apparent absence of Drosha/Pasha and Exportin5, one has to question the ability of Giardia Dicer to digest relatively large RNA hairpins. Giardiavirus, a dsRNA virus with a genome size of 6,277 bps, is known to multiply vigorously in the cytoplasm of infected Giardia trophozoites [49],[50]. This fact appears to rule out the possibility that Giardia Dicer could digest relatively long dsRNA. The snoRNAs identified in Giardia have lengths ranging from 60 to 160 nts [48], which appear to be the right size of RNA to fold into hairpin structures suitable for digestion by Dicer. Thus, the mature snoRNA could be a Dicer substrate without additional processing. So, it is not entirely unlikely that snoRNAs and perhaps some other small RNAs may constitute the reservoir for miRNA in Giardia. They could be exported from the nucleus by some yet unidentified means and digested by the Dicer for miRNA. Future investigations will provide answer to these intriguing possibilities.
In light of Giardia's minimalistic nature, it is reasonable to hypothesize that Giardia utilizes a single RNA processing mechanism for the maturation of both snoRNAs and miRNA precursors with characteristics similar to snoRNA. In human cells, snoRNAs are transcribed by RNA pol II [51]. Maturation of snoRNAs from the primary transcript involves processing of the 3′ end and hypermethylation of the m7G cap to a trimethyl cap (m2,2,7G) [51]. Recently, studies have shown that snoRNAs have a cytoplamic component during biogenesis. snoRNAs injected into the cytoplasm of Xenopus oocytes are imported into the nucleus, indicating a mechanism for nuclear transport [52]. Additionally, snoRNAs have been shown to associate with the nuclear export complex [53],[54] and reach their maturation in the cytoplasm [55]. Although this mechanism has only been shown for snoRNAs known to be involved in rRNA maturation, it is possible that “orphan” snoRNAs may also mature utilizing the same mechanism.
The outcome from our FISH experiments indicated the localization of GlsR17 primarily in the nucleolus and miR2 in the cytoplasm (Figure 4). A faint signal of GlsR17 was, however, also detectable in the cytoplasm. It is unclear if this signal is due to the presence of GlsR17 in the cytoplasm or background. It is possible that Giardia snoRNAs have a cytoplasmic component during maturation. After processing of the pre-snoRNA in the cytoplasm, the mature snoRNA may either be rapidly imported into the nucleus or processed by Dicer. Therefore, only small undetectable amounts of snoRNA may remain in the cytoplasm. The export of snoRNAs could provide an ideal method for bypassing the requirements for Drosha/Pahsa in producing pre-miRNAs.
Although the proposed mechanism for production of pre-miRNAs is unique in Giardia, the remaining aspects of the miRNA pathway appear similar to that found in higher eukaryotes. These similarities include the requirement of Dicer for the production of mature miRNAs and an essential role of Argonaute in translational repression. In the absence of Dicer, mature miR2 and miR3 were virtually abolished. Additionally, both miR2 and miR3 are 26 nts in size, reflecting the expected size of the Giardia Dicer cleavage products based on its crystal structure [17]. GlAgo plays an essential role in miR2-mediated translational repression. The apparent formation of a GlAgo-miR2 complex in vitro and the specific binding of GlAgo to m7G-Sepharose suggests that it functions like a bona fide Argonaute by becoming incorporated into RISC with miRNA and competing for the cap with eIF4E [13], which results in the inhibition of translation initiation. When GlAgo was partially depleted from Giardia, translation of reporter mRNA RL-TS was increased by ∼20% with or without the exogenously introduced miR2. This indicates an inability of endogenous as well as the exogenously introduced miR2 to repress the translation of RL-TS when GlAgo was knocked down. Thus, the cytoplasmic component of the miRNA pathway in Giardia functions similarly to those seen in other eukaryotes.
A third of the mRNAs of annotated proteins identified to carry a potential target site for miR2 encode either VSPs or trophozoite cysteine-rich surface antigens. Giardia contains approximately 235–275 VSP genes in clusters of two to nine in a head to tail orientation [14]. Only a single VSP, however, is expressed on the cell surface at any given time [29]. Expression of multiple VSPs is only observed during VSP switching or excystation [29],[56]. Recently, Kulakova et al. showed that the activation of VSP expression is regulated by acetylation of the histone upstream of the gene [57]. It is, however, still unclear how VSP exchange occurs. During VSP exchange, expression of the currently expressed VSP must be repressed and the expression of a new VSP has to be up-regulated. In this genetic turmoil, it is possible that multiple VSP mRNAs are transcribed. The role of miR2 may be to differentially limit the translation of a particular family of 22 VSP mRNAs. It is possible that the 235–275 VSP transcripts can be classified into several families each possessing a similar 3′-UTR targeted by a specific miRNA. When the level in one of these miRNAs is lowered, a specific family of VSP will be over expressed, though it is still not clear how only one of the family members will eventually be expressed on the cell surface. The 22 identified VSP genes carry a similar miR2 target site, but they are not identical. Differences in the free energy of binding between miR2 and individual VSP target sites could differentially regulate the expression of each VSP so that only one of them eventually becomes expressed on the cell surface when the level of miR2 is lowered to a certain specific range. This may contribute to the pathogenicity of Giardia by limiting the number of antigens exposed to the immune system at a given time and thereby increasing the number of VSPs novel to the host. As a consequence, Giardia may be able to evade the host immune response. The precise role of miR2 in regulating expression of the 22 VSP genes needs to be investigated further.
There has been an accumulation of experimental data over the past years indicating that the two nuclei in Giardia are virtually identical. Kabnick K.S. and Peattie D.A. showed that both nuclei contain equivalent rDNA based on in situ hybridization and that both nuclei are transcriptionally active based on the incorporation of [3H] uridine [31]. Yu et al. used FISH and demonstrated that each of the two nuclei has a complete copy of the genome and are partitioned equally during cytokinesis [32]. Therefore, the localization of GlsR17 to the nucleoli of a single nucleus was surprising. Localization of the other identified snoRNAs will determine if this is a unique occurrence or a general phenomenon in Giardia.
Differences in the number of chromosomes [58] as well as discrepancies in the number of nuclear pores [59] between the two nuclei have recently been described in Giardia. Since the number of nuclear pores has generally been correlated to transcription activity in other eukaryotes [59], it suggests that one of the two nuclei may have higher transcriptional activities than the other. It is possible that one of the two nuclei may be solely responsible for production and export of some or all of the snoRNAs. A fascinating aspect may concern the apparent exclusive import of the snoRNAs back into the same original nucleus and the mechanism dictating it. Clarifications of this unusual finding will have to wait for further investigation.
Giardia is a deeply branched eukaryote showing a combination of prokaryotic and eukaryotic features [14]. The presence of the miRNA pathway in this organism suggests that miRNA-mediated repression of translational initiation could be an ancient mechanism of gene regulation and that the snoRNAs may represent the original miRNA precursors. During evolution, snoRNAs, while maintaining their original function in ribosomal RNA maturation, may have become involved in gene regulation. These small non-coding RNAs would make ideal substrates for a primative RNase III enzyme before the evolution of new enzymes for the production of miRNAs from other much larger precursors. Since snoRNAs have been routinely discarded from the libraries of potential miRNA precursors in the past, it is not inconceivable that some of the snoRNAs in higher eukaryotes may still assume the role of miRNA precursors today, and could be readily tested.
The Giardia Argonaute was PCR amplified from Giardia genomic DNA using primers Ago full U-2 (5′ GAG CCC GGG TCA CTA GTG CCA TGG TAG CAG ATG TTG TCA C 3′) and Ago full L-2 (5′ GAG CTC GAG GCG GCC GCC TAG TGG TGG TGG TGG TGG TGT ATG AAG AAT GGT CTG TAC T 3′). The amplified product was cloned into the pGEM-T Easy vector (Promega), sequenced, and subcloned into the pET28b (Novagen) expression vector using NcoI/XhoI. The pET28b GlAgo was electroporated into BL21(DE3) cells containing the pG-KJE8 plasmid (Takara), which provides over-expression of chaperone proteins. An overnight culture was diluted 1∶100 into fresh LB media containing 10 ng/ml tetracycline and 4 mg/ml arabinose and incubated at 37°C for 1 hour to allow for the expression of chaperone proteins. Expression of GlAgo was induced with 0.1 mM IPTG and incubated at room temperature for 5 hours. Pelleted cells were lysed with BugBuster (Novagene) and bound to Ni-NTA beads in the presence of 40 mM imidazole, 5 mM ATP, and 10 mM MgCl2 at 4°C overnight. Beads were washed with 15 ml of wash buffer (20 mM sodium phosphate, pH 60; 500 mM NaCl) containing 60 mM imidazole and the protein was eluted in 4 ml of wash buffer containing 150 mM imidazole, concentrated and transferred to storage buffer.
Small RNAs were cloned following the protocol from the Ambros lab (http://banjo.darthouth.edu/lab/MicroRNAs/Ambros_microRNAcloning.htm). In short, Giardia cells were lysed and small RNAs were enriched using the mirVana kit from Ambion. The purified small RNAs were further size fractionated using the Ambion FlashPAGE fractionator to select small RNA less then 40 nts. The small RNAs were then linked to a 3′ linker (AMP-5′p-5′p/CTG TAG GCA CCA TCA AT di-deoxyC- 3′) and size fractionated on a 15% urea-polyacrylamide gel. Those with attached 3′ linkers (∼40 nt) were electroeluted from the gel and ethanol precipitated. The 3′ linked small RNAs were then ligated to a 5′ linker (5′- ATC GTrA rGrGrC rArCrC rUrGrA rArA –3′; “r” denotes RNA). This step is designed to select for small RNAs with a 5′ monophosphate, as would be expected from Dicer processing. The reaction was cleaned by phenol/chloroform extraction followed by ethanol precipitation. The cleaned product was used in an RT-PCR reaction to make cDNA, which was gel purified, digested with BanI and ethanol precipitated. The cleaned product was used in a ligation reaction to concatamerize the PCR product together. The reaction was run on an agarose gel and fragments between 600–1000 bp were isolated. The isolated fragments were re-amplified using PCR and cloned into pGEM-T Easy using the pGEM-T Easy kit (Promega). Colonies containing insert were sequenced.
The Hammerhead ribozyme was incorporated into the GlAgo antisense RNA using recombinant PCR. Briefly, PCR1 containing the 5′ complementary sequence and the ribozyme was amplified using Ago HHR PCR1 F (5′ CGC GCG CTC GAG CTC CCA GAT TGA CCT GGG ATC 3′) and Ago HHR PCR1 R (5′ CTG CCC CTG AAC TAT AGA GTG CTG ATG AGT CCG TGA GGA CGA AAC TCT GAA AAC CTT TCC GTT G 3′; ribozyme underlined). PCR2 containing the 3′ complementary sequence was amplified using Ago HHR PCR2 F (5′ CAC TCT ATA GTT CAG GGG CAG 3′) and Ago HHR PCR2 R (5′ GCG CGC GAG CTC CTT CAA TGG TAA CTA TAC GAG 3′). Purified PCR1 and PCR2 were then used as the template for the recombinant PCR using Ago HHR PCR1 F and Ago HHR PCR2 R. The recombinant PCR was cloned into pGEM-T Easy and sequenced. The Hammerhead ribozyme surrounded by 500 nt of complementary GlAgo sequence was then cloned into pC631pac using the XhoI restriction site. This clone was linearized with NruI and used for in vitro transcription. The transcribed RNA was transfected into giardiavirus-infected Giardia. Puromycin (100 µg/ml) was used to select for expression of the ribozyme.
A similar approach was used to incorporate a Hammerhead ribozyme into the Dicer antisense. PCR3 was obtained using primers Xho-Dicer 5′ (5′ GCC TCG AGT TTA GTA GGA ATG CAT GCT TTG G 3′) and Dicer RZ2 (5′ GGG TAG AAT CGA TCC CAA GAA CCT GAT GAG TCC GTG AGG ACG AAA CAT AAA GAG ACC AGC 3′; ribozyme underlined). PCR4 was obtained using primers Xho-Dicer 3′ (5′ GCC TCG AGG GAT ATT ACA CTA CGC ATC AGC 3′) and Dicer RZ1 (5′ GCT GGT CTC TTT ATG TTT CGT CCT CAC GGA CTC ATC AGG TTC TTG GGA TCG ATT CTA CCC 3′; ribozyme underlined). Purified PCR3 and PCR4 were used as templates for recombinant PCR using Xho-Dicer 5′ and Xho-Dicer 3′. After cloning into pGEM-T Easy and sequencing, the PCR product was cloned into the pC631neo vector using XhoI and the in vitro transcript transfected into giardiavirus-infected Giardia as described above. Neomycin (800 µg/ml) was used to select for expression of the ribozyme.
Giardia WB trophozoites were grown in modified TYI-S-33 media to a density of 107 per ml, washed twice in phosphate buffered saline (PBS), once in electroporation buffer (10 mM K2HPO4–KH2PO4 (pH 7.6), 25 mM HEPES (free acid), 120 mM KCl, 0.15 mM CaCl2, 2 mM EGTA, 5 mM MgCl2, 2 mM ATP, 4 mM Glutathione), and finally resuspended in electroporation buffer. RL-TS mRNA (3.5 µg), yeast tRNA (125 µg), and, if needed, 1 µg of 5′-phosphate-miR2 RNA (miR2) (IDT) were added to the cell suspension, incubated on ice for 10 minutes and then subjected to electroporation using a Bio-Rad Gene Pulser Xcell (Voltage: 450 V, Capacitance: 500 µF, Resistance: ∞). Cells were then incubated on ice for 10 minutes and added to pre-warmed culture medium. The transfected cells were incubated at 37°C for 5 hours, pelleted, washed once in PBS, and lysed using the Renilla luciferase assay kit (Promega). The lysate was centrifuged at 12,000 g for 2 min to remove cellular debris. The cleared lysate was used to test for Renilla luciferase activity. The protein concentration of the cleared lysate was measured by the Bradford method (Bio-Rad) and used to normalize the luciferase activity.
Total RNA was isolated from Giardia using Trizol (Invitrogen) while RNA <200 nts was isolated using the mirVana kit (Ambion). Further fractionation of total RNA to <40 nts was accomplished by standard denaturing PAGE and electroelution or by using the Ambion FlashPAGE followed by ethanol precipitation.
Splinted ligation was preformed as previously described [26]. Size fractionated RNA was incubated with 100 pmoles of the “bridge” oligo B1 (5′ C3 spacer-GAA TGT CAT AAG CGG GAC TAT AGG GGC GGT GAT TAG GCT G–C3 spacer 3′) containing the miR2 binding site (underlined) and 100 fmoles of the “linker” oligo L1 (5′ CGC TTA TGA CAT TCddC 3′) with 20 mM Tris-HCl (pH 8.0) and 75 mM KCl. The reaction mixture was incubated at 95°C for 1 min, 65°C for 2 min and 37°C for 10 min. Finally, 1X T4 DNA ligase buffer and 10 U of T4 DNA ligase (NEB) was added to the reaction and incubated at 30°C for 1 hour. The ligase was heat inactivated by incubation at 75°C for 15 min. The reaction mixture was loaded on to a pre-run 15% denaturing urea-polyacrylamide gel and quantified using a PhosphorImager.
Giardia WB trophozoites were harvested by placing culture tubes on ice for 15 minutes and centrifuging to pellet the cells. The cells were suspended in 1 ml of modified TYI-S-33 culture medium, placed on cover slips pretreated with 0.1% poly-L-lysine, and incubated at 37°C for 30 minutes to allow the trophozoites to adhere. They were then fixed in 4% paraformaldehyde for 30 minutes at room temperature and washed with PBS and 2X SSC (300 mM NaCl, 30 mM sodium citrate) for 5 minutes each. The cells were permeabilized with 0.5% Triton X-100 for 5 minutes at room temperature and dehydrated in 70% ethanol followed by 100% ethanol for 5 minutes each. The dehydrated cells were denatured with 2X SSC in 70% formamide for 2 minutes at 70°C and dehydrated again with cold 70% and 100% ethanol. Salmon DNA (10 µg), yeast tRNA (25 µg), and 100 pmoles of the GlsR17 RNA 5′-UTR probe (5′ 6-carboxyfluorescein (FAM)–CCC GGA UCC UCA CCA CGA GUA AAC CC 3′) and miR2 RNA 3′-UTR probe (5′ Cy3–GGA CUA UAG GGG CGG UGA UUA GGC UG 3′) (IDT) were added to 100 µl of 100% formamide, heated at 75°C for 10 minutes and placed immediately on ice. The probe was mixed with an equal volume of hybridization buffer (4X SSC, 20% dextran sulfate, and 4 mg/ml BSA), added to the cover slip and incubated overnight at 37°C. The cover slips were washed 3 times with 0.1% SSC in 50% formamide at 50°C for 5 minutes each to remove unhybridized probe, placed facedown on clean glass slides with 1 drop of Vectashield (Vector Labs) mounting media with DAPI (4′,6 diamidino-2-phenylindole) and sealed with paraffin wax. Cells were examined using a Nikon TE2000E motorized inverted microscope equipped with 60X bright-field and epifluorescence optics. Images were acquired with the NIS-Elements Advanced Research software (Nikon) and analyzed with ImageJ (http://rsbweb.nih.gov/ij/index.html).
|
10.1371/journal.pcbi.1000525 | Combining Fungal Biopesticides and Insecticide-Treated Bednets to Enhance Malaria Control | In developing strategies to control malaria vectors, there is increased interest in biological methods that do not cause instant vector mortality, but have sublethal and lethal effects at different ages and stages in the mosquito life cycle. These techniques, particularly if integrated with other vector control interventions, may produce substantial reductions in malaria transmission due to the total effect of alterations to multiple life history parameters at relevant points in the life-cycle and transmission-cycle of the vector. To quantify this effect, an analytically tractable gonotrophic cycle model of mosquito-malaria interactions is developed that unites existing continuous and discrete feeding cycle approaches. As a case study, the combined use of fungal biopesticides and insecticide treated bednets (ITNs) is considered. Low values of the equilibrium EIR and human prevalence were obtained when fungal biopesticides and ITNs were combined, even for scenarios where each intervention acting alone had relatively little impact. The effect of the combined interventions on the equilibrium EIR was at least as strong as the multiplicative effect of both interventions. For scenarios representing difficult conditions for malaria control, due to high transmission intensity and widespread insecticide resistance, the effect of the combined interventions on the equilibrium EIR was greater than the multiplicative effect, as a result of synergistic interactions between the interventions. Fungal biopesticide application was found to be most effective when ITN coverage was high, producing significant reductions in equilibrium prevalence for low levels of biopesticide coverage. By incorporating biological mechanisms relevant to vectorial capacity, continuous-time vector population models can increase their applicability to integrated vector management.
| It has recently been proposed that mosquito vectors of malaria may be controlled by biopesticide sprays containing spores of fungi that are pathogenic to mosquitoes, causing reduced blood feeding activity and eventual death. This technique has been shown to have strong potential to reduce malaria transmission rates, and may be most effective when combined with other interventions as part of an integrated vector management strategy. I develop a model to quantify the total impact of combined interventions that can affect mosquitoes at different ages and stages in their lifecycle. As a case study, I consider the combined use of fungal biopesticides and insecticide- treated bednets (ITNs), a widespread and important vector control method. The model demonstrates that these interventions combined can have strong effects on malaria transmission even in situations where each intervention acting alone has relatively little impact. In situations difficult for malaria control due to high transmission intensity and widespread insecticide resistance, the performance of the combined interventions is improved by synergistic interactions between the interventions, whereby the ITN intervention improves the performance of the fungal biopesticide intervention. The results suggest that the combined use of ITNs and fungal biopesticides may be an efficient and effective method of malaria control.
| Malaria is a major contributer to the global disease burden, and disproportionately affects low-income countries with climates suitable for transmission [1]–[3]. Vector control strategies have proven effective in reducing malaria transmission and prevalence [4]–[6], and are a key element of current malaria control initiatives [7]–[9]. Indoor residual spraying (IRS) and insecticide-treated bednet (ITN) interventions have been and remain the dominant methods of controlling malaria vectors [4], [9]–[12], but problems of public health and insecticide resistance associated with chemical insecticides have increased interest in alternate methods, including novel biological methods [13], [14]–[16]. Because the incubation period of the malaria parasite is relatively long in comparison to the average adult mosquito lifespan, biological methods of vector control that have sublethal and lethal effects at different points in the mosquito life cycle may substantially reduce the potential for malaria transmission [17]–[21]. Such methods may be most effective when combined with established methods in a strategic manner [8],[22],[23].
In order to impact on malaria prevalence it is necessary to reduce transmission to very low levels [24],[25]. Vector management strategies that combine multiple mosquito control interventions would therefore benefit from tactical design to alter mosquito life history in ways that are likely to maximise the impact on malaria transmission, given the resources available. This paper presents an age-structured model that explores the impact of interventions that affect multiple gonotrophic and demographic processes in the mosquito on malaria transmission and prevalence. As a case study, the combined use of fungal biopesticide and ITN interventions is considered.
Biopesticides containing spores of entomopathogenic fungi are a novel strategy for controlling malaria vectors that have shown potential to cause substantial reductions in malaria transmission in laboratory and field studies [17]–[20]. The biopesticide targets adult mosquitoes, infecting them with a fungal pathogen that does not kill instantly, and can generate a wide range of mortality patterns, some early-acting while others showing a distinct delay [17],[26],[27]. Fungal pathogen-induced mortality rates typically increase with the fungal infection age, with the average times to death due to fungal infection less than 10 days [17]–[21]. This slow-acting mortality suggests that high fungal infection rates in adult mosquitoes would be required to affect malaria prevalence, however sublethal effects of fungal infection on mosquitoes have been observed which may considerably reduce their transmission potential [21]. Fungal infection can cause a reduction in the blood feeding rate and lifetime fecundity [26]. There is also evidence that co-infection with the fungal pathogen and the malaria parasite can cause greater than additive mortality and reduced transmissibility of the malaria pathogen [15],[17].
In contrast to fungal biopesticides, ITNs are an established and widely used vector control method [6], [28]–[30] that has proven successful in reducing malaria transmission and prevalence in situations where high levels of community-wide ITN coverage are achieved [6], [29]–[31]. They therefore have a focal role in current vector control initiatives [7],[10],[12]. ITNs work by targeting the adult host-seeking mosquito population, increasing the time taken for mosquitoes to find a blood meal and increasing the mortality risk while host-seeking. Both factors interact positively to reduce the likelihood that mosquitoes live long enough to contract and transmit malaria. The effect of ITNs on mosquito mortality rates depends on levels of insecticide resistance [14], [32]–[34], the persistence of the insecticide treatment, and the excito-repellency properties of the insecticide [35],[36].
This study considers mechanisms by which ITN and fungal biopesticide interventions may affect mosquito populations at the scale of the gonotrophic cycle and at within-gonotrophic cycle time scales. The gonotrophic cycle in female adult mosquitoes is often conceptualised in terms of a host-seeking stage, during which mosquitoes actively search for a blood meal, and a non-host-seeking stage, during which blood from a recent blood meal is digested, oocytes are developed and eggs are oviposited, after which host-seeking activity begins again [37],[38]. While the ITN intervention reduces the rate of host-seeking success throughout the adult mosquito population as a whole, the fungal biopesticide may also extend the host-seeking stage in fungal pathogen-infected mosquitoes due to a deterioration in flight and blood-feeding capabilities [26]. The non-host-seeking stage in fungal pathogen-infected mosquitoes may also be protracted due to impaired metabolic functioning [26]. Within a given gonotrophic cycle, the period during which mosquitoes are exposed to a risk of fungal infection may not necessarily correspond to a particular gonotrophic cycle stage. For biopesticide application as a residual treatment in and around domestic dwellings, the fungal infection risk would conceivably be higher whilst mosquitoes are host-seeking, and also for some time after they obtain a blood meal when they often rest on nearby surfaces for a period of less than 24 hours [39],[40]. In fungal pathogen-infected mosquitoes, fungal infection age, and the corresponding risk of fungal pathogen-induced mortality, increases continually, with the mortality risk for all mosquitoes being augmented during the host-seeking stage by the presence of ITNs.
Similar to [21], the population dynamic model presented here is an age-structured Susceptible-Exposed-Infectious (SEI) model based on integral equations, considering fungal pathogen-induced age-dependent mortality in adult mosquitoes that can contract the fungal infection at any point in their adult life. This model reformulates that of [21] to explicitly incorporate gonotrophic cycles in the adult mosquito population by defining a recursive series of host-seeking and non-host-seeking classes of mosquitoes. The model thus retains the properties of existing continuous [21], [41]–[43] and discrete feeding cycle approaches [31],[44],[45]. Equilibrium analysis is used to validate the model for limiting cases similar to those represented by existing continuous-time models [21],[43]. Cases where the risk of fungal infection varies throughout the gonotrophic cycle and where fungal infection causes within-population variation in the lengths of both host-seeking and non-host-seeking stages are explored numerically.
The model is parameterized with literature data on mosquito-malaria interactions, and the effects fungal biopesticides and ITNs on mosquito populations. A series of questions relevant to mosquito control by fungal biopesticides, ITNs and both interventions in combination are explored. How do sublethal effects of fungal infection on rates of finding and processing blood meals affect the impact of biopesticides on malaria transmission rates? How is the fungal biopesticide performance affected by variation in the period of biopesticide exposure within a given gonotrophic cycle? How does the performance of fungal biopesticide and ITN interventions combined compare with that of each single intervention for varying levels of transmission intensity and insecticide resistance? Mechanisms important to the performance of fungal biopesticides, ITNs and both interventions combined are identified and discussed.
The dynamics of adult mosquito infection with the Plasmodium parasite are treated as an SEI process in continuous time with the mosquitoes being categorized as susceptible, exposed (but not yet infectious) and infectious [43]. Similar to [21], each of these three categories is divided into mosquitoes that are uninfected with the fungal pathogen and those that are infected. Mosquitoes that are infected with the fungal pathogen experience a fungal pathogen-induced mortality risk that is dependent on fungal infection age. Mosquitoes are further categorized according to their stage in the gonotrophic feeding cycle (Figure 1). Gonotrophic feeding cycle parameters may be altered by both fungal biopesticide and ITN interventions. These processes are now described in more detail and an age-structured population dynamic model is developed and analysed analytically for limiting cases. Model parameters are listed in Table 1 and model variables are listed in Table S3.
The blood-feeding patterns of adult female mosquitoes are assumed to follow a gonotrophic cycle whereby mosquitoes repeatedly seek a blood meal, obtain a single blood meal, and then stop seeking blood for a fixed time period after which they resume host-seeking activity and the gonotrophic cycle begins again. The model accordingly divides each gonotrophic cycle into two stages: a host seeking stage of variable duration, during which the mosquito is searching for a blood meal, and a non-host-seeking stage of fixed duration, during which the mosquito does not seek blood meals (Figure 1).
During the host-seeking stage, mosquitoes are assumed to feed soley on humans at a rate f (Table 1). This rate may be intuitively interpreted by the corresponding daily probability of finding a blood meal given that the mosquito does not die, (Table 1). During the host-seeking stage the mosquito mortality rate due to all mortality sources other than fungal biopesticide and ITN interventions is . Upon finding a blood meal mosquitoes enter the non-host-seeking stage, which lasts for days. During the non-host-seeking stage the mosquito mortality rate due to all mortality sources other than fungal biopesticide and ITN interventions is . For each stage of the malaria infection process (susceptible, exposed and infectious), the maximum number of gonotrophic cycles completed, including host-seeking and non-host-seeking stages, is nS, nE and nI respectively (Table S3).
In describing the model below, the rate of fungal infection in adult mosquitoes, F, is assumed to be a constant. This rate may be intuitively interpreted by the corresponding daily probability of fungal infection given that the mosquito does not die, (Table 1). However fungal infection risk may differ at different points in the gonotrophic cycle and so a version of the model where F varies throughout the gonotrophic cycle is analysed numerically in the Results.
The mortality rate of mosquitoes that are infected with the fungal pathogen is increased by an amount that may vary with u, the time that has elapsed since infection [21]. As in [21], , is modelled using the Weibull model , where and β are the rate and shape parameters respectively (Table S3). The measure of fungal pathogen virulence is the expected time until death due to fungal infection given no other mortality, , given by(1)[21].
Sublethal effects of fungal infection on mosquitoes are also considered, including a reduction in the host-seeking success and an increase in the time to the next host-seeking period following a blood meal in fungal pathogen-infected mosquitoes. This is modeled by a reduced blood feeding rate and an increased duration of the non-host-seeking stage in fungal pathogen-infected mosquitoes (Table 1).
Similar to [45], an ITN intervention is assumed to affect two gonotrophic cycle parameters, the mortality rate of host-seeking mosquitoes, , and the rate at which host-seeking mosquitoes take blood meals, fF or f for mosquitoes with and without fungal infection respectively. The blood-feeding rates fF and f are multiplied by a factor , where is the fraction of humans protected by ITNs [45]. This assumes that mosquitoes do not bite ITN users or non-human hosts. The parameter is referred to as the ITN coverage, in line with ITN literature [6],[31],[45]. The mortality rate is assumed to be directly proportional to such that when (Table 1). The effect of ITNs on mosquito mortality can be reduced by insecticide resistance in the mosquito population [32],[34], limited persistence of the insecticide treatment, or excito-repellency properties of the insecticide [36]. Therefore, the limiting case in which the ITN intervention has no effect on mosquito mortality is also considered.
The model is described by a system of integral equations for the mosquito density, defined as the number of mosquitoes per human [21],[43], as a function of time in each of a series of 12 stages. The stages correspond to the mosquito's malaria infection status (susceptible, exposed, infectious), gonotrophic cycle stage (host-seeking, non-host-seeking), and fungal infection status (infected, uninfected). Equations for stages of mosquitoes without fungal infection (Text S1) are considered separately from stages of fungal pathogen-infected mosquitoes (Text S2). A series of expressions are defined, , which give the probability that a mosquito remains in the same stage over different periods of time. These are described in more detail in Text S1 and Text S2 and are listed in Tables S1 and S2. Similar to [21],[43], the model assumes that mosquitoes are recruited to the adult population at a constant rate ε, that a constant fraction, x, of the human population is infected with malaria, the probability of transmission from an infected human or an infected mosquito is b, and that it takes the Plasmodium exactly TE days to mature in mosquito and become infectious (Table 1).
The equilibrium daily EIR, denoted , is given by(2)where and are the equilibrium densities of infectious, host-seeking mosquitoes with and without fungal infection respectively. To qualitatively estimate the relationship between the model-derived equilibrium daily EIR and the malaria prevalence in the human population, denoted , the best fit model obtained by [25] was used, with no change to the best fit parameters. This provides a conservative estimate of the prevalence, because the equilibrium EIR estimates from this model do not take into account reductions in prevalence that may occur as a result of the decrease in EIR resulting from fungal biopesticide and ITN interventions.
Expressions are derived for the equilibrium density of susceptible, exposed and infectious host-seeking mosquitoes for the limiting case in which there are no sublethal effects of fungal infection on mosquito feeding biology (, ) and the shape parameter of the fungal pathogen-induced mortality function (Text S1 and Text S2). The analytically derived equilibrium EIR agrees well with the equilibrium obtained by simulating equations (S1.1)–(S1.17) and (S2.1)–(S2.18) through time using a simulation algorithm coded in C++ (Figure 2). The simulation algorithm is used to obtain the equilibrium for the general case in which , and . The pattern in Figure 2 is similar to that produced by simpler models [21], whereby increasing the shape parameter β above 1 reduces the equilibrium malaria transmission rate.
The ITN intervention corresponds to a limiting case of the model in which the fungal infection rate F is zero. For this case, the dynamic system is much simpler, requiring only equations (S1.1)–(S1.17), and solution of the equilibrium EIR is much simpler, given by equations (S1.18)–(S1.27). The ITN intervention produces a rapid decline in the equilibrium EIR as the ITN coverage () increases, but does not have a strong impact on the equilibrium malaria prevalence in humans until the level of ITN coverage is moderate to high (Figure 3). These patterns are similar to those produced by [31],[45]. Figure 3 also shows the case where the ITN intervention has no effect on mosquito mortality, and thus affects only the blood feeding rates fF and f. This may represent widespread insecticide resistance in the mosquito population.
The aim of the models developed here is to explore generic issues relating to fungal biopesticide and ITN interventions rather than to parameterise a specific case. In order to use biological relevant parameters, a survey of the literature was conducted (Table 1). Using these parameters, the model gives an equilibrium annual EIR in the absence of the fungal biopesticide and ITNs of 47.8, which consistent with a number of regions in Africa with moderate to high malaria prevalence [24],[25]. This is similar to the value of 45 given by the continuous time models of [21],[43] with a human biting rate of and other parameters as in Table 1. Similar to [21], the equilibrium baseline EIR scales linearly with the recruitment rate ε.
The model was explored by asking a series of questions relevant to mosquito control by fungal biopestides, ITNs and the combined use of both these interventions, using the equilibrium EIR and the corresponding estimate of human malaria prevalence as a measure of successful intervention.
Fungal infection can substantially reduce mosquito blood-feeding activity [26]. Here, two possible effects of fungal infection on mosquito blood-feeding biology are considered, including a reduction in the blood-feeding rate in host-seeking mosquitoes, , and an increase in the duration of the non-host-seeking stage, (Table 1). Even when these effects act simultaneously, they have less potential to produce very low equilibrium EIR than decreasing the average time to death from fungal infection, (Figure 4). The strongest sublethal effects shown in Figure 4 represent more than a four fold increase in the average gonotrophic cycle length in fungal pathogen-infected mosquitoes, . For moderate to high daily probability of fungal infection, this has a similar effect to a 25% reduction in the average time to death from fungal infection (Figure 4B and C). When the daily probability of fungal infection is low, reductions in the equilibrium EIR obtained by either increasing the fungal pathogen virulence (by reducing ) or increasing the sublethal effects are considerably less, and the impact of strong sublethal effects on the EIR is of similar magnitude to that produced by strong reductions in (Figure 4A).
For a fungal biopesticide applied in and around human settlements, mosquitoes may not be exposed to a risk of fungal infection for their entire gonotrophic cycle. They may be most likely to contract the fungal pathogen when they are host seeking and for a time period directly after blood feeding. A shorter period of exposure to a risk of fungal infection within a given gonotrophic cycle may lead to lower fungal infection levels in the mosquito population, with implications for fungal biopesticide performance. Here, the effect of varying fungal infection risk throughout the gonotrophic cycle is explored by varying the fungal infection rate, F, such that non-host-seeking mosquitoes experience a constant fungal infection rate for an initial proportion of the non-host-seeking stage, denoted α, and a fungal infection rate of zero for the remainder of the non-host-seeking stage (Table 1). The corresponding daily probability of fungal infection during the period of biopesticide exposure is (Table 1). Host-seeking mosquitoes are assumed to experience fungal infection rate (and daily probability of fungal infection CE) throughout the entire host-seeking stage.
The equilibrium daily EIR is marginally higher if mosquitoes are exposed to fungal infection risk for half of the non-host-seeking stage () in comparison to exposure for the full stage duration () (Figure 5A, C, E). However, if mosquitoes are only exposed to fungal infection risk when they are host-seeking (), the equilibrium EIR is considerably higher. The corresponding estimate of the equilibrium malaria prevalence in humans varies more between the three exposure periods than the equilibrium EIR (Figure 5B, D, F), because it is sensitive to changes in EIR at low EIR values. The effect of varying the period of biopesticide exposure on prevalence is greater for more virulent biopesticides, which benefit more from the very low transmission levels achieved for longer duration of exposure to fungal infection risk (Figure 5A, B). Similar to [21], if mosquitoes are always exposed to fungal infection risk, there is a threshold level of the daily probability of fungal infection above which additional reductions in equilibrium EIR and prevalence are marginal. This is not the case if mosquitoes are only exposed to fungal infection risk when they are host-seeking, where prevalence continues to decline steadily as the daily probability of fungal infection is increased to high values (Figure 5).
ITNs can be an effective means of malaria control if high levels of community-wide ITN coverage can be achieved [31]. Fungal biopesticides may be used in combination with ITN interventions at varying levels of coverage to give greater reductions in malaria transmission and human prevalence. To quantify this effect, the impact of biopesticide application on the equilibrium EIR and human malaria prevalence is explored for fixed levels of ITN coverage . The daily probability of fungal infection during the period of biopesticide exposure (CE) is referred to as the fungal biopesticide coverage throughout this section. The conservative assumptions that mosquitoes are only exposed to a risk of fungal infection during the host-seeking stage (), at a constant rate , and that there are no sublethal effects of fungal infection on mosquito feeding biology () are adopted (Table 1).
Low values of equilibrium prevalence are not obtained by the fungal biopesticide intervention alone, or by the ITN intervention alone unless ITN coverage is high (Figure 6A). When both interventions are used in combination, low prevalence is obtained with moderate coverage of each intervention. The proportional reduction in equilibrium EIR obtained by the combined interventions found to be very close to the multiplicative effect of both interventions, or the product of the proportional reductions given by each intervention acting alone. The difference between the equilibrium prevalence obtained from the combining the two interventions, , and the prevalence corresponding to the multiplicative effect of the two interventions on the equilibrium EIR, denoted , is calculated as . As is small in this case (Figure 7A, open circles), there is negligible redundancy in combining both interventions, and also negligible synergistic effects, or no increase in the impact of one intervention on malaria transmission due to the presence of the other intervention.
The mortality in host-seeking mosquitoes caused by the ITN intervention may be reduced by the development of insecticide resistance in the mosquito population. To give a conservative estimate of intervention performance, the limiting case in which the ITN intervention causes no increase to mosquito mortality is examined. In this case the ITN intervention affects only the blood-feeding rate of host-seeking mosquitoes, by providing a physical barrier between mosquitoes and the human hosts that are protected. For this scenario, low values of equilibrium prevalence are still be obtained by the ITN and fungal biopesticide interventions combined, although higher coverage of each intervention is required (Figure 6B). The baseline equilibrium prevalence is reduced by approximately 50% by the two interventions combined with moderate coverage of each intervention, whereas a similar reduction requires high coverage of each intervention acting alone (Figure 6B).
When the ITN intervention has no mortality effect on mosquitoes, the reduction in equilibrium prevalence obtained by the interventions combined is greater than the multiplicative effect of both interventions (Figure 7A, crosses). The deviation from the multiplicative effect generally increases with increasing ITN coverage (Figure 7A), although not with increasing fungal biopesticide coverage (results not shown). Thus when ITN coverage is high, the addition of the fungal biopesticide has a large impact on prevalence even at low biopesticide coverage (Figure 6B). The combined effect of the two interventions, being greater than multiplicative, is indicative of synergistic interactions between the interventions. This synergism results from the increase in the average time required for mosquitoes to find a blood meal due to the presence of non-lethal bednets, which increases the period of exposure to the fungal biopesticide in a given gonotrophic cycle, and improves the performance of the fungal biopesticide intervention.
Variation in malaria transmission intensity will affect the efficacy of vector control strategies and may alter the appropriate choice of strategy. The case of high malaria transmission intensity is considered here by increasing the recruitment rate ε to give an annual EIR of 412. This value is similar to levels of transmission measured in the high transmission season in the Garki district of Nigeria, an area where malaria control has proven difficult [46].
When transmission intensity is high, low values of equilibrium prevalence are not obtained by either the fungal biopesticide or the ITN intervention acting alone, but are obtained when ITN and fungal biopesticide interventions are combined if ITN coverage is high and fungal biopesticide coverage is moderate (Figure 8A). When ITN coverage is high, fungal biopesticide application is again more effective, producing a sharp decline in prevalence even at low biopesticide coverage. For high transmission intensity, the combined effect of the two interventions is similar to the multiplicative effect of both interventions (Figure 7B, open circles)
As above, the ITN intervention is now considered to have no effect on mosquito mortality, representing widespread insecticide resistance. In this case, considerable reductions in prevalence are only obtained when ITN and fungal biopesticide interventions are combined (Figure 8B). Low equilibrium prevalence is still obtained by the combined interventions for high levels of coverage of each intervention. When the ITN intervention does not affect mosquito mortality, the combined effect of the two interventions exceeds the multiplicative effect of both interventions, to a greater extent than when transmission intensity was lower (Figure 7B, crosses). This demonstrates a case of considerable synergism between the two interventions, whereby the ITN intervention improves the performance of the fungal biopesticide intervention, particularly when ITN coverage is high (Figure 8B).
Low transmission intensity is simulated by decreasing ε to give an annual EIR of 1.5, a value in the lower part of the range reported in [25]. In this case, low values of equilibrium prevalence are obtained by the combined use of fungal biopesticide and ITN interventions for low to moderate levels of coverage of each intervention. If the interventions act alone the level of coverage required to achieve low prevalence is considerably higher, though still moderate (Figure 9).
This study models the effect of fungal biopesticide interventions on rates of malaria transmission with greater biological detail than [21]. While the models are not formally equivalent, the results of this model are consistent with those of [21] for the limiting case in which mosquitoes are continually exposed to a constant risk of fungal infection and there are no sublethal effects of fungal infection on the mosquito blood-feeding biology. The baseline EIR (in the absence of the interventions) is also similar for the two models. The same assumptions about mosquito-fungus interactions described in [21] apply to this model with two exceptions. Firstly, by incorporating gonotrophic structure, this model can consider the effects of fungal infection on host-seeking and non-host-seeking stages of the gonotrophic cycle. Secondly, this model allows the fungal infection rate to vary throughout the gonotrophic cycle.
Firstly, the model results indicate that high virulence will be important to the success of the fungal biopesticide intervention even if fungal infection has a strong effect on both the average time taken for host-seeking mosquitoes to find a blood meal and the time to the next period of host-seeking activity following a blood meal. Secondly, assuming realistically that mosquitoes are not exposed to the fungal biopesticide for their entire gonotrophic cycle, the results indicate that a high daily probability of fungal infection during the period of biopesticide exposure will be important to the success of the biopesticide intervention. This implies an important role for strategies designed to prolong the period of biopesticide exposure, such as the use of African water storage pots sprayed with the biopesticide [20].
Consistent with the goals of Integrated Vector Management frameworks [7],[8],[10], the model explores the total impact of multiple mosquito control interventions used in combination. The principle and practise of combining interventions to give substantial impacts on malaria transmission and manage insecticide resistance is not new [5],[8],[41],[47],[48]. This study explores interactions not previously considered, namely those between ITNs, an widely-used method of controlling malaria vectors, and fungal biopesticides, a novel method of biocontrol that has slow-acting and potentially complex effects on the mosquito life cycle.
In the absence of field data on the combined application of fungal biopesticides and ITNs, conservative assumptions were adopted. Fungal infection was assumed to affect only mosquito mortality, and the exposure of mosquitoes to both interventions was restricted to the host-seeking stage of the gonotrophic cycle. Similar to [31],[49], the equilibrium EIR estimates given by this model are conservative in that they do not take into account changes in human malaria prevalence that may result from the interventions. This assumption may be accurate in the short term given that the lifespan of the vector is considerably shorter than the duration of malaria infection in humans [25]. This SEI model could be extended incorporate human prevalence dynamics [50], however the relationship between human prevalence and rates of malaria transmission is known to be complex and heterogeneous [25].
Low estimates of the equilibrium prevalence were obtained by combining fungal biopesticide and ITN interventions even for scenarios where the impact of each intervention acting alone was relatively small, indicating that the combination is effective. In general, the impact of combining the two interventions on malaria transmission was at least as good as the multiplicative effect of both interventions, which intuitively demonstrates that the combination is efficient. This need not be the case, for example, if large numbers of fungal pathogen-infected mosquitoes are killed by ITNs, the effective coverage of the fungal biopesticide would be reduced. Figure 7 shows that the impact of the combined interventions is sometimes slightly less than multiplicative, which is indicative of this effect. A spatially heterogeneous process, whereby locations that are sprayed with the fungal biopesticide are also those that are best protected by ITNs, may lead to correlation between the probability of fungal infection and the probability of encountering an ITN. This may result in greater redundancy in the combined effect of both interventions compared to the multiplicative effect.
However, as conditions for malaria control become more difficult due to increasing transmission intensity or the development of insecticide resistance, interactions between the two interventions become increasingly synergistic, in that the performance of the fungal biopesticide is enhanced by the ITN intervention. This allowed low prevalence to be obtained for the combined interventions even for high transmission intensity and widespread resistance of mosquitoes to mortality from ITNs. The mechanism underlying this synergism, namely the protraction of the period of exposure to the fungal biopesticide due to the presence of non-lethal ITNs, may be more robust to spatial heterogeneity in the application of both interventions. Deflection of mosquitoes by non-lethal bednets may increase their likelihood of encountering the biopesticide whether it is sprayed at the same location or at another location within the range of mosquito diffusion.
The model results suggest that combining fungal biopesticide and ITN interventions can allow each intervention to be used at lower coverage to maintain a given level of malaria control. This may increase the persistence of each intervention, as lower coverage of each intervention may reduce the selection pressure for the evolution of resistance. However, cautious interpretation is again necessary. The mechanisms behind the success of high community-wide ITN coverage observed in the field may be more complex than those represented by homogenous models of ITN interventions, including the model presented here. Evidence that malaria transmission in human populations is highly heterogenous supports this suggestion [25]. It is encouraging, however, that if high ITN coverage is achieved, the model results indicate that fungal biopesticide application can be very effective even at low biopesticide coverage, particularly when transmission intensity is high and insecticide resistance is widespread.
The model presented here could also be extended to incorporate additional gonotrophic processes with important malaria transmission implications. Firstly, multiple blood feeding within a single gonotrophic cycle could considerably alter transmission patterns, depending on the mosquito life-stages in which it occurs. Multiple blood meals per cycle may be more common in newly emerged Anopheles mosquitoes [51], but can also be more prominent in mosquitoes harbouring infectious sporozoites [52], with the later having the most serious epidemiological implications. Secondly, the time required for blood meal digestion and oocyte development in mosquitoes decreases with increasing temperature, as does the Plasmodium incubation period, leading to substantial variation in these parameters across different field locations [38],[53]. Covariation in duration of the non-host-seeking stage and the Plasmodium incubation period may have a stronger effect on malaria transmission compared to varying each parameter in isolation.
By quantifying the impact of the combined use of fungal biopesticide and ITN interventions on malaria transmission and prevalence, the model indicates that these interventions combined may considerably improve malaria control even in situations each single intervention would have a relatively low impact. Modelling is no substitute for field studies, and attempts to make generalizations about vector biology need to be cautiously interpreted [37]. Recent vector control initiatives encourage the development of models that have the capacity to use field data to guide decision making [7]. This study demonstrates that biological mechanisms relevant to vectorial capacity can be built into existing continuous-time, population-level frameworks to allow direct parameterization from field and laboratory data on both established and novel interventions. This is a means by which models can increase their applicability to integrated vector management strategies.
|
10.1371/journal.ppat.1007515 | Probing the impact of nairovirus genomic diversity on viral ovarian tumor domain protease (vOTU) structure and deubiquitinase activity | Post-translational modification of host and viral proteins by ubiquitin (Ub) and Ub-like proteins, such as interferon stimulated gene product 15 (ISG15), plays a key role in response to infection. Viruses have been increasingly identified that contain proteases possessing deubiquitinase (DUB) and/or deISGylase functions. This includes viruses in the Nairoviridae family that encode a viral homologue of the ovarian tumor protease (vOTU). vOTU activity was recently demonstrated to be critical for replication of the often-fatal Crimean-Congo hemorrhagic fever virus, with DUB activity suppressing the type I interferon responses and deISGylase activity broadly removing ISG15 conjugated proteins. There are currently about 40 known nairoviruses classified into fourteen species. Recent genomic characterization has revealed a high degree of diversity, with vOTUs showing less than 25% amino acids identities within the family. Previous investigations have been limited to only a few closely related nairoviruses, leaving it unclear what impact this diversity has on vOTU function. To probe the effects of vOTU diversity on enzyme activity and specificity, we assessed representative vOTUs spanning the Nairoviridae family towards Ub and ISG15 fluorogenic substrates. This revealed great variation in enzymatic activity and specific substrate preferences. A subset of the vOTUs were further assayed against eight biologically relevant di-Ub substrates, uncovering both common trends and distinct preferences of poly-Ub linkages by vOTUs. Four novel X-ray crystal structures were obtained that provide a biochemical rationale for vOTU substrate preferences and elucidate structural features that distinguish the vOTUs, including a motif in the Hughes orthonairovirus species that has not been previously observed in OTU domains. Additionally, structure-informed mutagenesis provided the first direct evidence of a second site involved in di-Ub binding for vOTUs. These results provide new insight into nairovirus evolution and pathogenesis, and further enhances the development of tools for therapeutic purposes.
| Viruses utilize a variety of mechanisms to manipulate and suppress host responses to infection. One specific mechanism used by nairoviruses is the production of a deubiquitinating enzyme, termed the vOTU, that disrupts the innate immune response. This enzyme has been shown to play a key role in efficient replication of Crimean-Congo hemorrhagic fever virus (CCHFV), a severe human pathogen causing outbreaks with high case fatality rates. Recent genomic studies have revealed a high degree of sequence variation for the vOTU among nairoviruses, but knowledge relating to the functional impact of this diversity is lacking. Here we investigated the effects of this diversity on the structure and function of vOTUs from a wide range of nairoviruses. This revealed that vOTUs from different nairoviruses possess distinct preferences for certain host proteins. In addition, we found that different vOTUs possess distinguishing structural features, including a unique motif present in one that was previously undescribed. Utilizing this information, we were able to provide a rational basis for the observed differences in the vOTUs. This work provides a foundation to understand nairovirus evolution by providing insight into a mechanism that influences virus host adaptation and pathogenesis.
| Nairoviruses are negative sense single stranded RNA [(-) ssRNA] viruses within the order Bunyavirales. Initial classification of nairovirus species relied on antigenic cross-reactivity, leading to the clustering of viruses into seven serogroups; however, with the recent increase in the number of available viral sequences the classifications have shifted to a comparative genomics approach. This not only confirmed the diversity observed based on the serogroup classification, but also further accentuated how these viruses vary across the Nairoviridae family. The family Nairoviridae now consists of approximately 40 viruses that are currently classified into 14 species (Fig 1; [1–5]).
Most nairoviruses are tick-borne viruses infecting multiple vertebrate host species they parasitize in nature. Several have been implicated in human disease, the most notable being Crimean-Congo hemorrhagic fever virus (CCHFV), which has reported case fatality rates in humans that can exceed 30% [6]. Other nairoviruses associated with human disease include Dugbe virus (DUGV), Nairobi sheep disease virus (NSDV) and the Asian variant Ganjam virus (GANV), Erve virus (ERVEV), Issyk-kul virus (ISKV) and Kasokero virus (KASV). These viruses have been reported to cause a myriad of symptoms, some of which include fever, headache, and diarrhea [7–12]. Nairoviruses have also been observed to cause fatal animal disease. For example, NSDV has been reported to have a >90% mortality rate in sheep and goats making it a significant economic as well as human health concern [13]. A recently characterized nairovirus, Leopards Hill virus (LPHV), was isolated from bats and causes severe gastroenteric hemorrhaging and hepatic disease in mice [14]. Hazara virus (HAZV) was isolated from ticks collected from the Royle’s mountain vole and has been proposed as a model system to study CCHFV based on its ability to cause similar fatal disease in interferon (IFN)-receptor knockout mice [15, 16]. Beyond these viruses causing disease in mammalian hosts, other nairovirus have been associated with a broad taxonomic diversity of vertebrate hosts such as birds, fish, and reptiles. For example, viruses in the Hughes orthonairovirus species, such as Farallon virus (FARV), have been implicated in infecting birds [17].
Nairoviruses possess a tripartite genome consisting of small (S), medium (M), and large (L) segments that encode the viral nucleoprotein, glycoproteins, and RNA-dependent RNA polymerase, respectively. Interestingly, the nairoviral L segment also encodes a viral homologue of the ovarian tumor protease (OTU) at the N-terminus. This feature uniquely distinguishes the Nairoviridae family and genus Tenuivirus from other members of the order Bunyavirales. The viral OTU (vOTU) does not appear to play a direct role in genome replication and is dispensable in minigenome replication systems [18]. Instead, the vOTU’s primary function appears to be the reversal of post-translational modifications by ubiquitin (Ub) and the Ub-like protein interferon stimulated gene product 15 (ISG15). This vOTU-encoded deubiquitinase (DUB) and deISGylase activity has been implicated in evading the innate immune response [19–21]. Ub is an 8.5 kDa protein that is involved in a wide range of cellular processes, including key regulatory functions in innate immunity. Ub is conjugated to target proteins by means of a three step process involving activating (E1), conjugating (E2), and ligating (E3) enzymes, and can either occur as a single Ub moiety (mono-Ub) or in polymeric and branched forms (poly-Ub). These chains can be formed by linkage through either the N-terminus (linear) or one of seven lysine residues in Ub (K6, K11, K27, K29, K33, K48, and K63), with different forms often mediating different downstream effects. The most thoroughly studied forms, K48 and K63, play important roles in regulation of the innate immune responses. Specifically, K48-mediated proteasomal degradation has been associated with feedback control, while K63 polyubiquitination is required for pathway activation, including retinoic acid-inducible protein I (RIG-I), mitochondrial antiviral signaling protein (MAVS), Tumor necrosis factor (TNF) receptor associated factor 3 (TRAF3), TANK binding kinase 1 (TBK1), and IFN regulatory factor 3 (IRF3). This signaling cascade leads to the production of IFN-α/β, which ultimately results in the upregulation of numerous IFN-stimulated genes, including ISG15 [22, 23]. The role of ISG15 is complex and not well understood but is generally associated with mediating and regulating antiviral responses both as a co-translational modification and as free ISG15 in the cytosol and secreted form inducing the secretion of IFN-γ and IL-10 by binding cell surface receptor LFA-1 [24–29].
Initial studies on nairoviruses, including CCHFV, DUGV, and NSDV, established the potential immune modulatory effects of vOTU activity based on overexpression of the respective isolated OTU domain in cell culture [19–21]. The ability to probe the specific role of the vOTU during the viral replication cycle remained elusive, however, until the recent development of a reverse genetics system for CCHFV. These studies revealed distinct roles for DUB versus deISGylating activity during the course of a CCHFV infection [30]. Specifically, that CCHFV vOTU DUB activity is not as promiscuous towards ubiquitinated host proteins as it first seemed based on the overexpression studies, but appears to be restricted to a targeted subset of cellular substrates associated with suppression of RIG-I-mediated early cellular responses to infection. In particular, wildtype CCHFV was able to reduce the induction of several immune components, including RIG-I, while CCHFV with a vOTU specifically lacking DUB activity resulted in enhanced cellular responses to infection and establishment of a cellular antiviral state that reduced viral titers. In contrast, deISGylating activity appears to play a role in later stages of CCHFV infection. A recent study demonstrated a similar impact of DUB activity in viral immune suppression during the replication cycle of severe acute respiratory syndrome coronavirus (SARS-CoV) [31]. Specifically, when the DUB activity of the SARS-CoV papain-like protease (PLpro) was selectively disrupted, the virus showed increased sensitivity to IFN and slower growth kinetics. Furthermore, domain exchanges of PLpro’s between different SARS-CoV variants supported this observation, establishing DUB activity to be a distinguishing virulence trait. These emerging insights into the impact of DUB activity in the CCHFV vOTU and SARS-CoV PLpro during viral replication emphasizes the importance of robust DUB activity among pathogenic viruses. The demonstrated vOTU-associated DUB/deISGylase activity of other nairoviruses such as DUGV, ERVEV, and NSDV/GANV, further highlights a potentially substantial role of the vOTU in viral replication and immune suppression for viruses in the Nairoviridae family [19–21, 32].
Remarkably, the nairoviral vOTU domain shows a great degree of sequence diversity, with sequence identities that can drop below 25% between species (Fig 1). A particularly striking case of this diversity is found in members of the Hughes orthonairovirus species, such as FARV, which possess 26–30 additional residues in the middle of the OTU domain (Fig 1B). These sequence differences between vOTUs suggest a plasticity in the OTU domain that could play a role in evolutionary adaptation. Currently, exploration into the phenotypical effects of this diversity has been restricted to only a few taxa that include CCHFV, DUGV, NSDV/GANV, and ERVEV [32, 33]. These studies revealed that vOTUs possess different enzymatic and structural characteristics. In particular, vOTUs display a wide degree of variation in the efficiency with which they engage Ub and ISG15 that is driven by specific sequence and structural features. These substantial differences in viruses within closely related taxa raises questions on the impact of vOTU diversity across the Nairoviridae family. Specifically, how vOTUs from viruses in each species vary in structure and activity, and the implications of this for the potential to suppress the innate immune response and affect viral pathogenesis and host tropism.
To better understand the impact of vOTU diversity, we sought to obtain a more complete perspective of the functional and structural features of vOTUs within the Nairoviridae family. In vitro assays revealed that vOTUs across diverse taxa possess Ub activity, but that activity towards ISG15 appears more restricted. Further characterization of vOTU activity uncovered distinct trends and preferences for specific poly-Ub linkages. To better understand the molecular mechanisms driving Ub activity and specificity, novel X-ray crystal structures were solved revealing features that distinguish the vOTUs from each other, including a pocket that correlates with Ub specificity. Additionally, a structure of the FARV vOTU provides details into the structural nature of the additional residues in Hughes orthonairovirus vOTUs. Structure-informed mutagenesis of FARV vOTU identified residues involved specifically in di-Ub binding, representing the first report of the role of a second site involved in di-Ub binding in nairovirus vOTUs. This novel enzymatic and structural data not only provides insight into the nature of vOTU diversity, but also lays a foundation for understanding the impact of the vOTU interaction with the innate immune response and its connection to viral pathogenesis.
To gauge vOTU diversity across the Nairoviridae family, viruses representing the divergent species were selected and the OTU domain recombinantly expressed. Initially selected based on the traditional serogroups as well as emerging genetic characterization, these viruses include members of the most distantly related taxa and represent 12 of the currently recognized 14 species in the dynamic classification landscape of nairoviruses (Fig 1). Included were the vOTUs from CCHFV, NSDV and GANV, DUGV and Kupe virus (KUPEV), HAZV, Taggert virus (TAGV), ERVEV, FARV, Dera Ghazi Khan virus (DGKV), Huángpí Tick virus 1 (HpTV-1), LPHV, Qalyub virus (QYBV), and ISKV (Fig 1). To better understand the global diversity of nairoviral engagement with Ub and ISG15 substrates, these vOTUs were assessed for activity towards Ub and human ISG15 fluorogenic substrates. These specific activities were measured by the accumulation of the fluorescent molecule 7-amino-4-methylcoumarin (AMC) as a result of cleavage from the C-terminus of Ub or ISG15 (Fig 2).
Intriguingly, the vOTUs showed a diverse range of activity towards Ub. In general, vOTUs can be divided into groups possessing high (CCHFV, HAZV, NSDV/GANV, TAGV), moderate (DUGV, KUPEV, FARV, QYBV, ISKV), or low activity (ERVEV, DGKV, LPHV, HpTV-1) (Fig 2A). For some of these vOTUs, their deubiquitination activity mirrors that observed in DUB-deficient CCHFV mutants that impact cellular ubiquitination levels leading to an impaired ability to suppress the IFN response [30, 34]. To a large degree, viruses more closely related phylogenetically with CCHFV possess the most robust activity (Figs 1A and 2A). Beyond this, there is not an obvious phylogenetic trend to how well the vOTUs cleave Ub-AMC, with disparate taxa showing similar low to mid-range activity. Overall, engagement with Ub is observed to be a feature that can be present in diverse species in the Nairoviridae family, with some taxa demonstrating enhanced activity.
The patterns of activity for Ub are in stark contrast to those of ISG15-AMC, which shows a more dichotomous pattern as a substrate for the vOTUs (Fig 2B). Specifically, there appears to be an abrupt break phylogenetically between groups that contain vOTUs with deISGylating activity, compared to others for which activity is almost negligible. This break appears to exist at the node separating the Thiafora, Artashat, Sakhalin, Crimean-Congo hemorrhagic fever, Nairobi Sheep Disease, Dugbe, and Hazara orthonairovirus species from the remaining seven (Fig 1A). Interestingly, the presence of ISG15 activity does not encompass every vOTU in these species, suggesting individual factors may have driven the development or retention of ISG15 activity for viruses within this clade. Naturally, this also implies that DUB activity could be a more broadly utilized mechanism to evade cellular responses. This led us to further explore the dynamics of different nairovirus vOTU’s interactions with Ub.
While the Ub-AMC assay provides general information on the ability of vOTUs to engage monomeric Ub, cellular substrates are typically modified by poly-Ub chains through various linkage types [35]. Additionally, DUBs in general and vOTUs in particular have been observed to prefer some linkage types over others [32, 36–40]. To assess the patterns of Ub linkage preferences of diverse vOTUs, a subset of the vOTUs were analyzed against di-Ub FRET-TAMRA substrates (Fig 3A). HAZV vOTU and GANV vOTU were selected because they have been considered to be a potential model system for CCHFV and have significant health and economic impact, respectively. TAGV vOTU represents a more distantly related vOTU that also has substantial DUB activity, while the vOTUs encoded by QYBV, FARV, and DGKV display diminished activity towards mono-Ub. To reduce the influence of interactions with the FRET pairs that may disrupt interaction, multiple FRET pair configurations were assessed when available and the one displaying the highest activity selected (two positions each for K48 and K63; S1 Fig).
Comparison of vOTU activities towards different di-Ub FRET substrates reveals that each species’ vOTU has distinct preferences for specific di-Ub linkages. While HAZV vOTU and GANV vOTU both possess notable activity for K48 and K63 di-Ub, there appears to be more substantial activity towards K11. TAGV vOTU, on the other hand, prefers K63 and K11 to a greater extent, while not possessing as much activity towards K48. For FARV vOTU, the opposite is observed, with K48 being preferred. DGKV vOTU, consistent with its low Ub-AMC activity, possesses very low activity for the di-Ub FRET substrates, regardless of the linkage. Similar to the pattern observed for HAZV and GANV, QYBV vOTU shows the most activity towards K11, though at lower overall levels. Additionally, QYBV vOTU shows a more pronounced difference in the relative preference for K48 versus K63 linkages, with substantially more activity towards K48.
Regrettably, the commercial availability of FRET-TAMRA di-Ub substrates is restricted to these tested linkages. Additionally, limitations are known to exist due to how the positions of the donor-quencher pairs affects binding of these substrates with the proteases. To gain a more complete and natural perspective of di-Ub linkage preferences, these vOTUs were also assessed by SDS-PAGE for the ability to cleave unlabeled di-Ub substrates of all eight linkage types (Fig 3B). As expected, the vOTUs did not show equal preferences for the different linkages. Intriguingly, some of the results appeared to differ from the FRET analysis. Specifically, for both HAZV vOTU and FARV vOTU the gel cleavage assay would suggest the K63 activity to be the highest with K48 and K11 roughly equal, suggesting that the positions of the donor-quencher pairs may have hindered binding to some of the substrates. As expected, the HAZV, GANV, and TAGV vOTU showed substantial cleavage of several di-Ub substrates that is consistent with their high Ub-AMC activity, while the DGKV and QYBV vOTUs showed low level of substrate cleavage over the same time course. Intriguingly, FARV vOTU showed substantial cleavage of some of the substrates, despite not possessing high Ub-AMC activity. This may be reflective of differences in the assays in measuring DUB activity. Alternatively, it may also suggest the existence of an additional site of interaction with the proximal Ub that enhances the efficiency of di-Ub cleavage.
Assimilation with previously reported data reveals interesting trends and points of divergence between the vOTUs ([32]; Fig 3C). Linear and K29-linked di-Ub does not show any sign of cleavage with any vOTU. vOTUs demonstrate varying levels of low/detectable activity on K27-linked and K33-linked di-Ub. In contrast, vOTUs show a consistent pattern of higher activity towards K6, K11, K48, and K63 di-Ub. While most of the vOTUs show some level of enhanced activity towards these linkages, the specific linkage most preferred can differ. The CCHFV, HAZV, TAGV, and FARV vOTUs all show a distinct preference for K63 over K48 linkages, while DUGV and QYBV vOTU show more activity towards K48. GANV vOTU shows approximately equal preference for these linkages. Interestingly, GANV vOTU also shows more activity towards both K6 and K11 di-Ub at approximately equal levels. The high degree of preference towards these substrates extends to the majority of the vOTUs, as even for DGKV vOTU, which shows minimal or no cleavage of most of the substrates, cleavage of K6-linked di-Ub can be identified within an hour (Fig 3B). The other vOTUs, with the exception of ERVEV, all show detectable levels of K6 cleavage, with most also cleaving K11. The CCHFV, HAZV, DUGV, and FARV vOTUs all showed a greater relative preference for K6 over K11, while TAGV vOTU was the opposite. QYBV vOTU, similar to GANV, showed approximately equivalent activity towards K6 and K11, though overall activity was lower. Overall, these patterns of activity suggest that vOTUs do not merely cut any Ub moiety, but that they are specific to a subset of linkages that may influence specific aspects of cellular biology.
To gain a better understanding of how sequence diversity translates into structural differences, X-ray crystal structures were sought of vOTUs from divergent species. The vOTUs from DGKV, QYBV, and TAGV readily crystallized and were solved to 1.62 Å, 1.65 Å, and 2.05 Å, respectively (Table 1). These vOTUs represent diverse nairovirus species, and possess extensive variation in Ub activity with the DGKV, QYBV, and TAGV vOTUs possessing low, medium, and high activity towards Ub-AMC, respectively (Fig 2). TAGV provides a glimpse into the Sakhalin orthonairovirus species, a taxon that is more closely related to the ERVEV and CCHF/NSD/Dugbe/Hazara cluster, while DGKV and QYBV are from the Dera Ghazi Khan and Qalyub orthonairovirus species, respectively, and are much more distantly related (Fig 1).
These structures reveal global similarities among the vOTUs (Fig 4 and S2 Fig). Each vOTU possesses a seven-stranded beta sheet as the core feature, with five major alpha helices framing the rest of the structure. The catalytic triads perfectly superimpose over each with the exception of DGKV vOTU (Fig 4A). In DGKV vOTU, aspartate is replaced by a glutamate that alters the spatial dynamics of the catalytic triad, possibly contributing to a less rigid structure that allows the histidine to adopt the alternate conformation. While atypical for the vOTUs, it does not appear to be the cause of DGKV’s low DUB activity, as mutating the glutamate to an aspartate only further diminished activity. Looking beyond the catalytic triad, a structural overlay of the vOTUs highlights a point of difference in the overall structure that distinguishes the proteases from each other. Specifically, there appears to be substantial variability in the region encompassing the α3 (“selectivity”) helix that has been associated with substrate preference, and the loop between the β1 and β1a strands (Fig 4A; [32, 33]). Comparing the root mean square deviation (R.m.s.d.) for the positions of the main chain atoms of these different structures further emphasizes how they deviate from structure to structure (S2 Fig).
Closer examination of the structures reveals distinct molecular interactions that account for these observed structural differences within the vOTUs. Specifically, particular amino acid differences can be identified that form interactions that would promote the observed conformation of each protease, suggesting these differences to not merely be a consequence of dynamics or crystal packing. Intriguingly, these residues are not limited to just the selectivity helix and β1-β1a loop but extend to other secondary structural elements in local proximity, forming an ensemble of interactions that drive the noticeable variability of the α3 region (Fig 4B and 4C). The first area of prominent influence centers around position 73 of CCHFV vOTU (Fig 4C, Panel I). This position is strongly conserved in possessing an aromatic residue, consisting of a phenylalanine or a tyrosine in CCHFV and more closely related viruses, while consisting of a tryptophan in the rest of the vOTUs studied (Fig 1B). While subtle, this change results in distinctly different local interactions that influence the positioning of the selectivity helix. In the CCHFV vOTU, Tyr73 forms hydrogen bonds with the backbone of Leu84 within the α3-α4 loop. In contrast, the tryptophan residues in the other vOTUs fill into a hydrophobic cleft that involves residues within the selectivity helix. In DGKV vOTU, Trp73 packs with the methylene group of Ser86 on one side and the aliphatic portion of Lys80’s side chain on the other. Lys80 itself is stabilized in this permissive conformation by hydrogen bond pairing with the carbonyl of Gly74. In QYBV vOTU, Trp72 packs with Pro85. Additionally, it is in proximity to His79, suggesting potential stacking of the rings. Such an interaction could have an indirect effect on the positioning of Pro76 at the surface of the vOTU-substrate interface. In TAGV vOTU, Trp73 fits between Leu86 and Thr79.
The second area centers around helix α5 and shows a great degree of variation between the vOTUs (Fig 4C, Panel II). It can encompass interactions that can extend to the α2 and α3 helices as well as the β1a sheet with the potential to influence the local structural architecture. CCHFV vOTU possesses a number of interactions within this region, including unique lysine pairings consisting of Lys71, Glu111, Lys110, and Glu76 that accommodate hydrophobic packing with Tyr72 and Phe133. Along with hydrogen bond pairing of Glu78 with Thr102, these work in conjunction in orienting the position of the selectivity helix, a region that has been implicated in substrate preference [32, 34]. The other vOTUs, in contrast, possess fewer interactions but still could influence the structure. In DGKV vOTU, Gln78 within the selectivity helix is central to the interaction, pairing with both Ser102 and Tyr133 by hydrogen bonding. Similarly, for TAGV vOTU Thr76 and Glu78 form electrostatic interactions with Gln110 and Tyr133, respectively. In contrast, for QYBV vOTU there do not appear to be any direct interactions with the selectivity helix. Instead, Lys109 and Ser108 form electrostatic interactions with Asp15 and Asn16, respectively, suggesting a role in manipulating the positioning of the β1-β1a loop.
The third region consists of the selectivity helix and β1-β1a loop themselves (Fig 4C, Panel III). Specifically, direct interactions, or the lack thereof, work in conjunction with the other interactions to complete the structural features. This is most notable in QYBV vOTU, in which Phe14 appears to pack with Pro76. In CCHFV vOTU, the corresponding residue is Ile13, which does not appear to be able to bridge the distance and form an interaction. In TAGV vOTU, Asn14 and Thr81 are in general proximity, but appear to be too distant to form a strong interaction with each other. Similarly, DGKV vOTU appears to lack any direct interactions. Overall, these interactions contribute to structural features that influence spatial and chemical presentation of the vOTU interface, potentially affecting how these vOTUs engage substrates.
Binding with Ub often centralizes around the specific hydrophobic residues Leu8, Ile44, and Val70 [41]. Looking at X-ray crystal structures of the CCHFV and DUGV vOTUs bound to Ub reveals that Leu8 in particular has to be spatially accommodated in a pocket deep within the interaction interface (Fig 5A). To confirm that this interaction with Leu8 is likewise involved in Ub binding with these other vOTUs, isothermal titration calorimetry (ITC) was performed using the TAGV vOTU to determine the relative binding efficiency of alanine and asparagine Ub mutants (Ub-L8A and Ub-L8N) compared to WT Ub (Table 2, S3A Fig). This revealed a stark difference in the affinity. While WT Ub bound strongly with a dissociation constant (KD) of 11.5 ± 2.5 μM, Ub-L8A showed no detectable binding under similar conditions. The Ub-L8N mutant faired only slightly better than the Ub-L8A mutant with a 20 times weaker dissociation constant, KD of 295.3 ± 39.7 μM, compared to WT. These results further underscore the importance of vOTUs being able to accommodate Ub Leu8 in order to have robust deubiquitinating activity. Examining the analogous residues in the other vOTUs reveals a diverse composition for this pocket that could influence how well Leu8 can be accommodated. In TAGV vOTU, this pocket is largely hydrophobic possessing two tyrosines as well as a valine. In contrast, the DGKV and QYBV vOTUs possess more polar residues, including asparagine, glutamate, and threonine for DGKV and glutamine and lysine in QYBV. When considering the activity towards Ub based on the AMC assay, a general trend emerges that correlates the degree of an enzyme’s ability to engage mono-Ub with the hydrophobicity of this pocket.
Looking more closely at this interface suggests an additional nuance to the ability to accommodate particular substrates. Specifically, spatio-chemical characteristics could largely influence what defines a good or acceptable pocket composition for binding a given substrate. In the vOTUs that most effectively engage with Ub, such as CCHFV, HAZV, NSDV/GANV, and TAGV, the residue that most directly interfaces with Ub’s Leu8 is an isoleucine, valine, or threonine that corresponds to position 131 in CCHFV vOTU (Figs 1B, 2 and 5A). This correlation is consistent across the Nairoviridae family. Despite being phylogenetically distant from CCHFV and the other robust vOTU DUBs, QYBV demonstrates substantial Ub activity and possesses Ile130 that could pack with Ub’s Leu8.
In other vOTUs that have poor Ub activity this residue is typically polar, such as ERVEV’s Asn134 (Fig 5B, Panel I). This creates an environment that would discourage binding with Ub. Mutation of this residue in ERVEV to the corresponding hydrophobic residues in CCHFV has been observed to generate robust Ub activity [33]. The FARV and ISKV vOTUs appear to have similar characteristics, with both encoding a glutamine at this position. Intriguingly, other vOTUs may go even further in discouraging Ub binding. These include LPHV and HTV-1, which possess an arginine and lysine, respectively, at their equivalent positions to Asn134. Modeling in an arginine at this position, such as what LPHV vOTU possesses, reveals that this type of change would be prohibitive for Ub binding (Fig 5B, Panel II).
To test the central role of this pocket for Ub activity, a series of mutants were made in the DGKV, QYBV, and TAGV vOTUs and tested against Ub-AMC (Fig 5C). As expected, the disruptive mutants I130R in QYBV and Y129R and V131R in TAGV completely knocked out the ability to process Ub-AMC, in keeping with a previous report demonstrating that the presence of arginine hindered ERVEV vOTU Ub activity [33]. Further, increasing the hydrophobicity of the pocket in QYBV vOTU was able to enhance Ub-AMC cleavage, boosting activity by 150% and 50% for the Q19V and Q128A mutants, respectively. Interestingly, changing the pocket in DGKV vOTU failed to improve activity. This suggests that some vOTUs lacking any outright DUB activity may have evolved to a degree that prevents the generation of this activity through simple changes to better accommodate Leu8 in Ub. This leaves the presence of a Ub Leu8 accommodating pocket in vOTUs as a major marker for deubiquitinating activity and, if present, the ability to dictate variable levels of activity based on the hydrophobicity.
While the vOTUs show a range of sequence diversity, most of them possess domains of approximately the same size with one notable exception. Viruses in the Hughes orthonairovirus species possess an additional 26–30 amino acids in the middle of the vOTU. When aligned with the other vOTUs, this extra sequence corresponds to the region of the α3/α4 helices. To assess how this additional sequence influences the structure of vOTUs in this subset of nairoviruses, a crystal structure of FARV vOTU was solved to 2.22 Å (Table 1). Inspection of the structure immediately revealed the impact of this additional sequence on the protease’s structure (Fig 6A). While possessing the familiar core domain and secondary structure features, the protease possesses extended α3/α4 helices that are connected by several intervening residues. Thirteen residues could not to be built due to a lack of well-defined electron density in the crystal structure, and those that could be modeled possess high B factors, suggesting this region to have a high degree of flexibility. This contrasts with the vOTUs from CCHFV and other nairoviruses that possess relatively small α3/α4 helices connected by a short loop (Fig 6B). Additionally, the α3/α4 helices of FARV vOTU appear to interact with each other in a manner resembling a coiled-coil motif. This is facilitated by hydrophobic packing between Ile86, Val105, Ala82, and the aliphatic portion of the Arg108 sidechain. This relationship between the helices is further promoted by electrostatic interactions, including a salt bridge between Arg81 and Asp116, as well as a hydrogen bond between Tyr78 and Asp116. Beyond this interaction, Tyr78 is also positioned to hydrophobically interact with Tyr113, which together create an environment in which Trp68 can insert. Trp68 further promotes the interaction between these helices through a hydrogen bond with Asn75, as well as through additional hydrophobic packing with Lys79 and Leu110.
The presence of the extra sequence/structural motif in the Hughes orthonairovirus species raises the question of whether it could be involved in substrate interaction. A model of how FARV vOTU could interact with Ub further accentuates this possibility, suggesting the α3/α4 helices to be in close enough proximity to participate in binding (Fig 7). Such an interaction could potentially offset other factors in FARV vOTU are not ideal for binding. Looking at the selectivity pocket of FARV vOTU reveals it to possess more of a hydrophilic character and contains a relatively bulky Gln155 residue in the equivalent site to position 131 in CCHFV (Fig 7A, Panel I). Additionally, FARV vOTU possesses a potential steric hindrance to efficient binding with the presence of Arg170 (Fig 7A, Panel II). This residue may be less accommodating for the Leu73 in Ub than other vOTUs, such as CCHFV and TAGV which contain a histidine at this site. Further, FARV vOTU may lack a significant interaction that CCHFV vOTU possesses with Arg42 of Ub (Fig 7A, Panel III). In contrast to Gln16 in CCHFV vOTU that is able to form a hydrogen bond, FARV vOTU possesses a leucine that is unable form this interaction. To test the influence of these sites on DUB activity, mutations were made to Arg170 and Leu13 in FARV vOTU to histidine and glutamine, respectively. As anticipated, R170H was able to improve Ub-AMC activity, boosting it by ~250%. Making the reverse mutation in TAGV vOTU, H146R, essentially knocked out this activity suggesting this residue to have a key impact in diminishing FARV vOTU’s activity compared to other vOTUs. Interestingly, the L13Q mutation in FARV vOTU led to a large reduction in Ub-AMC cleavage. Looking more closely at this region shows that Leu13 is in the middle of a large hydrophobic region in FARV vOTU (S3B Fig). The swap to a large polar residue may impact the structural integrity of the β1-β1a region, further underscoring the nuances created by the variability of this region.
To probe the potential significance of the α3/α4 motif in offsetting these other effects in FARV vOTU, a construct was synthesized lacking residues 79–107 (“FARV vOTUΔ79–107”) and assessed for activity against Ub substrates (Fig 7B). Removing this region reduced activity towards Ub-AMC by almost 60%, suggesting that this motif could play a significant role in Ub binding (Fig 7B, Panel I). Interestingly, when tested against K48 and K63 FRET di-Ub substrates a more modest reduction in activity is observed, with only about a 30% and 40% reduction in activity, respectively. This is further borne out with unlabeled di-Ub, with there being no substantial difference between the WT and Δ79–107 vOTUs over the longer reaction time course (Figs 3B and 7B, Panel II).
Although the di-Ub cleavage assays are able to differentiate linkage preference, the structural architecture is still relatively simple. To gauge whether this motif can engage with more complex poly-Ub structures, the WT and Δ79–107 FARV vOTUs were tested with K48 and K63 linked tri-Ub (Fig 7B, Panel III). Interestingly, both constructs showed a clear preference for K48 over K63 tri-Ub. This is in contrast to the gel cleavage assay for di-Ub, which showed a slight preference for K63 di-Ub (Figs 3B and 7B, Panel II). Beyond this, both constructs showed similar patterns of activity for these substrates.
Despite possessing low to moderate activity towards Ub-AMC, FARV vOTU possesses substantial activity towards some di-Ub linkages (Figs 2 and 3). This suggests an additional site of interaction with the proximal Ub molecule that substantially increases the overall efficiency. To ascertain where this site may be located, a model of how FARV vOTU may bind di-Ub was generated (Fig 8). Examining the potential interface with the proximal Ub, two residues in FARV vOTU, Arg30 and Lys32, immediately stand out as potential contributors. These residues are just beyond the active site, and are part of a region that likely forms the closest contact with the proximal Ub. Beyond these two residues, Thr147 of FARV vOTU also stands out as being in an area with a higher R.m.s.d. between the vOTU structures, which in FARV vOTU positions it closer to the general area of the proximal Ub (Fig 1B and S2 Fig).
To assess whether these sites could play a role in the FARV vOTU’s interaction with di-Ub, mutations at these positions were designed in an attempt to alter activity towards K48 and K63 FRET di-Ub substrates. As a control, each mutant was also run with mono-Ub substrates. Due to the proximity of Arg30 and Lys32 to the space that would be occupied by the fluorogenic molecule, assays were performed with both Ub-AMC and Ub-Rhodamine110 (Rh110) to mitigate artifacts. Excitingly, these mutations substantially altered the rate of di-Ub cleavage, often towards both substrates (Fig 8). Individually mutating Arg30 and Lys32 to leucine reduces activity towards K48 di-Ub to 43–55% of wildtype and activity towards K63 to 56–70%. Although the mono-Ub activity appears to suffer as well in the case of R30L, the ~30% difference in the Ub-AMC versus Ub-Rh110 mono-Ub substrates suggests this to be an artifact of interactions with the AMC fluorophore. Otherwise, these mutants had little or no effects on mono-Ub activity. When a charge flip was introduced at position 30, activity was reduced to 18% for K48 and 35% for K63 while not substantially altering the activity for mono-Ub. A charge flip at position 32 had the most pronounced effect, driving it down to 7% for K48 and 38% for K63. However, there is also a substantive corresponding reduction in both the AMC and Rh110 mono-Ub substrates, indicating a potentially large disruptive interaction with the hydrophobic fluorophores or possible influence on the local fold. Interestingly, while mutating Thr147 to valine did not appreciably change the activity, introducing an arginine at this position increased the activity by 15% for K48 and doubled it for K63, suggesting the longer sidechain may be able to form a new interaction.
Ub is among the most conserved and important cellular regulatory components, influencing almost every key aspect of cell biology. Ub itself is tightly regulated by an array of endogenous DUB enzymes that specifically curb and tailor its effects. The realization that viruses also possessed enzymes with DUB activity introduced a paradigm in which these normal regulatory mechanisms could be manipulated to suppress immune responses and enhance viral propagation [19, 42–45]. Further investigation into these mechanisms continues to uncover how these viral DUBs disrupt cellular responses to infection. In particular, the role of robust DUB activity in promoting viral replication and conferring virulence in CCHFV and SARS-CoV emphasizes the impact of the respective proteases and highlights the emerging importance of understanding their effects when considering potential pathogenicity and therapeutic strategies.
With the almost perfectly conserved sequence of Ub, it is not surprising that tick-borne nairoviruses from disparate taxa possess notable DUB activity. Such a mechanism could provide broad utility in infecting hosts beyond the primary arthropod reservoir by enabling a route of horizontal as well as vertical transmission that amplifies viral replication. The diversity in the observed activity, however, raises questions as to the specific effects relating to arthropod versus vertebrate hosts. In general, the vOTUs from viruses most closely related to CCHFV appear to have the most substantial DUB activity based on the Ub-AMC assay (Figs 1 and 2). These viruses are known to cause viremia in vertebrate hosts, including mammals. This raises the prospect that increased Ub activity may be an adaptive mechanism allowing these nairoviruses to infect a wider host range. While ticks are known to possess RNAi and Toll sensing-mediated antiviral responses, there is little information pertaining to whether Ub plays a significant role in arthropod responses to viral infection [46–48]. Further characterization of Ub systems in arthropods will be needed to shed light into these questions and would clarify the significance of vOTU enzymatic diversity in nairovirus adaptation for arthropod versus vertebrate hosts.
In contrast to some mammalian DUBs, OTU proteases generally show poly-Ub linkage specificity that ranges from moderate to highly specific [38]. Nairovirus vOTUs reflect this tendency, possessing activity towards poly-Ub linkages that is neither highly promiscuous nor completely selective for a single linkage type with each vOTU possessesing its own respective preferences for the different linkages. While showing individual variation, the vOTUs consistently show the ability to process K6, K11, K48, and K63 linked di-Ub. The fact that nairovirus vOTUs generally show activity towards K48 and K63 di-Ub is significant. These well-studied forms of di-Ub have clearly established roles for cellular processes in general, as well as in antiviral responses specifically. It can be easily envisioned how disruption of K48 and K63-mediated functions could dampen antiviral responses. It is intriguing, however, that vOTUs as a whole also possess substantial activity towards K6 and K11 linkages. These forms of Ub have been studied much less extensively, with roles that have typically been associated with DNA damage responses and cell cycle regulation [49, 50]. As part of the L protein, vOTU activity would be restricted to the cytosol, raising questions as to whether these observed activities of vOTUs are incidental, or if there are important cytosolic functions of K6 and K11 linkages that could be manipulated. Recent studies have begun to expand knowledge of these linkages. Specifically, K11 poly-Ub has been associated with TNF signaling, providing a direct link to the innate immune response [51]. Even more recently, K6 has emerged as a key component in regulating mitophagy [52, 53]. Given the key role of mitochondria in innate immunity, this raises an interesting question of how vOTU activity could impact this process, and whether such manipulation could provide benefits for the virus [54]. What other functions remain to be identified for these linkages is still an open question, as well as how vOTUs may engage with them to modify cellular responses. The differences in linkage preferences between vOTUs implies potential differences in the degree to which specific viruses may influence these pathways. Alternatively, it’s possible that the relative importance of the linkages may differ in different hosts, and that different vOTU preferences reflects virus adaptation to their specific preferred hosts.
The new vOTU structures reveal an array of conserved and divergent features. The conserved elements of Nairovirus vOTU structure distinguishes these from other OTU proteases, as highlighted by how they cluster together in a structure-based phylogenetic tree (Fig 1A, inset; [34, 55, 56]). Most notably, this includes the presence of two additional beta sheets and a helix at the N-terminus of vOTUs that are absent from eukaryotic OTUs. While possessing these characteristic features of the vOTU fold, the nairoviruses show distinguishable differences from each other that can be traced to specific residue differences. This is particularly noteworthy when looking at the relationship between the selectivity pocket and the observed Ub activity by a given protease. It is significant that vOTUs possessing the highest activity for Ub all possess highly hydrophobic residues in this region. While many of the vOTUs possessing robust DUB activity are closely related phylogenetically, the presence of substantial activity in the more distantly related QYBV vOTU demonstrates that it is not exclusive to this subset of viruses. This suggests the vOTU fold to be a flexible platform that has allowed DUB activity to evolve independently to the benefit of each virus. Beyond the central role of this pocket that is deep in the binding interface, the vOTUs also display structural diversity in more peripheral regions. This includes areas that have been observed to influence substrate binding in vOTUs, such as the α3 selectivity helix, suggesting a potential impact on how vOTUs engage with other proteins.
Viruses in the Hughes orthonairovirus species possess a motif previously unobserved for OTU domains. This raises the question of whether this structural feature could have a functional impact. In particular, whether this motif could impact engagement with substrate. The effect of removing this motif from the FARV vOTU on Ub-AMC activity suggests that it can at least contribute to mono-Ub binding. This is consistent with what is observed when comparing FARV vOTU to a Ub-bound structure of CCHFV vOTU, where elements of this motif are in proximity for potential interactions with Ub (Fig 7A). This additional interface provided by the motif likely compensates for the presence of other less optimal factors for Ub binding, including an arginine that hinders interaction with the tail of Ub. Overall, these structural features suggest a mixture of elements that either promote or hinder interaction, with some that may carry a more dominant effect. The involvement of the structural motif in FARV vOTU formed by the two helices and intervening loop, which we also refer to as a substrate interacting bundle (SIB), suggests it may form a region that introduces potential to engage with otherwise inaccessible surfaces.
Interestingly, removing the SIB motif from FARV vOTU appears to have a lesser impact on di-Ub activity compared to mono-Ub (Fig 7B). This could be accounted for by the presence of an additional site of interaction in FARV vOTU that interacts with the proximal Ub. The existence of one or more subsites has been postulated as a mechanism for discriminating different di-Ub linkages based on the proximal Ub, and has been demonstrated in several mammalian OTUs [38, 39]. While not definitively observed in vOTUs, the ability to distinguish between different linkages implies a similar mechanism. The FARV vOTU mutants provide the first reported direct evidence identifying such a site in a vOTU, confirming that vOTUs can utilize this mechanism to distinguish various linkages. In addition to supplying potential leads for elucidating such sites in other vOTUs, it also demonstrates a case where this site can have a major impact on activity towards substrate, even when factors hindering binding with the distal Ub are present.
Although di-Ub wouldn’t directly interact with the SIB motif in a manner that would directly influence cleavage, it is possible that a more complex poly-Ub structure could engage with it. Modeling how tri-Ub might bind suggests that a K48 linkage could place one of the Ub molecules in close proximity to the structural motif (Fig 7C). In contrast, for the other linkages FARV vOTU most readily cleaves—K6, K11, and K63—this Ub would likely be too distant to form any interaction. Surprisingly, removal of the SIB motif has no noticeable impact on K48 tri-Ub cleavage, despite the apparent proximity the tri-Ub could have. It’s possible that tri-Ub may not possess a large enough architecture to be influenced by the SIB motif, and that a longer poly-Ub may interact with it. Additionally, Ub is able to form complex chains consisting of multiple linkage types [57]. It may be that the SIB motif can engage more effectively with these “heterotypic” Ub chains. Alternatively, the primary role of the SIB motif may go beyond Ub and facilitate interactions with other binding partners. The vOTU domain exists in the context of the multifunctional L protein. Apart from the vOTU domain, the structural features and dynamics of the nairovirus L protein are currently unknown. This leaves open the possibility that the SIB motif could be involved in binding another feature of the L protein to stabilize the overall architecture, or in facilitating interactions with other proteins. In addition, the SIB motif could potentially bind to other host factors in the innate immune system. Viruses in the Hughes orthonairovirus species have been isolated from birds or from ticks that infest them. The immune system of birds, including antiviral responses, possesses considerable differences from mammals in terms of what elements are present and how they are regulated (reviewed in [58] and [59]). This includes the apparent absence of an ISG15 homologue in birds. These differences from mammals raises the possibility that the SIB motif could play a role in adaptation to the avian innate immune system, perhaps by facilitating interactions with proteins other than Ub. In addition, the lack of ISG15 in birds leaves open the possibility that the motif could engage with other Ub-like entities that are involved in regulating the innate immune response.
While divergent vOTUs possess the ability to engage with Ub, it is possible that this may not be the only, or even predominant function of all vOTUs. In the case of ERVEV, it has been observed that it possesses poor activity towards Ub, while showing potent ability to engage with ISG15 (Fig 2; [32, 33]). This raises the possibility that other vOTUs that possess poor Ub activity may be able to engage with other Ub-like entities. While none of the new vOTUs assessed possess notable deISGylase activity, the availability of AMC-derived substrates is limited to human ISG15 (hISG15). In contrast to Ub, ISG15 shows considerable species-species variances that have been shown to impact binding with viral proteins, including vOTUs from nairoviruses [33, 60, 61]. This leaves open the possibility that vOTUs, while not engaging with hISG15, may still possess the ability to interact with ISG15 from species they productively infect. The presence of arginine, lysine, or glutamine in the selectivity pocket of several of the vOTUs, while not ideal for Ub, may still allow them to engage with other substrates. The structure of the ERVEV vOTU bound to mouse ISG15 (mISG15) has a gap in the area that Ub’s Leu8 would typically occupy (Fig 5B, Panel I). Modeling suggests that this feature would also be more permissive of binding with vOTUs possessing a bulky residue such as arginine at position 131 (Fig 5B, Panel II). This gap is caused by a pairing of Glu87 with Lys148 in mISG15 that pulls the sidechain of Glu87 away from the interface, and suggests a possible mechanism that could allow vOTUs with hydrophilic or bulky residues to effectively engage with non-Ub moieties. As highlighted by the lack of ISG15 in birds, however, it’s also possible that vOTUs, particularly in the Hughes orthonairovirus species, may engage with other Ub-like entities that can modulate the immune response. The lack of either Ub or ISG15 activity in a number of vOTUs further accentuates this possibility, implying possible biochemical functions that have yet to be characterized among vOTUs. Further developments shedding light on these questions could yield key insights into these influential virus-host interactions.
The recent increase in genomic characterization of nairoviruses has uncovered a wealth of diversity among them. While our knowledge of nairoviral sequence diversity has expanded, much is still unknown on how this variability affects virus-host relationships. The exact range of vertebrate hosts and their disease state upon infection is not presently known for all members of the Nairoviridae family. This novel characterization of nairovirus vOTUs reveals a diversity in the ability to engage mono- and poly-Ub that mirrors the genomic diversity. Additionally, this study uncovers motifs that appear to play a predominant role in determining these preferences, making it feasible to begin predicting DUB activity in uncharacterized or newly discovered nairoviruses. Given the presence of robust DUB activity in nairoviruses known to infect humans, including CCHFV and NSDV, this could serve as an early flag for assessing the risk posed by emerging viruses, and may shed light on the evolutionary trends leading to some viruses to having this capability over others. Further, these new structure and activity insights provide a platform to continue the development of robust tools, such as poly-Ub specific vOTUs, that can be paired with reverse genetics systems to better understand the role of the vOTU in the course of a viral infection and how differences in certain activities impact nairoviruses. Such knowledge could help propel the field in fully elucidating the detailed functional mechanism of the vOTU in the viral life cycle, potentially aiding in the development of better disease model systems. In addition, it provides insight that will further gauge the prospects of the vOTU as a therapeutic target for nairovirus-caused diseases such as CCHF, either through the development of specific inhibitors or live attenuated virus vaccines. Further, the diversity of the vOTU suggests a potential relationship with viral host adaptation, and that the role of the vOTU may extend beyond its well-known function in engaging with Ub and/or ISG15.
The vOTUS were constructed and expressed as previously described in published methods [32, 55]. Purification of QYBV, TAGV, and DGKV were carried out as previously described. For FARV, a slightly different approach altered from the previously described method was used to optimize the expression. E. coli strains with vOTUs from FARV were grown at 37°C in 6 L of Luria-Bertani broth with 100 μg/ml ampicillin. Once the optical density reached 0.6–0.8, 0.8 mM isopropyl-β-D-thiogalactopyranoside (IPTG) was added to induce gene expression. The temperature was then dropped to 25°C and expression continued overnight. The culture was subsequently centrifuged at 5000xg for 10 minutes and the pelleted cells stored at -80°C.
Assays were carried out as described previously [32, 33]. Briefly, assays were run in 100 mM NaCl, 50 mM HEPES [pH 7.5], 0.01 mg/mL BSA, 5 mM dithiothreitol (DTT) at 25°C. Reactions were run in 96-well plates with a 50 μl reaction volume using a CLARIOstar plate reader (BMG Labtech, Inc.). For Ub-AMC, all vOTUs were assessed at a final enzyme concentration of 4 nM. For ISG15-AMC, vOTUs were assessed at a final enzyme concentration of 20 nM with the exception of NSDV, GANV, and ERVEV, which were run at a final enzyme concentration of 4 nM due to the high activity towards the substrate. Both Ub-AMC and ISG15-AMC assays were run at a final substrate concentration of 1 μM. Assays with Ub-Rh110 were run under the same reaction conditions as Ub-AMC with instrument settings adjusted to optimize detection of the fluorophore. For DGKV vOTU additional assays were run with the WT and E152D mutant using the peptide Z-RLRGG-AMC (Bachem) substrate with protease concentrations of 4 μM and a substrate concentration of 50 μM.
Assays with FRET TAMRA/QXL pair tagged K11, K48, and K63 di-Ub substrates were performed as previously described with 4 nM vOTU and 1 μM substrate [32]. Untagged poly-Ub cleavage assays were adapted from the previously published method. Briefly, 4 nM of each vOTU was tested against 10 μM Linear (M1), K6, K11, K27, K29, K33, K48, and K63 linked di-Ub (Boston Biochem, MA). Reactions were initiated by the addition of vOTU and incubated at 37°C in reaction buffer (100 mM NaCl, 5 mM HEPES [pH 7.5], 2 mM DTT). The reactions were stopped at the time points indicated by mixing 5 μl of each reaction with 2x Laemmli sample buffer and heat killed by boiling at 98°C for 5 minutes. The cleavage over time was visualized using 8–16% Mini-Protean TGX precast gels (Bio-Rad) by Coomassie staining. Assays with K48 and K63 linked tri-Ub were run in the same manner except that tri-Ub was present at 20 μM.
All four vOTUs were screened against a series of Qiagen NeXtal suites in a 96-well hanging drop format with a TTP LabTech Mosquito (TTP Labtech, Herfordshire, United Kingdom). QYBV vOTU was screened at 11.36 mg/ml, TAGV vOTU at 12.70 mg/ml, DGKV vOTU at 10.96 mg/ml, and FARV vOTU at 10.96 mg/ml. Initial hits were optimized along salt, precipitant, and pH gradients as applicable. The TAGV and FARV vOTU hits were also optimized with an Additive HT Screen from Hampton Research. Final optimized crystals for all four vOTUs were flash frozen in cryoprotective solutions. For QYBV vOTU, the final optimized crystals were in 0.3 M magnesium acetate and 16% PEG 3350, with 0.3 M magnesium acetate, 20% PEG 3350, and 18% of a 1:1:1 solution of ethylene glycol, dimethyl sulfoxide, and glycerol (EDG) as the cryoprotectant. The final crystals for TAGV vOTU were grown in 0.15 M magnesium formate, 22% PEG 3350, with 0.25 M TCEP as an additive, with a cryoprotectant solution consisting of 0.15 M magnesium formate, 22% PEG 3350, 18% EDG. Final optimized crystals for DGKV vOTU were found in the condition with 0.1 M citric acid pH 3.5, 13% PEG 6000 and were flash frozen in 0.1 M citric acid pH 3.5, 20% PEG 6000, 18% EDG. For FARV vOTU, the final optimized crystals were grown in 0.3 M magnesium chloride, 0.1 M MES pH 6.5, and 8% PEG 4000, and flash frozen in 0.3 M magnesium chloride, 0.1 M MES pH 6.5 and 20% PEG 4000 as the cryoprotectant. For selenomethionyl (Se-Met) derivative QYBV vOTU crystals, bacterial cells were grown in minimal media to OD 0.6 and induced with 0.8 mM IPTG at 37°C for 4 hrs. Prior to induction, the cultures were supplemented with eight amino acids (Leu, Ile, Val, and Trp at 0.05 g/L; Thr, Lys, Phe, and Cys at 0.1 g/L) as well as selenomethionine (0.12 g/L). Cells were harvested and protein purified as previously described. Final crystals were grown in 0.3 M magnesium acetate, 16% PEG 3350, in drops formed from 1 μl of solution and 2 ul of 9.45 mg/ml protein. Native datasets of the QYBV, DGKV, TAGV, and FARV vOTUs were collected at a wavelength of 1 Å. A Se-Met single anomalous dispersion (SAD) dataset for QYBV vOTU was collected at the absorption edge of Se at 0.9792 Å.
The data sets were indexed, integrated and scaled with HKL-2000 [62]. Experimental phasing of the Se-Met-SAD dataset was performed using the Phenix suite of programs [63]. HySS was utilized to locate the Se-Met sites, with Phaser solving the experimental phases [64–66]. Initial model building was performed using AutoBuild, with subsequent cycles of Refinement and model building carried out in Phenix and Coot ([63, 67, 68]. This structure was then used as a search model to solve the QYBV vOTU native dataset by Molecular Replacement in Phaser [66]. The other vOTUs were solved by Molecular Replacement. A QYBV vOTU-based homology model was used to solve DGKV vOTU, while homology models based on DUGV vOTU (PDB entry 4HXD) were used to solve TAGV vOTU and FARV vOTU. All the structures were built with Autobuild, followed by alternating manual building and refinement in Coot and Phenix. Structures were validated using the MolProbity server [69].
Mutations were made using the QuikChange Lightning Kit according to the manufacturer’s protocol (Agilent Technologies, Inc.). The PCR-generated plasmids were transformed into Escherichia coli NEB-5α cells by heat shock. The mutant plasmids were confirmed by sequencing and transformed into T7 Express cells (New England Biolabs).
T7 Express cells expressing Ub, Ub-L8A, and Ub-L8N in pET-15b were grown to OD 0.6–0.8 at 37°C. Expression was induced with 0.5 mM IPTG and continued at 18°C overnight. The cells were pelleted and stored as described above. The pellet was resuspended in 500 mM NaCl, 50 mM Tris [pH 7.5] supplemented with lysozyme at 4°C for 30 minutes. The cells were sonicated on ice at 70% power with a 50% duty cycle for a total of 6 minutes, followed centrifugation at 48,000xg for 45 minutes. The supernatant was filtered through a 0.8 μm and applied to a gravity flow Ni-NTA column (GoldBio) pre-equilibrated with 500 mM NaCl, 50 mM Tris [pH 7.5]. The column was washed with the same buffer containing 30 mM imidazole, followed by elution with 300 mM imidazole. Thrombin was added to cleave the 6X His-tag and the elution dialyzed overnight in 250 mM NaCl, 25 mM HEPES [pH 7.5], 2 mM DTT at 4°C. After dialysis the protein was filtered through a 0.22 μm membrane and run over a Superdex 200 column (GE Healthcare) equilibrated with 100 mM NaCl, 5 mM HEPES [pH 7.5], 2 mM DTT. The fractions were pooled based on the chromatogram and concentrated to ~2–2.5 mM, supplemented with 5% glycerol, and flash frozen in liquid nitrogen followed by storage at -80°C until further use. TAGV vOTU was purified as previously described and dialyzed alongside Ub and Ub-L8N in 150 mM NaCl, 50 mM HEPES [pH 7.5], 1 mM TCEP overnight at 4°C. ITC was performed with a Microcal PEAQ-ITC (Malvern, Worcestershire, UK). Ub or Ub-L8N were titrated into the cell in series of 19 injections, 2 μL each with a spacing of 180 seconds. The temperature was kept constant at 25°C with a reference power ranging from 6–10 μcal/s. For the TAGV vOTU binding with WT Ub the vOTU was present in the cell at 114–134 μM with Ub at 1.26–1.29 mM in the syringe. For TAGV vOTU binding with Ub-L8A the vOTU was present in the cell at 111–114 μM with Ub-L8A at 1.32–1.35 mM in the syringe. For TAGV vOTU binding with Ub-L8N the vOTU was present in the cell at 234–235 μM in the cell and Ub-L8N at 4.67–4.74 mM in the syringe. The data was processed in the Microcal PEAQ-ITC Analysis Software and fit to an independent model. Values for Ub and Ub-L8N represent the average and standard deviation of three independent runs for each experiment.
Final protein structures were deposited in the Protein Data Bank with IDs 6DWX, 6DX1, 6DX2, 6DX3, and 6DX5 for Se-Met QYBV vOTU, native QYBV vOTU, DGKV vOTU, TAGV vOTU, and FARV vOTU respectively.
|
10.1371/journal.pbio.1000148 | Repression of Flowering by the miR172 Target SMZ | A small mobile protein, encoded by the FLOWERING LOCUS T (FT) locus, plays a central role in the control of flowering. FT is regulated positively by CONSTANS (CO), the output of the photoperiod pathway, and negatively by FLC, which integrates the effects of prolonged cold exposure. Here, we reveal the mechanisms of regulation by the microRNA miR172 target SCHLAFMÜTZE (SMZ), a potent repressor of flowering. Whole-genome mapping of SMZ binding sites demonstrates not only direct regulation of FT, but also of many other flowering time regulators acting both upstream and downstream of FT, indicating an important role of miR172 and its targets in fine tuning the flowering response. A role for the miR172/SMZ module as a rheostat in flowering time is further supported by SMZ binding to several other genes encoding miR172 targets. Finally, we show that the action of SMZ is completely dependent on another floral repressor, FLM, providing the first direct connection between two important classes of flowering time regulators, AP2- and MADS-domain proteins.
| Flowering is a pivotal event in the life cycle of many plants and is therefore under tight control. The ability to detect the daily photoperiod is of particular importance in many plant species, as it enables them to enter the reproductive phase in response to seasonal changes in day length. When the photoperiod is permissive to flowering, a signal is produced in leaves that is transported to the shoot meristem, where it initiates the formation of flowers. It is now widely accepted that an important component of this long-distance signal is the flowering protein FT. Here, we show that the AP2-like transcription factor SMZ, which represses flowering and is a target of the regulatory miRNA172 microRNA, functions together with related proteins to directly regulate FT expression. Using chromatin immunoprecipitation coupled to genome tiling arrays, we find that SMZ binds directly to the FT genomic locus and to several other key flowering-related loci. Unexpectedly, the ability of SMZ to repress flowering strictly depends on the presence of the MADS-domain transcription factor FLM. In addition, SMZ binds to its own regulatory sequences and those of three closely related genes, providing evidence of strong negative feedback between SMZ and the other AP2-like miRNA172 targets.
| Throughout their lives, plants progress through distinct developmental phases, from germination and vegetative growth to flowering and, finally, senescence. The transition from vegetative growth to flowering is of particular importance because the correct timing of this switch is mandatory to ensure reproductive success. Plants have therefore evolved an elaborate genetic network that integrates endogenous and environmental signals to guarantee that flowering commences when conditions are most favorable.
Genetic and molecular analyses in Arabidopsis thaliana and other plants have identified several distinct genetic pathways that are involved in regulating the floral transition [1],[2]. On the basis of genetic interactions, one can distinguish between the gibberellic acid pathway, the autonomous pathway, and the vernalization pathway. Finally, light, and especially day length, is an important stimulus that is integrated into the flowering time regulatory network by the photoperiod pathway [3],[4]. A. thaliana is a facultative long-day plant, which means that it will flower more rapidly when day length exceeds a critical minimum. Interestingly, plants measure photoperiod in the leaves and not at the shoot apex where the new flowers will be formed. It has therefore been long postulated that the light-exposed leaves produce a flower-forming substance to regulate the formation of flowers at the shoot apex [5],[6]. This ultimately led to the formation of the “florigen” hypothesis, which postulated that a substance, “florigen,” is produced in leaves under inductive photoperiod and is transported to the shoot apex to induce flowering [7]. It was later demonstrated that such a flower-inductive substance could be transmitted from one plant (donor) via grafting to another plant (receptor) that had been cultivated under noninductive conditions.
An important factor that allows Arabidopsis to discriminate between short day (SD) and inductive long day (LD) is the B-box–type zinc finger protein CONSTANS (CO) [8]. The regulation of CO at both the mRNA and protein levels ensures that the protein will accumulate and activate flowering only under LD conditions [4],[9],[10]. Interestingly, CO appears to carry out its function in leaves, where it acts in the phloem companion cells to regulate a systemic signal that induces photoperiodic flowering [11],[12].
Several lines of evidence suggest that the protein FLOWERING LOCUS T (FT) acts as a florigen to convey flowering time signals from the leaves to the apex [13],[14]. First, it was established that FT is the major target of CO in leaves [15],[16]. It was further demonstrated that the FT protein interacts at the shoot apex with another flowering time regulator, the bZIP transcription factor FD, to induce downstream flower-specific targets such as the MADS-domain proteins APETALA1 (AP1) and FRUITFULL (FUL) [16],[17]. The finding that FT is transcribed in leaves but acts at the apex implied that FT can move, either as mRNA or as protein. Later experiments were unable to detect FT mRNA movement but provided evidence that FT protein is able to reach the apex when expressed in the vasculature [18]–[23].
Interestingly, the induction of flowering under LD by CO/FT is counteracted by several factors that either prevent FT expression in the leaf or act downstream of FT to modulate its function at the shoot apex. In particular, MADS-domain transcription factors have been shown to act as repressors of flowering. The most prominent of these is FLOWERING LOCUS C (FLC), which represses flowering in winter annual accessions of Arabidopsis before the plants have been exposed to a prolonged period of cold [24]. It has recently been shown that FLC, when expressed either from the phloem-specific SUC2 promoter or the meristem-specific KNAT1 promoter, efficiently represses flowering and that these effects are additive. Further, it was demonstrated that FLC directly binds to the regulatory regions of three positive regulators of flowering, FT, FD, and SUPRESSOR OF OVEREXPRESSION OF CONSTANS 1 (SOC1), presumably to repress these genes [25]. Two other MADS-domain transcription factors, FLM and SVP, have also been shown to repress flowering. In contrast to FLC, which is involved in the vernalization pathway, these two genes seem to be involved predominately in the photoperiod pathway, and FLM and SVP act as partners [26]–[28]. There is, however, also evidence that implicates SVP and FLM in temperature-dependent regulation of flowering in Arabidopsis, and SVP has recently been shown to interact with FLC in a repressor complex [29]. In addition, SVP has also been shown to directly bind to regulatory regions of FT and SOC1 [29],[30].
More recently, two more transcription factors, TEMPRANILLO 1 (TEM1) and TEM2, have been shown to redundantly repress flowering [31]. In contrast to FLC, FLM and SVP, which are MADS-domain transcription factors, each TEM gene encodes an AP2 domain as well as a B3-type DNA binding domain. TEM1 is most strongly expressed in leaves, where its expression is regulated in a circadian fashion [31]. TEM1 was further shown to directly bind to the 5′ UTR of FT [31]. This is in contrast to FLC, which bound most strongly to the first intron of FT, indicating that FT is regulated by different repressors in different regions.
Yet another family of six AP2-like transcription factors also act as repressors of flowering. This clade of proteins comprises APETALA 2 (AP2) itself, the three TARGET OF EAT (TOE) proteins (TOE1, TOE2, and TOE3), and SCHLAFMÜTZE (SMZ) and its paralog SCHNARCHZAPFEN (SNZ) [32]–[34]. All six genes have in common that they are predicted targets of microRNA172 (miR172), expression of which is regulated by GIGANTEA (GI) to control flowering in a CO-independent manner [35]. It has previously been shown that TOE1 and TOE2 act as repressors of flowering: toe1 mutants are significantly early flowering, and this effect is enhanced in a toe1 toe2 double mutant [33],[35]. However, plants that expressed miR172 constitutively were found to flower much earlier than even the toe1 toe2 double mutant, indicating that the other AP2 family members most likely act redundantly with TOE1 and TOE2 to repress flowering [33],[36]. A good candidate for such a repressor is SMZ, which was originally identified in an activation-tagging screen because of its dominant late-flowering phenotype [34]. Additionally, SNZ, a paralog of SMZ, has been shown to repress flowering when expressed at high levels [34]. However, it was unclear whether SMZ and SNZ normally act as repressors of flowering.
Here, we show that the miR172 targets SMZ and SNZ are bona fide floral repressors and act redundantly with TOE1 and TOE2 to delay flowering specifically under LD conditions. Plants expressing SMZ at high levels are late flowering, which is due to an almost complete block in FT induction. The effects of SMZ on FT expression appear to be direct, as chromatin immunoprecipitation coupled to hybridization to tiling arrays (ChIP-chip) identified FT as a target of SMZ. In addition, several other known regulators of flowering time were identified as SMZ targets. Among them are SMZ itself, SNZ, AP2, and TOE3, suggesting a complex feedback regulation among miR172 targets. Finally, we found that repression of flowering by SMZ is independent of the potent floral repressor FLC, but requires FLM for its function, providing a direct connection between two important classes of flowering time regulators, AP2- and MADS-domain proteins.
smz-D was originally isolated as a dominant late-flowering mutant in an activation-tagging screen under LD conditions. SMZ is expressed in young seedlings and is developmentally regulated, as deduced from microarray data (Figure 1A) and confirmed by a genomic SMZ∶GUS reporter (Figure 1B–1D). Expression of SMZ declines with increasing age, but SMZ is induced again in seeds during maturation. In addition, analysis of publicly available microarray data (“The diurnal project”; http://diurnal.cgrb.oregonstate.edu/) revealed that SMZ exhibits a diurnal expression with a maximum at Zeitgeber 15 under LD conditions [37].
To better understand where SMZ functions in respect to the known flowering time pathways, we first investigated the flowering time behavior in this mutant under different day lengths (Figure 2A and Table 1). We found that smz-D delays the onset of flowering specifically under inductive LD conditions, where it produced 45.1±1.7 leaves before flowering compared to wild-type (15.5±0.6 leaves). In contrast, under noninductive SD conditions, smz-D (79.1±2.2 leaves) plants flowered similarly to wild-type (73.0±1.6 leaves), indicating that smz-D represses flowering specifically under inductive LD, but has little effect under SD conditions.
To investigate whether SMZ is indeed functioning as a floral repressor, we isolated homozygous SMZ loss-of-function alleles from T-DNA insertion collections. Neither of the individual smz mutant lines displayed any obvious phenotypes; in particular, the total number of leaves did not significantly differ from that of wild-type plants (Figure 2B and Table 1). Most notably, the plants were not early flowering, as one would have expected to result from the loss of a putative floral repressor. Also, a double mutant lacking SMZ and its closest paralog SNZ was found to be indistinguishable from the wild type (Figure 2B).
Together with AP2, TOE1, TOE2, and TOE3, SMZ and SNZ form a clade of six AP2-domain transcription genes. Because functional redundancy has been observed within this clade in respect to the timing of floral induction of the toe1 toe2 double mutant, we first focused on the function of TOE1, TOE2, SMZ and SNZ rather than that of AP2 and TOE3, which are predominately expressed at the meristem, and created a mutant line that lacks toe1 toe2 smz snz functions. This quadruple mutant was found to flower significantly earlier than Col-0, toe1, and even toe1 toe2 double-mutant plants (Figure 2B and Table 1; p<0.001 in all comparisons). This result confirms that SMZ and its paralog SNZ are indeed acting as floral repressors redundantly with TOE1 and TOE2. This effect is only apparent in the sensitized toe1 toe2 mutant background, as TOE1 normally masks the effects of smz and snz loss of function.
The early flowering we observed in certain combinations of smz, snz, toe1, and toe2 loss-of-function alleles was associated with a reduced number of rosette leaves, whereas the number of cauline leaves remained constant (Figure S1). It is interesting to note that even the toe1 toe2 smz snz quadruple mutant still flowers significantly later than plants that constitutively express miR172, which have been reported to produce on average two to three rosette leaves before bolting [33],[36]. This strongly suggests that the two remaining miR172 targets, AP2 and TOE3, also act to repress flowering, which is especially interesting given that these two genes are predominately expressed at the meristem.
We conclude from these results that SMZ and its homolog SNZ are bona fide floral repressors that, partly redundant with other members of the miR172 target family, act to delay flowering in A. thaliana under LD conditions.
Two genes play key roles in the photoperiod pathway: CO, which constitutes the main readout of the circadian clock, and FT, which has been shown to be an important part of the mobile signal that conveys the information to induce flowering from the leaves to the apex [18]–[23].
To test the genetic position of SMZ in relation to these two factors, we introduced smz-D into established plant lines that expressed CO or FT under the control of the constitutive 35S promoter (Figure 2C). Lines expressing either of these two genes at a high level are extremely early flowering. We observed a substantial delay in flowering in smz-D 35S::CO plants (11.3±0.4 leaves) compared to the CO overexpressing line (5.4±0.3 leaves) (Figure 2C and Table 1). In contrast, smz-D had a much smaller effect on the flowering of plants expressing FT at high levels (6.3±0.3 leaves; compared to 4.9±0.2 leaves observed in 35S::FT). These findings are compatible with the idea that SMZ acts as a repressor of flowering and counteracts the flower-promoting activity of CO.
Next, we tested the dependence of SMZ on the presence of functional FLC, as FLC is a well-described repressor of flowering, integrating environmental signals such as vernalization and ambient temperature [24].
FLC has been shown to directly bind to regulatory sequences of the FT gene as well as to the promoters of SUPRESSOR OF OVEREXPRESSION OF CONSTANS 1 (SOC1 = AGL20) and FD [29]. We therefore tested whether smz-D acts through FLC to repress flowering. When we introduced smz-D into the strong flc-3 deletion mutant background, which lacks part of the 5′UTR and the first exon, and is a genetic null allele of FLC, we observed no difference in flowering time in smzD flc-3 (43.8±1.9 leaves) when compared to smz-D (45.6±1.9 leaves), showing that smz-D represses flowering independently of FLC (Figure 2D and Table 1).
As mentioned above, SMZ and SNZ share a miR172 target site. To test whether the mRNAs of these two genes are indeed targeted for degradation and are cleaved at the predicted positions, we carried out RACE-PCR to map the 5′ end of miR172 cleavage products. We found that the SMZ mRNA was cleaved at the predicted site in all clones analyzed (n = 12; Figure S2). Similarly, correct cleavage of SNZ mRNA was observed in 12/13 (Figure S2) cases, confirming that both SMZ and SNZ are indeed miR172 targets.
To further investigate whether the regulation of SMZ by miRNA172 plays a role in controlling the floral transition, we introduced into plants a version of SMZ mRNA (rSMZ) that carried silent mutations in the miR172 complementary site, rendering the mRNA resistant to miR172-directed cleavage (Figure S2). Strong expression of rSMZ from the constitutive 35S promoter caused plants to remain vegetative throughout their life (Figure 3C and 3G, and Table 2). In addition, the leaves of these plants displayed a crinkled phenotype and remained smaller than those of either wild-type controls or plants transformed with the SMZ ORF. The failure of 35S::rSMZ plants to initiate flowering suggests that miR172-directed cleavage of SMZ mRNA in smz-D and 35S::SMZ plants limits the effects of overexpressing the native version of SMZ.
Besides FLC, two other MADS-domain proteins, SHORT VEGETATIVE PHASE (SVP) and FLOWERING LOCUS M (FLM = MAF1), have been shown to function as floral repressors [26],[27],[30]. On the basis of genetic analysis of mutant alleles, it has been suggested that FLM and SVP act as coregulated partners in the same pathway [28].
To test whether either of these two genes is required for SMZ function, 35S::SMZ and 35S::rSMZ constructs were transformed into established svp and flm T-DNA mutant lines. Loss of SVP did not affect the SMZ overexpression phenotype, i.e., svp plants carrying either the 35S::SMZ or the 35S::rSMZ transgene flowered just as late as control plants (Figure 3G and Table 2).
In contrast, the late flowering, which usually would result from SMZ overexpression, was completely abolished in the flm mutant background (Figure 3E and 3G, and Table 2). Even constitutive expression of the miR172-resistant form of SMZ (rSMZ) was no longer able to delay flowering (Figure 3F and 3G, and Table 2) in flm mutants. This is in contrast to the extreme effect of (r)SMZ on flowering in wild-type control transformations (Figure 3B, 3C, and 3G, and Table 2). Interestingly, we also observed that expression of rSMZ in flm resulted in reduced growth and crinkly leaves similar to what we had observed in wild-type control plants transformed with 35S::rSMZ (Figure 3C), despite the fact that the plants now flowered with a normal number of leaves. In addition, apical dominance was reduced in these lines, giving the plants a bushy appearance. The fact that expression of rSMZ causes phenotypes even though no flowering time defects were observed rules out that the transgene was silenced in the flm background. In addition, levels of transgene expression were found to be comparable in Col-0 and flm background when analyzed by quantitative reverse transcription PCR (qRT-PCR) (unpublished data).
Because high levels of FLM have been shown to delay flowering, we therefore examined the expression of FLM in smz-D by qRT-PCR, but did not find any evidence that FLM levels were increased (unpublished data). Therefore, it does not appear that SMZ is simply up-regulating FLM transcription.
Our previous results demonstrated that SMZ acts as a floral repressor in the photoperiod pathway. A characteristic of this pathway is the spatial separation of the perception of inductive photoperiod in the leaves and the formation of flowers at the shoot apical meristem. Recently, the FT protein has been shown to be an important component of the signal that transmits flowering time information from the leaves to the apex [18]–[23]. In the course of these studies, it was shown that for FT to exert its function in the photoperiod pathway, it is both necessary and sufficient for this gene to be expressed in leaf phloem companion cells [19],[20],[22]. Therefore, if the function of SMZ in flowering time is indeed to negatively regulate FT expression, misexpression of SMZ in phloem companion cells and the resulting repression of FT in this tissue should be sufficient to recapitulate the late-flowering phenotype observed in smz-D.
As expected, expression of SMZ from the phloem companion cell-specific SUC2 promoter efficiently delayed flowering and resulted in late-flowering plants that were phenotypically indistinguishable from 35S::SMZ plants (Figure 4A and 4B, and Table 2). Similarly, rSMZ driven from the phloem-specific promoter SUC2 did not cause plants to flower later than the miR172-susceptible SMZ ORF driven from the same promoter (Figure 4A and 4B, and Table 2). Furthermore, SUC2::rSMZ plants did not display any of the additional defects in leaf or shoot morphology that were evident in 35S::rSMZ (Figure 4A), indicating that these phenotypes were caused by misexpression of rSMZ in tissues other than the vasculature. An alternative explanation could be that miR172 is normally not expressed in phloem companion cells, in which case, SMZ and rSMZ overexpression would have similar effects. This seems unlikely, however, as miR172 has been cloned from phloem exudates in Brassica [38].
In contrast, expression of SMZ from the shoot meristem-specific FD promoter had only the most modest effect on bolting time (Figure 4A and 4B, and Table 2). Even in FD::rSMZ plants, the number of rosette leaves (14.2±1.5) was similar to that in controls, indicating that high levels of SMZ at the shoot apex are not sufficient to delay the onset of flowering. The number of cauline leaves, however, was vastly increased in FD::rSMZ plants (13.8±6.3) compared to wild type (2.9±0.2) (Figure 4B and Table 2). Additionally, these plants displayed a shoot phenotype reminiscent of a double mutant lacking both the meristem identity gene LEAFY (LFY) and AP1, in that flowers were replaced by leaf-like organs, which were frequently subtended by bracts (Figure 4A). Taken together, these results suggest that SMZ can affect different sets of target genes in leaves and at the shoot meristem. An alternative explanation would be that the FD promoter becomes active too late in development to delay flowering but in time to repress flower development, causing this shoot phenotype.
As we showed earlier, genetic analyses clearly place SMZ in the photoperiod pathway and tissue-specific misexpression of SMZ suggests that regulation of flowering time by SMZ occurs predominantly in the leaves. To analyze the molecular cause for the late flowering of SMZ overexpressing plants, we carried out quantitative RT-PCR on the putative target gene FT, which is normally induced in leaves under LD.
As expected, FT mRNA was not detectable in noninductive SD conditions in flc-3 mutants, which served as a background for this experiment, irrespective of the presence or absence of smz-D (Figure 5A). FT transcription was rapidly and strongly induced in flc-3 1 d after plants were shifted to inductive LD conditions (Figure 5A). Levels of FT mRNA increased even further after exposure to four consecutive LD. In contrast, smz-D flc-3 plants completely failed to induce FT, even after 4 LD. These results indicate that the late flowering observed in smz-D is largely caused by the inability of smz-D plants to induce FT even under inductive LD. Furthermore, the ability of SMZ to repress FT did not depend on the presence of a functional FLC allele, as already suggested by genetic analyses (Figure 2D).
Genetic analyses had demonstrated that, in contrast to FLC, FLM is strictly required for SMZ to repress flowering. We therefore tested FT expression in a flm mutant and compared it to that in a flm 35S::SMZ line (Figure 5B). As expected, FT was readily induced in flm after 3 d of inductive LD. FT levels were actually higher in flm than in Col-0 wild-type control plants, suggesting that FLM is normally involved in FT repression. In 35S::SMZ plants, however, FT induction was strongly attenuated, and FT levels reached only 8% of those observed in Col-0. In contrast, FT was strongly expressed in a 35S::SMZ flm line, indicating that SMZ requires functional FLM in order to suppress FT induction. This is in perfect agreement with our genetic analyses, which had shown that a mutation in FLM completely suppresses the late flowering of SMZ overexpression.
To test whether SMZ normally represses FT, we analyzed its expression in the toe1 toe2 smz snz quadruple mutant. We found that FT is expressed at high levels in this mutant background when compared to wild-type control plants throughout the first 2 wk of development (Figure 5C). This supports the idea that SMZ, together with the other AP2 family members, represses flowering by regulating FT expression.
To determine the effects of SMZ overexpression on the transcriptome in greater detail, we performed a microarray analysis in leaves and at the shoot meristem of flc-3 and smz-D flc-3 plants. SMZ was significantly (RankProducts, percentage false positives [pfp]<0.01) up-regulated in smz-D plants at all time points in both tissues investigated (Figure 6A). In contrast, expression of GIGANTEA (GI) and CO, which both act upstream of FT in the photoperiod pathway, remained unchanged in smz-D plants (Figure 6A). Furthermore, the diurnal expression normally observed of CO and GI was unaltered (Figure S3), indicating that SMZ is not regulating flowering by modulating expression of these genes. Analysis of FT expression by qRT-PCR had revealed that FT expression is strongly attenuated in smz-D (Figure 5A). In agreement with this, we found that FT was significantly (pfp<0.01) induced in flc-3 leaves 1 and 3 d after plants were transferred to inductive LD but that FT induction in leaves was completely blocked by smz-D (Figure 6B). FT mRNA was not detectable at the shoot meristem at any time point in all samples.
Interestingly, TWIN SISTER OF FT (TSF) followed the expression of FT in that it was substantially up-regulated in leaves of flc-3 plants, but was not induced at any other time point in any tissues investigated (Figure 6B). This supports the idea that FT and TSF act partially redundantly in promoting the floral transition. Statistical analysis revealed that there was only one other transcript besides FT that was significantly (pfp<0.01) induced in leaves of flc-3 plants in response to inductive photoperiod 1 and 3 d after shift to LD that was not also up-regulated in smz-D. This gene encodes a β-amylase (BMY1; At4g15210) and has not previously been implicated in the regulation of flowering.
Taken together, these results indicate that FT, and to a lesser extent TSF, are major targets of SMZ in leaves. Such a repression of FT could either be due to a direct effect of SMZ on the FT locus or through the activation of other floral repressors. To investigate the possibility that SMZ is acting indirectly on FT by activating transcription of another floral repressor, we examined our smz-D microarray data. Many of the known repressors of flowering encode MADS-domain transcription factors, but neither FLC, FLM, nor SVP RNA was induced in smz-D flc-3, when compared to flc-3 (Figure 6C). This is of particular importance, as both FLC and SVP have been shown to bind to the FT locus. These results indicate that the delay in flowering observed in smz-D plants is not simply caused by FLC and/or SVP activation. It is, however, possible that SMZ activates or stabilizes FLC and/or SVP protein by an unknown mechanism.
Recently, two AP2-domain transcription factors, TEM1 and TEM2, have also been shown to regulate FT. In fact, direct binding of TEM1 to a regulatory region of FT has been demonstrated [31]. Although TEM1 expression is unaltered in both the leaves and at the shoot meristem of smz-D plants, TEM2 was found to be significantly (pfp<0.01) up-regulated in leaves and shoot meristem samples of smz-D at all time points (Figure 6C). This suggests that at least part of the effect of SMZ on flowering may be mediated by TEM2.
Interestingly, expression of SNZ, the closest paralog of SMZ, is significantly (pfp<0.05) down-regulated in leaves of smz-D flc-3 1 and 4 d after plants were shifted to LD (Figure 6C). Indeed, even before the shift (day 0), a substantial repression of SNZ can be observed. Similarly, levels of AP2, TOE1, and TOE3 mRNA are reduced in SMZ overexpressing lines, although to a lesser extent, suggesting widespread feedback regulation among the miR172 target genes (Figure S4).
Positive regulators of flowering such as FUL, LFY, CAL, and SOC1 were all induced significantly (pfp<0.05) at the apex of flc-3 plants after the shift to LD (Figure 6D). Similarly, FD, which physically and genetically interacts with FT, was substantially (but not significantly) induced at the meristem (Figure 6D). Neither of these genes was, however, induced in smz-D flc-3, indicating that high levels of SMZ are sufficient to completely block the transition to flowering at the shoot apex. In agreement with this finding, homeotic genes such as AP1, PI, and AG are also not induced in smz-D flc-3, but are readily detectable at the meristem of flc-3 (Figure 6E).
To determine whether SMZ acts as a regulator of transcription, we established lines that express SMZ in fusion with an N-terminal GFP tag and drove them in the leaves by the 35S promoter. To sustain SMZ function, it was necessary to place a flexible linker consisting of ten Gly-Ser pairs between the GFP and SMZ. Among the T1 lines, several late-flowering individuals were recovered. Late flowering in these lines was confirmed in the T2 generation (Figure 7), indicating that functional GFP∶SMZ protein persisted in these lines at high levels. As expected for a putative transcription factor, the GFP signal was predominately nuclear localized in GFP∶SMZ plants even though signal intensity was rather low when compared to control plants expressing a nuclear-localized 3xVENUS YFP (Figure S5).
The nuclear localization of the GFP fusion protein enabled us to perform chromatin immunoprecipitation on whole-genome tiling arrays (ChIP-chip) in order to identify regions in the Arabidopsis genome bound by SMZ. In total, 434 regions in the genome exhibited statistically significant enrichment to GFP∶SMZ at a false discovery rate (FDR) of 5% or less when compared to a line expressing a nuclear-localized YFP (Dataset S5). Of these 434 peak regions, only 33 were not directly associated with genes (±2.5 kb of the coding sequence [CDS]), whereas the great majority (401 peaks) fell within 2.5 kb of annotated genes. The latter were associated with 395 unique Arabidopsis gene models with six annotated loci having two peaks of significant binding.
We observed several interesting enrichments for particular gene ontology (GO) categories among the 307 of 395 genes for which assignments exist. Indeed, genes associated with flower development (GO:0009908) were significantly overrepresented at a FDR p<0.0005 among the list of potential SMZ target genes, demonstrating a functional specificity to binding and a nonrandom distribution of the peaks identified by ChIP-chip. The second biological process found to be overrepresented at a FDR p<0.0005 comprises genes involved in “response to stimuli” (GO:0050896), in particular to water (GO:0009415, GO:0009414) and jasmonic acid (GO:0009753).
Among the genes bound by SMZ were many known regulators of flowering, suggesting that the effects of SMZ on flowering time are rather direct. Most notably, the second most strongly enriched locus in the entire genome analysis was located less than 2 kb upstream of the transcription start site of the miR172 target gene TOE3 (Figure 8C). Binding at the TOE3 locus was highly statistically significant (FDR<0.000001). Closer inspection of the list of high-confidence SMZ targets (FDR<5%) revealed that three other miR172 targets were also significantly bound. Among them were SMZ itself, its paralog SNZ, and AP2 (Figure 8A, 8B, and 8D). In fact, SMZ was one of only six loci genome-wide that was the closest locus to two peaks of high-confidence binding. No high-confidence binding by SMZ was detected to the last two members of the clade of miR172 targets, TOE1 and TOE2. The finding that four out of six miR172 targets were significantly bound by SMZ is more than expected by chance (Fisher exact test; p<0.0001). It should be noted that the expression levels of SNZ, AP2, and TOE3 were reduced in leaves of smz-D plants as measured by microarrays (Figure 6C and Figure S4). Taken together, these results strongly suggest a complex negative regulatory feedback mechanism among the miR172 targets. In addition, two other known repressors of flowering, TEM1 and FRI, were also bound by SMZ (Figure 8E and 8F).
Most interestingly, FT was also among the genes that showed significant binding by SMZ approximately 1.5 kb downstream of the FT CDS (Figure 8G). Assuming that SMZ acts as a transcriptional repressor, the binding of SMZ to the FT locus readily explains the failure of smz-D and 35S::SMZ plants to induce FT (Figure 5A and 5B) and the high levels of FT expression in the toe1 toe2 smz snz quadruple mutant (Figure 5C). The regulatory landscape around the FT locus appears to be rather complex, with TEM1 binding to the 5′UTR, FLC binding to the first intron, and finally, SMZ binding downstream of the coding region [25],[29],[31]. Further, the effect of high SMZ levels on flower development is most likely not only due to a repression of FT, but also to SMZ repressing other flowering time regulators as well. This idea is supported by the finding that, besides FT, the floral integrator and flower development genes SOC1 and AP1 were also identified by ChIP-chip as high-confidence (FDR<5%) targets of SMZ (Figure 8H and 8I). In each of these cases, binding occurred directly upstream of the transcription start site, suggesting strongly that expression of these genes is under direct negative regulation by SMZ. Similar to what we had observed for FT, smz-D plants did not induce SOC1 or AP1 either in leaves or at the shoot meristem when shifted from SD to inductive LD (Figure 6D and 6E). Enrichment of loci identified by ChIP-chip was confirmed by quantitative PCR for all genes discussed (Figure 8J).
To test whether SMZ and its paralogs affect expression of the genes bound by SMZ, we analyzed their expression by quantitative PCR in the toe1 toe2 smz snz quadruple mutant. As already described for FT (Figure 5C), AP1 and SOC1 were strongly induced in the quadruple mutant (Figure S6). So were LFY and FUL (Figure S6), which are not or only weakly bound by SMZ. The latter most likely is due to indirect activation of these genes. The expression of AP2 and TOE3 was also increased. This is in agreement with the proposed negative feedback regulation among the miR172 target genes. In contrast, FRI was only marginally up-regulated (Figure S6). However, the FRI allele of the Col-0 accession, is recessive and essentially nonfunctional, making it hard to interpret this result [39].
Our results strongly support the notion that SMZ, and by extension the other miR172 targets as well, act as direct repressors of the transition to flowering. Consistent evidence from both gene expression and ChIP-chip experiments suggests that SMZ directly represses the transcription of a range of known flowering time genes and that the delay in flowering time caused by high levels of SMZ is most likely a result of repression of a number of flowering time regulators in both the leaves and the shoot meristem.
To ensure reproductive success, plants have evolved a complex regulatory network that integrates various endogenous and environmental factors to ensure that flowering occurs when conditions are most favorable. Many of the key regulators that control flowering time have been identified and the majority of them are putative transcription factors. Extensive epigenetic regulation of several key regulators of flowering complicates the situation even further [40]. Based on genetic analyses, pathways that control the transition to flowering have been defined, but the details of how this transcription factor network functions at a molecular level is poorly understood. Here, we have combined genetic analysis with high-throughput microarray technologies to understand in detail how the AP2-like transcription factor SMZ represses flowering. A model summarizing our findings regarding the genetic interactions of SMZ and its position in the network regulating flowering in response to photoperiod is represented in Figure 9.
SMZ is predominately expressed in young leaves, suggesting that this is the tissue where it normally functions [35],[41]. Leaves play a crucial role in the perception of day length, and it has recently been demonstrated that the information to induce flowering can be conveyed from the leaves to the shoot apex via transport of the FT protein [18]–[23]. The regulation of FT expression is therefore of the utmost importance for a plant to ensure the correct timing of flowering. Plants achieve this control by a combined effect of activators of FT expression, such as CO, and repressors, such as FLC, FLM, SVP, and the TEM proteins, some of which have been shown to directly bind to regulatory regions of the FT locus [25],[29],[31]. Expression of SMZ from a leaf-specific promoter recapitulated the late-flowering phenotype of constitutive SMZ overexpression, indicating that presence of SMZ in the vasculature was sufficient to repress flowering. Molecular analyses indicate that SMZ directly contributes to the regulation of FT in leaves. The evidence for this is 2-fold: first, plants expressing SMZ at high levels fail to induce FT in response to LD, and second, SMZ binds directly to the FT locus. Taken together, these results strongly indicate that SMZ acts as a floral repressor and that FT is a major transcriptional target of SMZ in leaves.
Whether FT constitutes the sole mobile signal that conveys the instruction to flower from leaves to the apex is still an open question. Several other classes of molecules have been implicated as long-distance flowering signals in various plant species [42],[43]. Carbohydrates in general, and sucrose in particular, have been suggested to play a role in the induction of flowering in Arabidopsis [44],[45]. The mechanism by which sugars affect flowering is not entirely clear, but the finding that SMZ binds to and represses BMY1, which encodes a cytosolic β-amylase, may provide new insights into this issue.
Misexpression of miR172-resistant SMZ from a meristem-specific promoter had a marked effect on flower development, and FD::rSMZ plants phenotypically resembled ft lfy, fd lfy, or lfy ap1 double mutants [16],[17],[46]. Although we did not detect high-confidence binding of SMZ to LFY, strong, significant binding was observed to SOC1 and AP1, so it is possible that the reduced abundance of these factors at the meristem may be at least partly responsible for the observed phenotype. Along these lines, it has recently been reported that a soc1 ful double mutant reverts to a vegetative state after flowering had been induced, resulting in a perennial growth habit of the double mutant and demonstrating the importance of these genes in robustly inducing flowering [47].
Furthermore, SMZ does not act alone in repressing flowering but instead redundantly with related proteins, all of which are miR172 target genes. It has previously been shown that miR172 overexpression did not dramatically alter the mRNA levels of its targets. This has been interpreted as evidence for translational repression being more important than mRNA cleavage [36]. Later, however, it was shown that at least one of these genes, AP2, can repress its own transcription, demonstrating that a negative feedback very much confounds this conclusion [48]. It was, however, not clear whether this repression was direct or indirect. Also, it was unclear just how widespread this negative feedback regulation among the miR172 targets really was. Our genome-wide ChIP binding and gene expression studies indicate that SMZ is not only binding to its own genomic region, but regulates at least three other family members as well, demonstrating that the negative feedback is direct and common among the miR172 targets.
The function of SMZ appears to strictly depend on the presence of the MADS-domain transcription factor FLM. The mechanistic details of this interaction remain unknown, but one can imagine several possible scenarios. SMZ might directly interact with FLM protein to form a repressor complex. However, at least when tested in yeast, we did not find any indication for direct interaction between SMZ and FLM (unpublished data). An alternative would be that FLM needs to be present at target loci in order to facilitate either SMZ binding or activity, but without physical interaction between the two. The dependence of SMZ on FLM seems to be rather specific, as inactivation of the MADS repressors FLC nor SVP, both of which have been shown to directly repress FT, does not prevent SMZ function. It has been suggested that FLM and SVP genetically act as partners in repressing flowering time [28]. However, at least in the case of repressing SMZ activity, FLM and SVP functions are clearly separate and not interchangeable.
In addition to the genes discussed so far, TEM1, a known repressor of flowering, was also bound by SMZ (Figure 8E). How binding of TEM1 by SMZ could possibly regulate flowering is currently unclear. It should be noted that the expression levels of TEM1 were not changed in leaves and at the meristem by smz-D (Figure 6C). In contrast, the expression levels of TEM2, the closest paralog of TEM1, were significantly up-regulated in leaves and meristem samples (Figure 6C). One may hypothesize that, similar to what we have observed among the miR172 targets, control of TEM1 and TEM2 expression also involves a regulatory feedback mechanism.
Regardless of the precise mechanisms that control TEM expression, the increased TEM2 levels in smz-D very likely contribute to the repression of FT. Finally, FRIGIDA (FRI) was also identified among the genes most strongly bound by SMZ (Figure 8F). FRI is a potent activator of FLC, and together these two genes are to a large extent responsible for the winter-annual behavior of certain Arabidopsis accessions [24],[39],[49]. Col-0 carries a recessive FRI allele, and it is therefore unlikely that the binding of SMZ to the FRI promoter is responsible for the delay in flowering we observe in smz-D. However, SMZ and the other miR172 targets could very well contribute to the control of FRI expression in late-flowering accessions that carry a functional FRI allele. This would provide Arabidopsis with a way to regulate FLC levels by modulating FRI expression. To test such a scenario, one would need to analyze the effect of gain- and loss of function of miR172 targets on flowering time and especially FLC expression levels in a FRI dominant background.
Our results indicate that the miR172/SMZ module functions as a rheostat in flowering time by SMZ binding to several genes encoding miR172 targets and other flowering time regulators. The importance of this regulatory module is highlighted by the finding that overexpression of miR172 strongly accelerates flowering [33],[36], whereas constitutive expression of SMZ (or AP2, TOE1, or TOE2) has the opposite effect [33],[34],[36]. In nature, to tightly control flowering time, Arabidopsis must achieve a careful balance between miR172 and its targets. Negative feedback of SMZ onto the other miR172 targets likely contributes to this regulation. miR172 and its targets are not specific to Arabidopsis, but are conserved in other dicotyledonous as well as monocotyledonous plant species, suggesting that these genes play an important role in plant development in general [50]. In maize, for example, miR172 promotes vegetative phase change and onset of reproductive development [51], indicating that the function of the miR172/AP2 module is largely conserved. In addition, miR172 and its target indeterminate spikelet1 (ids1) have been shown to participate in sex determination and meristem cell fate in maize [52]. Thus, our findings about the regulatory module consisting of AP2-like transcription factors and their microRNA will likely be relevant to many other plants.
In summary, we provide evidence for a complex regulatory feedback mechanism among the miR172 target genes that directly controls the expression of FT. In addition, we show that several other known flowering time regulators such as SOC1 and AP1 are also directly targeted and repressed by SMZ. The intricate regulatory interactions we uncovered by just looking at just one single factor, SMZ, demonstrate how complex regulation of flowering time at the molecular level is. To fully understand this complex trait, a concerted effort of the flowering time community will be required to systematically study the genes and proteins involved in floral transition on a genome-wide scale.
Sequences of oligonucleotide primers used in this work are given in Table S2.
Wild-type plants were of the Columbia (Col-0) accession. All T-DNA insertion mutants used in this work are in Col-0 accession [53],[54]. flc-3, flm-3, svp-31, toe1-2, and toe2-1 have been described before [24],[30],[33],[55]. Two T-DNA insertion lines for SMZ (smz-1 and smz-2) and one SNZ loss-of-function allele (snz-1) were isolated as part of this work (Table S1). smz-2 was used for genetic analysis throughout this work. Mutant plants were confirmed by PCR-based genotyping.
All plants were grown in growth chambers in a controlled environment (23°C, 65% relative humidity). Plants were raised on soil under a mixture of Cool White and Gro-Lux Wide Spectrum fluorescent lights, with a fluence rate of 125 to 175 µmol m−2 s−1. All light bulbs were of the same age. Long day (LD) is defined as 16 h light, 8 h dark, and short days (SD) as 8 h light, 16 h dark. For flowering time measurements, plants were randomized with the respective controls, and the flowering time phenotype was determined without prior knowledge of the genotype. All flowering time assays were performed at least twice.
ORFs were amplified using the Pfu DNA polymerase (New England Biolabs), cloned into Gateway entry vectors using T4 DNA ligase and subsequently recombined into Gateway-compatible binary vectors suitable for plant transformation. Constructs for constitutive and tissue-specific expression of SMZ were obtained by amplification of the SMZ ORF using oligonucleotide primers G-3323 and G-5638, cloning the PCR product into pJLSmart, resulting in pJM9, and recombination into Gateway-compatible plant binary vectors, providing promoters for expression in plants, generating pJM34 (35S::SMZ), pJM66 (SUC2::SMZ), and pJM50 (FD::SMZ). To generate the miRNA172-resistant form of SMZ (rSMZ), synonymous mutations were introduced into the miR172 binding site by site-directed mutagenesis using oligonucleotide primers G-2050 and G-2051, resulting in pFK37. Cloning of rSMZ into Gateway-compatible entry and destination vectors was as described above, resulting in pJM36, pJM68, and pJM52 (for 35S::rSMZ, SUC2::rSMZ, and FD::rSMZ, respectively). To separate SMZ from the GFP tag, a Gly-Ser linker was added to the N-terminus of the SMZ ORF in a two-step PCR. In a first PCR, the SMZ ORF was amplified using primers G-16615, which replaced the start codon of SMZ with 30 bases encoding for five Gly-Ser pairs, and G-16616. In a second step, the Gly-Ser linker was extended to its final length of 60 bases, encoding for ten Gly-Ser pairs, using G-18665 and G-16616. The resulting PCR product was cloned into the SmaI site of the pJLSmart Gateway-compatible entry vector by blunt end ligation (pFK478). The GS10∶SMZ ORF was subsequently recombined into a pGREEN-IIS based Gateway-compatible destination vector (pFK247), which provided 35S promoter for expression in plants and an in-frame fusion with an N-terminal eGFP, resulting in pFK480. For the genomic SMZ∶GUS reporter, a 10.6-kb EcoRV-SacI, including the whole SMZ 5′ and 3′ regions, was cut from BAC T15C9 and cloned into the pGREEN-IIS plant binary vector. The GUS ORF, including stop codon, was amplified by PCR using primers G-8237 and G-8238, introducing AgeI sites in the process. The GUS ORF was cloned in frame with the SMZ start codon in an AgeI site present in the first exon of SMZ. All sequences amplified by PCR were confirmed by sequencing. All enzymes used were purchased from Fermentas unless otherwise indicated. Complete sequences of constructs used are available on request. For sequences of the primers used to amplify ORFs, see Table S1.
For plant transformation, constructs were transformed into Agrobacterium tumefaciens strain ASE by electroporation. Arabidopsis plants of the Col-0 accession were transformed by the floral-dip procedure [56]. Transgenic plants were selected with 0.1% glufosinate (BASTA) on soil or 50 µg/ml kanamycin on plates. At least 20 T1 plants were analyzed for each construct.
Total RNA was extracted from plant tissue using either the Plant RNeasy kit (Qiagen) or Trizol reagent (Invitrogen) according to the manufacturer's instructions. 2 µg of total RNA was DNase I-treated and single-stranded cDNA was synthesized using oligo(dT) and the RevertAid First Strand cDNA Synthesis Kit (Fermentas). Quantitative real-time PCR was performed on an Opticon Continuous Fluorescence Detection System (MJR) using the Platinum SYBR Green qPCR Supermix-UDG (Invitrogen). Gene expression was calculated relative to β-Tubulin using the ΔΔCT method. Results are reported for triplicate measurements of one of several biological replicates. For each genotype and replicate, a minimum of 10 seedlings was pooled for RNA extraction. Oligonucleotide primers used for qRT PCR are listed in Table S1.
For the analysis of the leaf transcriptome in Arabidopsis flc-3 and smz-D flc-3, plants were grown under SD conditions for 14 days and shifted to LD to induce flowering synchronously. Rosette leaves one to three from 10 plants were collected zero, 1 and 4 d after the plants were shifted to LD in duplicate, and total RNA was extracted using Qiagen Plant RNeasy columns (Qiagen). Biotinylated antisense RNA was prepared from 1 µg of total RNA using the MessageAmp II-Biotin Enhanced Kit (Ambion) according to the manufacturer's instructions. A total of 13.5 µg of fragmented amplified RNA (aRNA) was hybridized to an Arabidopsis ATH1-121501 gene expression array (Affymetrix). Arrays were washed and stained on a GeneChip Fluidics Station 450 (Affymetrix) and scanned on an Affymetrix GeneChip scanner GS300 7G. Analysis of the shoot meristem transcriptome was carried out as described above except that plants were grown for 25 days under SD before transfer to LD. RNA from shoot apices was isolated as described [34]. For visualization, normalized expression estimates were obtained by directly importing .CEL files into GeneSpring 10 using gcRMA (Agilent Technologies) and baseline transformation as a normalization routine. All microarray data are freely available from the ArrayExpress database (http://www.ebi.ac.uk/arrayexpress; accession numbers: E-MEXP-2040 (leaf samples) and E-MEXP-2041 (apices)). Lists of statistically significantly expressed genes (Datasets S1, S2, S3, and S4) were calculated for pairwise comparisons between time points within a given genotype or between genotypes at a given time point using RankProducts (version 2.6.0) implemented in R (version 2.4.0; GUI 1.17) on gcRMA (version 2.6.0) normalized expression estimates [57],[58].
The entire ChIP-chip experiment from sonication through array analysis was performed on technical duplicate samples from both 35S::NLS-3xVenusGFP and 35S::GFP-SMZ seedlings and then repeated on biological replicate samples. Briefly, seedlings grown for 9 LD were fixed at the end of the day as described previously [59]. Frozen tissue was ground, filtered three times through Miracloth (Calibrochem), and washed as described previously thorough buffers M1, M2, and M3 [59]. Nuclear pellets were resuspended in sonic buffer as described (1 mM PEFA BLOC SC [Roche Diagnostics] was substituted for PMSF), split into duplicate samples, and sonicated with a Branson sonifier at continuous pulse (output level 3) for eight rounds of 2×6 s and allowed to cool on ice between rounds. Immunoprecipitation (IP) reactions were performed by incubating chromatin with 2.5 µl of anti-rabbit GFP antibody (ab290, Abcam) overnight at 4°C, as described [59]. The immunoprotein–chromatin complexes were captured by incubating with protein A-agarose beads (Santa Cruz Biotechnology), followed by consecutive washes in IP buffer and then elution as described [59]. Immunoprotein-DNA was then incubated consecutively in RNase A/T1 mix (Fermentas) and Proteinase K (Roche Diagnostics) as described, after which DNA was purified using Minelute columns (Qiagen) [59]. Recovered DNA was amplified using the Sigma WGA GenomePlex kit (Sigma-Aldrich), after we performed a comparison to other systems, which showed this protocol gives improved amplification consistency and minimal amplification bias, in accordance with a previous study [60]. One microgram of DNA was fragmented, labeled, and hybridized to Affymetrix Arabidopsis tiling 1.0F arrays (Affymetrix). Chromatin size distribution and fragmentation performance was confirmed on an Agilent Bioanalyzer prior to array hybridization (Agilent Technologies). Regions found to be enriched by ChIP-chip were confirmed by manual ChIP. We performed triplicate qPCR on chromatin samples from 35S::SMZ-GFP and 35S::GFP-NLS plants. As a negative control, we used a region 3.5 kb away from the FT peak that was not enriched in ChIP-chip analysis.
Tiling array data were processed using the CisGenome suite [61]. Briefly, raw .CEL files were quantile normalized and peaks were called using TileMapv2. Analysis was performed in MA mode with window size 5, and only peaks detected with a FDR of better than 0.05 were analyzed. EasyGO was used to do GO-based enrichment analysis [62]. Genome-wide visualization was performed with Affymetrix Integrated Genome Browser after normalization with Affymetrix Tiling Array Software (Affymetrix). All tiling array data are freely available from the ArrayExpress database (http://www.ebi.ac.uk/arrayexpress; accession numbers: E-MEXP-2068).
|
10.1371/journal.pbio.0050321 | Incomplete and Inaccurate Vocal Imitation after Knockdown of FoxP2 in Songbird Basal Ganglia Nucleus Area X | The gene encoding the forkhead box transcription factor, FOXP2, is essential for developing the full articulatory power of human language. Mutations of FOXP2 cause developmental verbal dyspraxia (DVD), a speech and language disorder that compromises the fluent production of words and the correct use and comprehension of grammar. FOXP2 patients have structural and functional abnormalities in the striatum of the basal ganglia, which also express high levels of FOXP2. Since human speech and learned vocalizations in songbirds bear behavioral and neural parallels, songbirds provide a genuine model for investigating the basic principles of speech and its pathologies. In zebra finch Area X, a basal ganglia structure necessary for song learning, FoxP2 expression increases during the time when song learning occurs. Here, we used lentivirus-mediated RNA interference (RNAi) to reduce FoxP2 levels in Area X during song development. Knockdown of FoxP2 resulted in an incomplete and inaccurate imitation of tutor song. Inaccurate vocal imitation was already evident early during song ontogeny and persisted into adulthood. The acoustic structure and the duration of adult song syllables were abnormally variable, similar to word production in children with DVD. Our findings provide the first example of a functional gene analysis in songbirds and suggest that normal auditory-guided vocal motor learning requires FoxP2.
| Do special “human” genes provide the biological substrate for uniquely human traits, such as language? Genetic aberrations of the human FoxP2 gene impair speech production and comprehension, yet the relative contributions of FoxP2 to brain development and function are unknown. Songbirds are a useful model to address this because, like human youngsters, they learn to vocalize by imitating the sounds of their elders. Previously, we found that when young zebra finches learn to sing or when adult canaries change their song seasonally, FoxP2 is up-regulated in Area X, a brain region important for song plasticity. Here, we reduced FoxP2 levels in Area X before zebra finches started to learn their song, using virus-mediated RNA interference for the first time in songbird brains. Birds with experimentally lowered levels of FoxP2 imitated their tutor's song imprecisely and sang more variably than controls. FoxP2 thus appears to be critical for proper song development. These results suggest that humans and birds may employ similar molecular substrates for vocal learning, which can now be further analyzed in an experimental animal system.
| Genetic aberrations of FOXP2 cause developmental verbal dyspraxia (DVD), which is characterized by impaired production of sequenced mouth movements and both expressive and receptive language deficits [1–4]. Brain imaging studies in adult FOXP2 patients implicate the basal ganglia as key affected regions [5–7], and FOXP2 is prominently expressed in the developing human striatum [8]. These findings raise the question whether the speech and language abnormalities observed in individuals with DVD result from erroneous brain development or impaired function of differentiated neural circuits in the postnatal brain, or a combination of both. Human speech and learned vocalizations in oscine birds bear behavioral and neural parallels [9]. Thus songbirds are a suitable model for studying the neural mechanisms of imitative vocal learning, including speech and its pathologies. The FoxP2 expression patterns in songbird and human brains are very similar, with strong expression in the basal ganglia, thalamus, and cerebellum [8,10,11]. Moreover, FoxP2 expression in the basal ganglia song nucleus, Area X, which is important for normal song development [12,13], transiently increases at the time when young zebra finches learn to sing. In adult canaries, FoxP2 expression in Area X is elevated during the late summer months, coincident with the incorporation of most new syllables to their seasonally changing song [10]. FoxP2 is down-regulated in Area X when adult zebra finches sing slightly variable, undirected song, but not when they sing more stereotyped female-directed song [14]. Together, these correlative findings raise the question whether FoxP2 and vocal plasticity are causally related.
Using lentivirus-mediated RNA interference (RNAi) during song development, we now show that zebra finches with reduced FoxP2 expression levels in Area X imitated tutor songs incompletely and inaccurately. This effect was already evident during vocal practice in young birds. Moreover, the acoustic structure and the duration of song syllables in adults were abnormally variable, similar to word production in children with DVD [15]. These findings are consistent with a role of FoxP2 during auditory-guided vocal motor learning in songbird basal ganglia.
Vocal learning in zebra finches proceeds through characteristic stages. In the sensory phase that commences around 25 d after hatching (post-hatch day [PHD]), young males memorize the song of an adult male tutor. Concomitantly, they start vocalizing the so called “subsong,” consisting of quietly uttered, poorly articulated, and nonstereotypically sequenced syllables [16]. Following intensive vocal practice and improvement toward matching the tutor song during the period of “plastic song,” they eventually imitate the song of their tutor with remarkable fidelity around PHD90. The structural and temporal characteristics of adult “crystallized” song remain essentially stable throughout adult life. To study the function of FoxP2 during song learning of zebra finches, we reduced the levels of FoxP2 expression bilaterally in Area X in vivo, using lentivirus-mediated RNAi. In this approach, short interfering hairpin RNA (shRNA) containing sense and antisense sequences of the target gene connected by a hairpin loop are expressed from a viral vector. The virus stably integrates into the host genome, enabling expression throughout the life of the animal [17].
We designed two different shRNAs (shFoxP2-f and shFoxP2-h) targeting different sequences in the FoxP2 gene. Both hairpins strongly reduced the levels of overexpressed FoxP2 protein in vitro (Figure 1F), but did not change the levels of overexpressed protein levels of FoxP1, the closest homolog of FoxP2. For further control experiments, we generated a shRNA designed not to target any zebra finch gene (shControl). As expected, this nontargeting shRNA did not affect expression of either FoxP2 or FoxP1 in vitro (Figure 1F). Since shFoxP2-f and shFoxP2-h targeted FoxP2 with similar efficiency, both of them were interchangeably used for subsequent in vivo experiments (shFoxP2-f/-h).
On PHD23, at the onset of sensory-motor learning [16] we injected either FoxP2 knockdown or control viruses (shControl and shGFP; see below) stereotaxically into Area X to achieve spatial control of knockdown (Figure 1A). Starting on PHD30, each young bird, here called pupil, was kept in a sound isolation chamber, together with an adult male zebra finch as tutor. At PHD65, PHD80, and between PHD90 and PHD93, we recorded the pupils' vocalization for subsequent song analysis (for timeline of experiments, see Figure S1). After the last song recording, brains were histologically analyzed for correct targeting of virus to Area X. All lentiviral constructs expressed the green fluorescent protein (GFP) reporter gene, allowing the detection of infected brain areas by fluorescence microscopy (Figure 1B). On average, 20.3% ± 9.9% (mean ± standard deviation [STDV]; n = 24 hemispheres from 12 animals) of the total volume of Area X was infected. Importantly, there was no difference in the volume of Area X targeted with FoxP2 knockdown or control viruses (two-tailed Mann-Whitney U test, p > 0.5; shFoxP2-f/-h, n = 6, shControl, n = 7). Quantification of Area X volume targeted by virus injection in an equally treated group of birds, but sacrificed at PHD50, confirmed the results obtained for PHD90 (mean volume 20.4% ± STDV 4.0%; two-tailed Mann-Whitney U test, p > 0.6; shControl n = 3 hemispheres from 3 animals; shFoxP2-f/-h, n = 3 hemispheres from 3 animals).
To quantify the neuronal extent of lentivirus expression in Area X, we used immunohistochemical staining with the neuronal marker Hu [18] (Figure S2). Of all virus-infected cells, 78.5% ± 3.5% were neurons (mean ± standard error of the mean [SEM]; no significant difference between shFoxP2 and shControl, two-tailed Mann-Whitney U test, p > 0.7; shControl injections n = 3 hemispheres from 3 animals, shFoxP2 injections n = 4 hemispheres from 4 animals;). This result is consistent with Wada et al. [19], who used the same viral constructs in the zebra finch brain in vivo. Among the infected cells were FoxP2-positive spiny neurons, which are assumed to be the most common cell type in Area X [20] (Figure 1C–1E).
To quantify FoxP2 knockdown in vivo, we determined FoxP2 protein levels in Area X on PHD50, the time of peak FoxP2 expression [10] in birds injected on PHD23 with shFoxP2-f/-h in one hemisphere and shControl into the contralateral hemisphere. The signal of the immunofluorescent staining with a FoxP2 antibody was significantly lower in knockdown Area X than in control Area X (Figure 1G and 1H). We also assessed FoxP2 mRNA levels after knockdown in Area X. Birds were injected on PHD23 with shFoxP2-f/-h in one hemisphere and shControl in the contralateral hemisphere. On PHD50, we punched out Area X of injected birds and measured FoxP2 mRNA levels by real-time PCR. FoxP2 levels were normalized to two independent RNAs coding for the housekeeping genes Hmbs and Pfkp. FoxP2 mRNA was reduced on average by approximately 70% in the shFoxP2-infected region of Area X compared to the shControl-infected region of Area X (Figure 1I). Of note, RNAi-mediated knockdown approximates FOXP2 levels in DVD patients, since haploinsufficiency, a 50% reduction of functional FOXP2 protein, is apparently the common feature of all reported human FOXP2 mutations [4,21].
To demonstrate that RNAi-mediated gene knockdown can persist in vivo throughout the entire song-learning phase, we used a virus expressing shRNA against the viral reporter GFP (shGFP) in conjunction with the virus expressing a shRNA lacking a target gene (shControl). We injected young zebra finches on PHD23 with equal amounts of equally infectious shGFP and shControl virus in the left and right hemisphere, respectively. More than 3 mo later, on PHD130, the GFP signal in the shGFP-injected hemisphere was still 70.5% ± 5.8% less intense than in the shControl-injected hemisphere (mean ± SEM; n = 2; Figure 1J).
To rule out potential side effects of FoxP2 knockdown on cellular survival in Area X, we investigated apoptosis in Area X 6 d after surgery with terminal deoxyribonucleotide transferase-mediated dUTP nick end labeling (TUNEL). The TUNEL method detects genomic DNA double-strand breaks characteristic of apoptotic cells. Of 1,149 GFP-positive cells counted in six hemispheres from three animals, only five were TUNEL-positive (Figure S3). ShControl-injected and uninjected animals had similar low levels of apoptotic cells (unpublished data). Thus, FoxP2 is not a gene essential for short-term survival of postmitotic neurons. Since the TUNEL method does not capture any long-term changes in neuronal viability that might follow after reduction of FoxP2, we used the neuronal marker Hu to determine neuronal densities in Area X 30 d after injecting either shFoxP2-f/-h or shControl virus (Figure S4). Neuronal densities in the infected region in Area X did not differ in knockdown and shControl-injected birds (two-tailed Mann-Whitney U test, p > 0.39; shControl, n = 4 hemispheres; shFoxP2-f/-h, n = 3 hemispheres). Density of neurons were also comparable inside and outside of the virus-infected region of Area X for all viruses (two-tailed Mann-Whitney U test, p > 0.6 for both shFoxP2-f/-h and shControl). In sum, these data demonstrate that virus-mediated RNAi can induce specific, long-lasting knockdown of gene expression in zebra finch Area X without causing cell death.
Adult zebra finch song consists of different sound elements, here called syllables, that are separated by silent intervals. Syllables are rendered in a stereotyped sequential order, constituting a motif. During a song bout, a variable number of motifs are sung in short succession. To obtain a first descriptive account of the song of knockdown and control pupils, we measured mean acoustic features for all syllables recorded from all pupils using the software Sound Analysis Pro (SAP) [22]. The features extracted were mean pitch, mean frequency, mean frequency modulation (FM; change of frequency in time), mean entropy, and mean pitch goodness (PG; periodicity of sound), as well as mean duration. The comparison of the distribution of these features across the repertoire of knockdown and control pupils did not reveal any significant differences, indicating that knockdown pupils, control pupils, and tutors sang syllables with similar acoustic features (Figure S5).
Next, we analyzed the behavioral consequences of bilateral FoxP2 knockdown in Area X for the outcome of song learning at PHD90. When a juvenile male finch is tutored individually by one adult male, the pupil learns to produce a song that strongly resembles that of his tutor [23]. We therefore determined learning success by the degree of acoustic similarity between pupil and tutor songs. Analysis of song recorded at PHD90 revealed that pupils with experimentally reduced FoxP2 levels in Area X imitated tutor songs with less fidelity than control animals did (see also Audio S1–S6). The comparison of sonograms from shControl-injected (Figure 2A) and shFoxP2-injected pupils (Figure 2B and 2C) with their respective tutors shows the characteristic effects caused by reduction of FoxP2. Typical features of FoxP2 knockdown pupils included syllable omissions (Figure 2B, syllables C, D, F, and G; Figure 2C, syllable B), imprecise copying of syllable duration (Figure 2B, syllable E longer; Figure 2C, syllable D shortened), and inaccurate imitation of spectral characteristics (Figure 2B, syllable E; Figure 2C, syllable D). In addition, in four out of seven knockdown pupils, the motif contained repetitions of individual syllables or syllable pairs (e.g., see Figure 2B and 2C). In contrast, none of the control or tutor motifs contained repeated syllables. Pupils did not reverse the sequential order of syllables in the tutor motifs, except for one control (unpublished data) and one FoxP2 knockdown pupil (Figure 3A).
Acoustic similarity between pupil and tutor song was measured with SAP by pairwise comparison of user-defined pupil and tutor motifs. SAP provides a similarity score that indicates how much of the tutor sound material was imitated by the pupil, regardless of syllable order. The distinction between imitated and non-imitated sounds in SAP is based on p-value estimates derived from the comparison of 250,000 sound interval pairs, obtained from 25 random pairs of zebra finch songs (see Materials and Methods and [22] for further details). The similarity score was significantly lower in FoxP2 knockdown than in control animals (Figure 2D). In addition, we also manually counted the number of user-defined syllables copied from the tutors, confirming that knockdown animals imitated fewer syllables (Figure S6).
Even though knockdown animals copied tutor syllables, their imitation appeared to be less precise than in control animals. Figure 3A illustrates the inaccurate syllable imitation (syllables A and B) in a knockdown pupil that learned from the same tutor as the shControl-injected pupil shown. To quantify how well the syllables of a motif were imitated on average, we obtained motif accuracy scores in SAP from pairwise motif comparisons between pupil and tutor. The motif accuracy score measures the extent to which the pupil's sounds are closer to the tutor than expected by chance. The average accuracy per motif was significantly lower in knockdown pupils than in shControl-injected pupils (Figure 3B). Of note, both shFoxP2 hairpins (shFoxP2-f and shFoxP2-h) affected motif similarity and motif accuracy scores to a similar degree (Figure S7), which is consistent with their comparable efficiency in reducing FoxP2 mRNA in vitro (Figure 1F). Neither the similarity score nor the accuracy score correlated with the volume of Area X targeted in the pupil. Possibly, there were too few values to observe such a correlation or the absolute volume targeted by shFoxP2 virus has only a small influence on the outcome of learning.
To investigate whether inaccurate imitation affected all or only some syllables, we compared corresponding syllable pairs between tutors and pupils using a syllable identity score. The syllable identity score reflects both the degree of similarity (i.e., quantity of imitation) and the degree of accuracy (i.e., quality of imitation) in a single measure. The frequency distribution of identity scores of all syllables from FoxP2 knockdown pupils was shifted towards lower scores compared to control pupils. This suggests that imprecise imitation was not skewed towards particular syllables or syllable types (Figure 3C), pointing to a generalized lack of copying precision. Consistent with the reduced accuracy of motif imitation (Figure 3B), we also found that syllable identity scores were significantly lower in knockdown pupils compared to control pupils (syllable identity score averaged for each animal, two-tailed Mann-Whitney U test, p < 0.02; n = 7 for both shFoxP2 and shControl).
To rule out that the lower imitation success of knockdown animals was related to specific song characteristics of the tutors or their lacking aptitude for tutoring, we used some males to tutor both knockdown and control pupils. Direct comparison of the motif similarity and accuracy scores from control and knockdown pupils tutored by the same male revealed significantly lower scores for knockdown compared to control pupils (average similarity score 82.6 ± 3.6 for shControl and 61.9 ± 5.6 for shFoxP2; average accuracy score 73.8 ± 0.7 for shControl and 71.7 ± 0.4 for shFoxP2; ± SEM; n = 5, two-tailed Mann-Whitney U test, p < 0.03 for similarity and p < 0.03 for accuracy; see also Figure 3A).
Because the shControl hairpin, in contrast to shFoxP2-f/-h, has no target gene, it might not stably activate the RNA-induced silencing complex (RISC) essential for knockdown of gene expression during RNAi. Because recent work suggests an involvement of the RISC in the formation of long-term memory in the fruitfly [24] we addressed a possible influence of RISC activation during song learning. For this, we compared song imitation in shGFP virus–injected pupils, in which virally expressed GFP is lastingly knocked down (Figure 1J), and shControl-injected pupils. Similarity and accuracy scores did not differ significantly between shGFP-injected and shControl-injected animals, ruling out that RISC activation contributed to the effects of shFoxP2 on song imitation (Figures 2D and 3B).
Finally, we investigated the precision of syllable imitation on the level of individual acoustic features by comparing the mean values of acoustic features of pupil syllables to those of their respective tutor. The divergence of imitated syllables from the tutor tended to be larger in all acoustic measures in the FoxP2 knockdown pupils than in the controls. For average syllable duration and mean entropy measures, the difference was significant (Figure 3D).
Area X is part of a basal ganglia–forebrain circuit, the anterior forebrain pathway (AFP), which bears similarities with mammalian cortical–basal ganglia loops [25]. The pallial target of the AFP, nucleus lateral magnocellular nucleus of the nidopallium (lMAN), may act as a neural source for vocal variability in juvenile zebra finches [13,26]. Similarly, in adult zebra finches, neural variability in AFP outflow is associated with the variability of song [27], and experimental manipulations inducing adult song variability require an intact AFP [28,29]. To explore AFP function in FoxP2 knockdown and control zebra finches, we investigated the variability of their song syllables.
The comparison of sonograms from different renditions of the same syllable revealed that knockdown pupils sang their syllables in a more variable fashion than control pupils (Figure 4A and 4B). Both the spectral (syllables I and III) and the temporal domain (syllables II and IV) were affected. Of note, the first three syllable examples shown in Figure 4A and 4B (syllables I, II, and III and I′, II′, and III′), stem from different animals, but were learned from the same tutor. To quantify the acoustic variability of syllables, we used the syllable identity score mentioned above. Pairwise comparison between different renditions of the same syllable revealed that shFoxP2-injected pupils sang syllables slightly, but significantly, more variably than control pupils or tutors (Figure 4C). As expected, shControl-injected pupils, shGFP-injected pupils, and tutors performed their syllables with equal stability (Figure 4C).
Next, we quantified the variability of syllable duration between different renditions of the same syllable. The coefficient of variation of syllable duration was significantly higher in knockdown than in control pupils and tutors, suggesting imprecise motor coordination on short temporal scales (Figure 4D). Notably, the timing of syllables in control pupils (shControl and shGFP) was as stable as in tutors (Figure 4D). The variability of syllable duration in tutor and control birds varied in the same range as reported previously [30], emphasizing how tightly adult zebra finches normally control syllable duration.
Finally, we analyzed the sequential order of syllables over the course of many motifs. To this end, we first annotated sequences of 300 user-defined syllables with the positions in their respective motifs. We then measured the stereotypy of a motif by calculating for each syllable the entropy of its transition distribution. Based on this entropy measure, we generated a sequence consistency score (1 − entropy), which reflects song stereotypy. An entropy score of 0 indicates random syllable order, whereas a score of 1 reflects a fixed syllable order. The mean sequence consistency was similar in shControl and shFoxP2-f/-h animals (Figure S8). Because stereotypy of motif delivery is a hallmark of “crystallized” adult song, it seems plausible that both knockdown animals and controls had reached the end of the sensory-motor learning period [31]. To investigate this question in more detail, we next analyzed the song of knockdown and control pupils recorded at earlier stages of song development.
To explore the developmental trajectory of song learning in knockdown and control pupils, we analyzed songs recorded during plastic song at PHD65 and towards the end of the learning phase at PHD80. Since syllables are not yet rendered in a stereotyped motif structure at PHD65, we quantified song imitation success and vocal variability on the level of the syllables only. To avoid the necessity of identifying individual syllables based on their morphology, we made use of an automated procedure provided by SAP to compare all song material from a given day to the tutor's typical motif. The vocalizations of pupils were first segmented into syllables. All segments were subsequently compared to the typical motif of the tutor in a pairwise fashion (between 1,000–3,000 comparisons per pupil per day). The output variable of these measurements is an accuracy score, which describes the extent to which the pupil's sounds match those of the tutor (see Materials and Methods and [22] for further details). We found that knockdown pupils imitated their tutors less accurately than control pupils already at PHD65 (Figure 5A). The frequency distribution of accuracy values also suggests that imprecise syllable imitation was not skewed towards particular syllables or syllable types (Figure S9). This result is in line with the observation made earlier for the syllables at PHD90 (Figure 3C). In contrast to control pupils, knockdown pupils did not improve in accuracy after PHD80, suggesting they had reached the end of the learning phase (Figure 5A).
For each pupil, we also calculated the change of accuracy from one age to the next (accuracy [agen − agen−1]). The change of accuracy from PHD65 to PHD80 was indistinguishable between knockdown and control pupils (two-tailed Mann-Whitney U test, p > 0.9; n = 5 for shFoxP2-f/-h and n = 7 for shControl), suggesting that up to this age, syllable imitation followed largely similar dynamics. However, from PHD80 to PHD90, accuracy of syllable imitation continued to improve only in control, but not in knockdown pupils (two-tailed Mann-Whitney U test, p < 0.04; n = 6 for shFoxP2-f/-h and n = 6 for shControl).
In order to investigate variability of syllable production during song development, we compared the variance of accuracy values between knockdown and control pupils. Whereas the variance was similar between the two experimental groups at PHD65 and at PHD80, it was significantly higher in knockdown pupils compared to controls at PHD90 (Figure 5B). This difference resulted from an increase of variance with age in shFoxP2-injected birds (Figure 5B). Of note, the similarity batch analysis, which does not require assumptions about the identity of individual motifs or syllables, confirmed the results on both lower imitation success and higher vocal variability obtained in our prior analysis of the songs from PHD90 (Figures 2D and 3B).
Our goal was to investigate the requirement of FoxP2 for normal song development in the zebra finch, a model for studying the basic principles of vocal learning. To this end, we analyzed the behavioral consequence of an experimental reduction of FoxP2 during song development. Using lentivirus-mediated RNAi for the first time in the songbird brain, we reduced FoxP2 mRNA and protein levels in Area X with either of two different knockdown constructs. We found that this prevented complete and accurate imitation of the tutors' song, an effect already evident during plastic song. Reduced FoxP2 levels also led to more variable performance of syllables in adults. In contrast, we observed no such abnormalities in birds with Area X injections of virus knocking down an exogenously expressed gene (GFP) or expressing a nontargeting control construct. In addition, we verified in vitro that knockdown of FoxP2 did not affect protein levels of FoxP1, the closest homolog of FoxP2. FoxP2 knockdown also did not cause apoptotic cell death in Area X, and it did not alter the density of neurons in this nucleus. Consistent with this, FoxP2 knockdown pupils showed different song abnormalities than birds with electrolytic lesions in Area X. Juvenile Area X lesions result in low sequence consistency, and the repertoire of birds with juvenile Area X lesions contains unusually long syllables [13], which were not observed in FoxP2 knockdown finches (Figure S5). Together, these data rule out that unspecific effects of RNAi induction, viral infection, or damage to Area X influenced our results. We further eliminated the possibility that specific song features of the tutor birds contributed to the behavioral differences. The outcome of song learning was affected by virus infection in approximately 20% of the volume of Area X. This result is consistent with a previous study on virally injected rats, in which blocking neural plasticity in 10%–20% of lateral amygdala neurons was sufficient to impair memory formation [32]. Taken together, these data strongly suggest that insufficient levels of FoxP2 in Area X spiny neurons lead to incomplete and inaccurate vocal imitation, implicating FoxP2 in postnatal brain function.
The incomplete and inaccurate vocal imitation of tutor song in FoxP2 knockdown pupils raises the question whether knockdown pupils were unable to generate particular sounds. Given that syllables with similar spectral features could be learned or omitted by the same pupil (e.g., in Figure 2B, tutor syllables E and G are similar; pupil imitated E, but not G), this does not seem likely. Also, omitted syllables did not differ in their spectral feature composition from those that were learned by knockdown animals (unpublished data). Consistent with this, the distributions of mean syllable feature values and mean duration across the syllable repertoire were indistinguishable between knockdown and control pupils (Figure S5). However, it is still possible that FoxP2 knockdown affected the motor control of singing. The fact that FoxP2 knockdown pupils produced syllables more variably than controls at PHD90 would be consistent with this. Importantly though, this increased variability of syllable rendition in FoxP2 knockdown pupils was not yet evident at PHD65, when tutor imitation was already less proficient (Figure 5B). Thus, the increased syllable variability is apparently not causally related to the observed tutor imitation deficit. Unfortunately, song analysis alone cannot ultimately distinguish between impairments in motor production and motor learning. Any motor production deficit likely affects the auditory feedback signal, which in turn is bound to reduce the quality of tutor imitation. Knockdown of FoxP2 in adult zebra finches might help to clarify the contribution of FoxP2 to motor control.
Although knockdown animals were apparently not unable to produce particular syllable types, given the involvement of the basal ganglia in the acquisition and performance of motor sequences [33], knockdown pupils might have been impaired in producing particular sequences of syllables, i.e., in moving from one syllable to the next. We found that knockdown pupils could in principle imitate adjacent tutor syllables in the same order (e.g., Figure 2B, syllables A and B, and H and I; Figure 2C, syllables C and D). There was also no preferred position (i.e., beginning or end of song) for imitated and non-imitated syllables. Moreover, potential sequencing problems might occur at different syllable transitions within the motif or intermittently in different renditions of the motif. Both scenarios would result in low sequence stereotypy, which we did not find (Figure S8). The limited imitation success of FoxP2 knockdown pupils could also result from an imprecise neural representation of the tutor model. There is evidence for an involvement of Area X in sensory learning at PHD35 [34], but the up-regulation of FoxP2 in Area X at PHD50 and PHD75 rather speaks for an involvement of FoxP2 in sensory-motor learning [10].
Under the assumption of a model of reinforcement-based motor learning mediated through the basal ganglia, the animal initially generates variable motor output. Progressively, particular motor actions are reinforced [33]. In view of this model, FoxP2 knockdown pupils might have either experienced a limitation in generating enough sound variability or difficulties with reinforcing the “right” motor patterns, a possibility that includes both difficulties in detecting similarity to the target or adjusting song appropriately. Since knockdown pupils sing as variable as control pupils early during song development and even more variable as adults (Figures 4 and 5B), we favor the hypothesis that knockdown pupils were impaired in adjusting their motor output according to the memorized tutor model in the course of song learning. This hypothesis is supported by the phenotypic overlap of song deficits observed in FoxP2 knockdown pupils and birds that were prevented from matching vocal output with memorized tutor song. For instance, perturbed auditory feedback provokes syllable repetitions [35], and deafening in juveniles brings about syllables with large acoustic variability [36]. Although we cannot ultimately rule out the possibility that the impairment observed after FoxP2 knockdown in juvenile birds was primary motor in nature, an interpretation involving a deficit with auditory-guided motor learning seems more consistent with the knockdown song phenotype.
What is the mechanism by which FoxP2 contributes to song development? In Area X, spiny neurons receive pallial glutamatergic input from Area X–projecting neurons in HVC [37]. These neurons process auditory information and are active during singing [38,39]. FoxP2 expressing spiny neurons also receive nigral dopaminergic input [10,40]. As has been suggested for motor learning in mammals [41], midbrain dopaminergic activity could act as reinforcement signal during song learning. Therefore, the integration of pallial and dopaminergic signals provides a candidate mechanism for tuning the motor output to the tutor model during learning. The increase of FoxP2 expression in Area X of zebra finches during times of vocal plasticity could be functionally related to this process. FoxP2 might mediate adaptive structural and functional changes of the spiny neurons while the song is learned. During the seasonal phase of vocal plasticity in canaries, increased FoxP2 expression in the fall months might similarly be involved in seasonal song modifications. Since FoxP2 is a transcription factor, it could act by positively or negatively regulating plasticity-related genes. If FoxP2 functions as a plasticity-promoting factor, knockdown pupils should have been less plastic during learning, resulting in impoverished imitation and abnormally invariant song. Syllable omissions of FoxP2 knockdowns are consistent with this notion, but more variable syllable production is clearly not. Alternatively, if FoxP2 restricts neuronal plasticity, knockdown pupils should sing more variable song. In fact, this is the case, but syllable omissions are not easily explained then. The identification of the downstream target genes of FoxP2 and the electrophysiological characterization of spiny neurons with reduced FoxP2 levels will shed light on the mechanisms by which FoxP2 affects the outcome of vocal learning.
The vocal behavior of FoxP2 knockdown zebra finches offers a new interpretation of the speech abnormality in individuals with genetic aberrations of FOXP2 [5], possibly extending to apraxia of speech in general [42]. The human core deficit affects the production of rapid, sequential mouth movements, which are required for speech articulation [43], and is thought to be caused by erroneous brain development. Perhaps the speech impairment results from a problem with motor learning rather than motor performance during speech learning, a hypothesis that is in line with recent theories on basal ganglia dysfunction in various developmental disorders [44]. Our results extend the similarities between learned birdsong and human speech to the molecular level, emphasizing the suitability of songbirds for investigating the basic principals of speech and its pathologies. It will be interesting to test, whether “dyspraxic song” is also perceived as different by other finches and interferes with communication, as DVD does in humans. Given female songbirds' preference for well-learned, experimentally unaltered song [45,46], we would expect this to be the case. Finally, the fact that a reduction of FoxP2 affects the outcome of both song learning and speech development provides further evidence for the hypothesis [4,21] that during evolution, ancestral genes and neural systems were adapted in the human brain and gave rise to the uniquely human capacity of language.
For FoxP2 nomenclature, we followed the convention proposed by the Nomenclature Committee for the forkhead family of genes (FOXP2 in Homo sapiens, Foxp2 in Mus musculus, and FoxP2 in all other species, including zebra finches) [47]. Proteins are in roman type, genes and RNA in italics.
Initially, we designed eight different constructs for the expression of short hairpin RNA (shRNA) targeting the zebra finch FoxP2 mRNA. All FoxP2 target sequences were located within the minimum common sequence of all isoforms (ORF of isoform IV), thus targeting all FoxP2 isoforms described in [10]. In order to minimize potential cross-reactivity of the hairpins, we chose target sequences that contained at least six dissimilar bases with FoxP1, the closest homolog of FoxP2. and were not located within the highly conserved forkhead box domain of FoxP2. This shRNA design is stringent in comparison to a recently published guideline [48] that recommends including at least three mismatches to untargeted sequences. The structure of the linear DNA encoding shRNA hairpins was sense-loop-antisense. The sequence of the loop was GTGAAGCCACAGATG. Each hairpin construct was tested for knockdown efficiency in HEK293 T cells in vitro by simultaneous overexpression with zebra finch FoxP2, tagged with the V5 epitope. Subsequent western blot analysis using a V5 antibody revealed two hairpins (shFoxP2-f, target sequence AACAGGAAGCCCAACGTTAGT, and shFoxP2-h, target sequence AACGCGAACGTCTTCAAGCAA) that strongly reduced FoxP2 expression levels. To demonstrate the sequence specificity of the hairpins to the FoxP2 gene, we also simultaneously overexpressed them with FoxP1, cloned from adult zebra finch brain cDNA and tagged with the V5 epitope. The DNA fragments encoding the hairpins shFoxP2-f and shFoxP2-h were subcloned into a modified version of the lentiviral expression vector pFUGW [17] containing the U6 promoter to drive their expression. To use as controls, we subcloned fragments encoding a hairpin targeting GFP (shGFP, target sequence GCAAGCTGACCCTGAAGTTCA) and a nontargeting hairpin (shControl, sequence AATTCTCCGAACGTGTCACGT) into the modified pFUGW. All viral constructs expressed GFP under control of the human ubiquitin C promoter. Recombinant lentivirus was generated as described in [17]. Titers were adjusted to 1–2 × 106/μl.
The general procedure for studying the behavioral consequences of locally reduced FoxP2 levels in Area X was as follows (Figure S1). Young zebra finches from our colony at the Max-Planck-Institute for Molecular Genetics were sexed as described [49] at approximately PHD10. By PHD20, fathers and older male siblings were removed from family cages to prevent experimental zebra finches from instructive auditory experience prior to the onset of tutoring. At PHD23, animals were anaesthetized with xylazine/ketamine and stereotaxically injected with recombinant lentivirus. The stereotaxic coordinates for Area X injections were anterior/posterior 3.6 and 4.0, medial/lateral 1.4 and 1.6, and dorsal/ventral 3.8 and 4.0. Per injection site, approximately 200 nl of lentiviral solution were injected over a period of 2 min with a hydraulic micromanipulator (Narishige). On PHD30, each pupil received an adult male song tutor, and both birds were kept together for 2 mo in a sound-isolated box with automated song-recording equipment. By PHD93–95, trained pupils were perfused with 4% paraformaldehyde in 0.1 M PB and their brains dissected for histological analysis (see Figure S1 for timeline of experiments).
We determined that the virus infected FoxP2 immunopositive neurons using immunostaining as described [10]. Moreover, we used immunohistological staining with antibody Hu (1:200; Chemicon) to stain neurons and quantify the percentage of them infected by virus. Immunofluorescent sections were analyzed with a 40× oil objective, using a Zeiss confocal microscope (LSM510) with the LSM-510 software package. On average, we counted 417 virus-infected cells in five to six sections per hemisphere (seven hemispheres from five animals) and determined how many of those were also Hu+. We quantified the neuronal density by counting the number of Hu+ cells in scanning windows of 230.3 μm × 230.3 μm (two scanning windows per section) inside and outside the injection site in Area X (presented as a number of cells/mm2).
To identify apoptotic cells, we used a fluorescein TUNEL assay (Roche) in 50-μm sagittal sections from PHD29 male zebra finch brains, injected with shFoxP2 virus on PHD23. To increase signal intensity, we stained the sections by fluorescent immunohistochemistry with an anti-FITC antibody, followed by incubation with an Alexa568-conjugated secondary antibody. TUNEL-positive cells were counted using a fluorescence microscope. In general, the total number of TUNEL-positive cells was very low (approximately eight cells per 50-μm brain section). There was no difference between knockdown and control animals in the total number of TUNEL-positive cells.
In order to quantify the volume of Area X targeted by virus injection, we measured the area of Area X in all brain sections (thickness, 50 μm) containing it, and quantified the region visibly expressing GFP within Area X under 5× magnification on a fluorescence microscope. We then summed the values from all sections for both areas separately and calculated the ratio of GFP-positive area to total Area X, which is equivalent to the ratio of GFP-positive volume to total Area X volume. The values from left and right hemispheres were averaged per animal. In one knockdown animal, GFP expression in Area X was detected only in the right hemisphere. Since this pupil had a motif imitation score of 50.8%, which is below the range of controls (68.1 ± 2.7% mean ± SEM), but better than knockdown pupils (39.6 ± 5.0 mean ± SEM), it could be that knockdown of FoxP2 in Area X of only the right hemisphere suffices to impair song learning consistent with right hemispheric dominance in zebra finches [50].
In six animals injected with either shFoxP2-f/-h or shControl virus, no GFP was detectable after histological analysis. We quantified imitation success in three of the six animals without GFP, and found it to be similar to zebra finches with shControl injection (similarity score = 90.7; accuracy score = 77.7; two-tailed Mann-Whitney U test, p > 0.8 for both similarity and accuracy).
Young male zebra finches received an injection of shFoxP2-f/-h virus in one hemisphere and an injection with control virus (shControl) in the contralateral hemisphere on PHD23 as described above. For the quantification of protein levels after FoxP2 knockdown, we performed an immunohistological staining with the FoxP2 antibody on 50-μm sections 30 d after virus injection. Immunohistological staining was performed as described [10], but using an antibody dilution of 1:5,000. All sections were processed at the same time with the same batch of antibody solution. Images of stained brain sections were taken with a digital camera using the Simple PCI software (Compix) at 40× magnification. For each section, we acquired multiple Z-stacked images of the virus-infected area (230.3 μm × 230.3 μm), and reconstructed a maximal projection. All images from the same bird were taken with the same microscope and software settings. Finally, we quantified fluorescence intensity levels in the images. The intensity of the green fluorescence from the viral GFP was not significantly different between shFoxP2-f/-h–injected and shControl-injected hemispheres (two-tailed Mann-Whitney U test, p > 0.3).
For the quantification of FoxP2 knockdown mRNA levels, young male zebra finches were injected with shFoxP2-f/-h virus in one hemisphere and control virus (shControl) in the contralateral hemisphere on PHD23, as described above. This permitted analysis of FoxP2 knockdown in the same bird while avoiding confounding differences in gene expression levels between birds. On PHD50, we sacrificed the birds and excised the GFP-expressing brain area with a 1-mm–diameter glass capillary (Brand) under a fluorescence dissecting microscope. RNA was extracted with TRIZOL (Invitrogen); yield was determined by UV spectroscopy at 260/280 nm with a Nanodrop device. FoxP2 expression was quantified by real-time PCR using SybrGreen (Applied Biosystems). We determined relative FoxP2 expression levels through normalization to the expression levels of two internal control genes, which were identified in a BLAST homology search for the mouse housekeeping genes Hmbs and Pfkp in the database from the Songbird Neurogenomics Initiative (http://titan.biotec.uiuc.edu/songbird/) and the Songbird Brain Transcriptome Database (http://songbirdtranscriptome.net/). The expression of Hmbs and Pfkp in the left and right hemisphere in both injected and untreated animals was equivalent (numbers indicate fold change between left and right hemispheres; untreated: Hmbs = 1.4 ± 0.5 and Pfkp = 1.3 ± 0.6; injected: Hmbs = 1.0 ± 0.4 and Pfkp = 1.1 ± 0.4, n=5 birds). Relative expression levels were determined with the comparative cycle time (Ct) method. All primers used in this study amplified the cDNA with similar efficiency (E = 1 ± 5%) in a validation experiment. Normalized Ct values from the same animal were calibrated to the shControl-injected hemisphere. FoxP2 expression levels are thus presented as the ratio of expression in shControl- to shFoxP2 -injected hemispheres.
Vocalizations were recorded between 9 am and 4 pm on PHDs 65, 80, and between 90 to 93 in absence of the tutor. Quantitative song analysis was performed using the SAP software, version 1.04 [22,51]. We analyzed song at the level of the syllables, the motif, and syntax. We define “syllable” as a continuous sound element, surrounded by silent intervals. The “typical song motif” was defined as the succession of syllables that includes all syllable types (except introductory notes), and occurs in a repeated manner during a song bout. Syntax refers to the sequence of syllables in many successive motifs.
Motif analysis. We quantified how well pupils had copied the motif of their tutor using a similarity score and an accuracy score obtained in SAP from ten asymmetric pairwise comparisons of the pupil's typical motif with the tutor motif. In asymmetric comparisons, the most similar sound elements of two motifs are compared, independent of their position within a motif. The smallest unit of comparison are 9.26-ms–long sound intervals (FFT windows). Each interval is characterized by measures for five acoustic features: pitch, FM, amplitude modulation (AM), Wiener entropy, and PG. SAP calculates the Euclidean distance between all interval pairs from two songs, over the course of the motif, and determines a p-value for each interval pair. This p-value is based on p-value estimates derived from the cumulative distribution of Euclidean distances across 250,000 sound-interval pairs, obtained from 25 random pairs of zebra finch songs. Neighboring intervals that pass the p-threshold value (p = 0.1 in this study) form larger similarity segments (70 ms). The amount of sound from the tutor's motif that was included into the similarity segments represents the similarity score; it thus reflects how much of the tutor's song material was found in the pupil's motif.
To measure how accurately pupils copied the sound elements of the tutor motif, we used the accuracy score from SAP. The accuracy score is computed locally, across short (9 ms) FFT windows and indicates how well the sound matched to the sound in the tutor song. SAP calculates an average accuracy value of the motif by averaging all accuracy values across the similarity segments.
Syllable analysis—manual counting of imitated syllables. For manual counting of imitated syllable types, two individuals who were blind to treatment counted all syllables that matched a tutor syllable by visual inspection of sonograms. Their interobserver reliability was 80%.
Syllable analysis—syllable acoustic features. We extracted the mean pitch, mean FM, mean entropy, and mean PG, as well as mean duration from 25 renditions of each syllable. To compare the similarity of individual spectral features between pupil and tutor syllables, we subtracted each mean feature value of each tutor syllable from the mean feature value of the corresponding pupil syllable. Next, we normalized the absolute differences between the values of tutor and pupil syllables to the values of the tutor syllable to obtain the difference of a pupil syllable in a given feature from the tutor syllable in percent.
To describe the variability of syllable duration between different renditions, we calculated the coefficient of variation of duration values among 25 renditions of each syllable.
Syllable analysis—syllable identity score. We quantified the acoustic similarity between different syllables using symmetric comparisons to obtain syllable identity scores. In contrast to asymmetric comparison, no similarity segments are identified during symmetric comparisons. Instead, the FFT windows are compared sequentially from beginning to the end of the two sounds. Thus, similarity reflects how many sound intervals were above p-value, and accuracy indicates the average (1 − p-value). To comprehensively capture the acoustic similarity between syllables in a single measure we used the product of similarity and accuracy to obtain the syllable identity score. As for the motif analysis the p-threshold value was set to p = 0.1.
To quantify how accurately pupils learned individual syllables, we performed ten symmetric comparisons of each pupil syllable with its corresponding tutor syllable. To assess how variable the same pupil performed a particular syllable in multiple renditions of his motif, we compared 20 renditions of each syllable, two at a time. Because minute temporal shifting of FFT windows is allowed in symmetric comparisons (10 ms in this study), the more variable duration of syllables in FoxP2 knockdown animals did not bias the identity score. The syllable identity score rather reflects spectral differences between syllables.
Syntax analysis. For each pupil, we manually annotated sequences of 300 user-defined syllables with the positions in their respective motifs. That is, each syllable of a motif was given a unique integer. Based on these data, we computed the Markov chain for each pupil, i.e., all transition probabilities between syllables. To measure the stereotypy of a motif, we calculated for each syllable the entropy of its transition distribution [52]. Because motif duration differed between birds, these entropy values were rescaled by the maximal possible entropy for each given motif duration. The entropy score for a pupil was then represented by the average of these fractions of maximal entropy over all syllables. Based on this entropy measure, we generated a sequence consistency score (1 − entropy measure), which reflects song stereotypy. An entropy score of 0 indicates random syllable order, whereas a score of 1 reflects a fixed syllable order.
Analysis of song development. To determine tutor similarity and vocal variability during plastic song and towards the end of the learning phase, we analyzed songs recorded on PHD65, PHD80, and PHD90–93 (PHD ± 1 d; in one control pupil, recordings were only available from PHD75 instead of PHD80). First, all sound files from one day were segmented into sounds in the feature batch mode of SAP. Here, the pupils' vocalization is separated from nonvocalization background using two thresholds (Wiener entropy and amplitude). The thresholds were adjusted for each pupil individually to obtain an optimal segmentation. We validated the segmentation for each pupil by visual inspection of the segments and confirmed that segments correspond to syllables. Next, all segments from a given day (between 1,000 and 3,000 segments) were automatically compared to the tutor motif. That is, in each comparison, SAP identifies the best possible match to the tutor motif for each segment. Of all segments analyzed from PHD65, PHD80, and PHD90, 11.0% ± 0.9% were less similar to the tutor model than two random zebra finch sounds are to each other, and thus did not receive any accuracy value in SAP. These sounds were found to represent cage noise, mostly. There were no differences between the amount of sounds excluded between knockdown and control pupils for any of the ages (two-tailed Mann-Whitney U test, p > 0.9 for PHD65; p > 0.8 for PHD80; p > 0.7 for PHD90).
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank) accession numbers for the genes and gene products discussed in this paper are FoxP1 (AY549152), FoxP2 isoform I (AY549148), FoxP2 isoform IV (AY549151), Hmbs (NM_013551), and Pfkp (NM_019703).
The Online Mendelian Inheritance in Man (OMIM; http://www.ncbi.nlm.nih.gov/sites/entrez?db=OMIM) accession number for FOXP2 is 605317.
|
10.1371/journal.pntd.0000409 | Inter-Cellular Variation in DNA Content of Entamoeba histolytica Originates from Temporal and Spatial Uncoupling of Cytokinesis from the Nuclear Cycle | Accumulation of multiple copies of the genome in a single nucleus and several nuclei in a single cell has previously been noted in Entamoeba histolytica, contributing to the genetic heterogeneity of this unicellular eukaryote. In this study, we demonstrate that this genetic heterogeneity is an inherent feature of the cell cycle of this organism. Chromosome segregation occurs on a variety of novel microtubular assemblies including multi-polar spindles. Cytokinesis in E. histolytica is completed by the mechanical severing of a thin cytoplasmic bridge, either independently or with the help of neighboring cells. Importantly, cytokinesis is uncoupled from the nuclear division cycle, both temporally and spatially, leading to the formation of unequal daughter cells. Sorting of euploid and polyploid cells showed that each of these sub-populations acquired heterogeneous DNA content upon further growth. Our study conclusively demonstrates that genetic heterogeneity originates from the unique mode of cell division events in this protist.
| Proliferating eukaryotic cells regulate their DNA synthesis, chromosome segregation, and cell division with great precision so that daughter cells are genetically identical. Our study demonstrates that in proliferating cells of the protist pathogen Entamoeba histolytica re-duplication of DNA followed by segregation on atypical and diverse microtubular structures is frequently observed. In this parasite, cell division is erratic, so that each daughter cell may contain one or more nuclei and sometimes no nuclei. This uncoupling of cell cycle events and survival of daughter cells with unequal DNA contents leads to genetic heterogeneity in E. histolytica. Our study highlights the inherent plasticity of the Entamoeba genome and the ability of this protist to survive in the absence of strict regulatory mechanisms that are a hallmark of the eukaryotic cell cycle.
| Eukaryotic cells undergo two major events during proliferation- the nuclear cycle, when the whole genome is duplicated and segregated equally into two nuclei and cytokinesis, the physical separation of a mother cell into two daughter cells. Sequential progression of events during the cell cycle is enforced by checkpoint proteins [1] that also ensure the spatial and temporal coordination of mitosis with cytokinesis [2]–[4]. While this paradigm is true for most organisms that have been studied, it is becoming increasingly clear that diverse groups of eukaryotes, including plants, ciliates, Drosophila, and protists, show striking differences in regulating the transmission of genetic information during proliferation [5]. Many of these organisms tolerate large variations in DNA content during different life cycle stages such that typical cell cycle checkpoints are altered. Entamoeba histolytica is one such protist parasite that proliferates in the intestinal lumen of human beings often causing severe disease in the host. Indeed this parasite is responsible for millions of cases of dysentery and liver abscess world wide especially in developing countries [6].
Axenically grown E. histolytica cells exhibit variations in DNA content, which may vary several folds within a population [7]. Multiple genome complements may be present in a single nucleus or distributed over multiple nuclei in a single cell [8],[9]. Re-duplication of the genome and de-linking of S-phase from cytokinesis were identified as two of the reasons for the generation of polyploidy in these trophozoites [8]. A significant number of E. histolytica trophozoites with multiple nuclei were observed in infected intestinal tissue suggesting this is an intrinsic property of this parasite and not due to in vitro culture conditions [10]. Remarkably, under different growth conditions such as switching from xenic to axenic growth leads to a significant increase in the nuclear DNA content of E. histolytica trophozoites [10]. Similar variation in nuclear DNA content was also observed in the related reptilian parasite E. invadens during conversion of cysts to trophozoites and vice versa [10]. These observations clearly indicate the inherent plasticity of the Entamoeba genome [10] and the ability of this protist to survive in the absence of strict regulatory mechanisms that are a hallmark of the eukaryotic cell cycle.
Most eukaryotes segregate their genomes once duplication is complete. While more than 2n genome content accumulates in some E. histolytica cells due to multiple rounds of DNA duplication, it has not been clear how polyploid amoebae partition their genomes. Bipolar microtubular spindles are not frequently visible in E. histolytica [11]. Rather, several electron microscopic studies have reported atypical organization of microtubules (MTs) during nuclear division [12]–[14]. Thus, the possible cell-cycle mechanism underlying the dynamic variation in DNA content of E. histolytica cells during their axenic growth include- a) multiple cycles of replication of the genome without nuclear or cell division, b) nuclear division without cytokinesis and c) polyploidization from over-replication of the genome followed by segregation on multipolar mitotic spindles.
In this study, we demonstrate: a) the formation of atypical microtubular structures during genome segregation; b) multiple microtubule organizing centers (MTOCs) and multi-polar spindles are formed in polyploid nuclei; c) cell division is spatially and temporally uncoupled from the nuclear cycle; d) cell division can be asymmetric, thereby producing either uni-nucleate, multi-nucleate or anucleate daughter cells. In summary, asynchrony between nuclear division and cytokinesis in a fraction of the population combined with segregation of the polyploid genome on multi-polar spindles accounts for the extreme variation of genomic DNA content in individual E. histolytica cells in axenic culture.
E. histolytica HM-1:IMSS trophozoites were maintained axenically and routinely sub-cultured every 48–72 h in TYI-S-33 medium containing 10% adult bovine serum at 37°C [15].
E. histolytica HM-1:IMSS cells were sub-cultured every 24 h for 3–4 days followed by serum starvation [8]. Adult bovine serum was added (10%) after 12–13 h of serum starvation. Cells were then withdrawn at different times as indicated and fixed.
Ethanol fixed cells were stained with 4′-6-Diamidino-2-phenylindole (DAPI, 0.1 µg/ml, Sigma) for 10 min, washed once with 1× PBS and then scanned for DNA content of individual nuclei [10] under a 40× oil objective (numerical aperture 1.3) of a Zeiss Axiovert 200 M fluorescence microscope fitted with the MetaCyte scanning cytometer (Zeiss, Germany). A minimum of 2000 nuclei was scanned for each sample and analyzed by Metafer4 software (Zeiss, Germany). The DAPI fluorescence values (x-axis) were represented as histograms. The scan yields varying numbers of nuclei with different fluorescence values. The number of nuclei on the y-axis was represented as a fraction of the highest number of nuclei obtained in any one sub-class of each scan.
E. histolytica HM-1:IMSS cells were grown on coverslips in 24-well plates at 37°C, fixed directly with warm 3.7% formaldehyde for 15 min and permeabilized with 0.1% Triton X-100 for 10 min. Fixed cells were stained with polyclonal anti-Eh β-tubulin antibody [11] followed by tetramethyl rhodamine isothiocyanate (TRITC) conjugated anti-rabbit secondary antibody (1∶200; Jackson Laboratories, USA). For visualization of actin filaments, cells were incubated with Alexa Fluor 488 conjugated phalloidin for 30 min (Molecular Probes, Invitrogen, USA). Images were acquired with (i) 63× Plan-Apochromat 1.4 oil differential interference contrast (DIC) objective (numerical aperture 1.4) in a Zeiss LSM 510 Meta confocal microscope equipped with a 488 nm argon laser and a 543-nm He/Ne laser and were analyzed with the LSM Meta 510 software package (Zeiss, Germany) or (ii) a 40× oil objective (numerical aperture 1.3) in an Axiovert 200 M fluorescence microscope using Z-stacking and analyzed by deconvolution (Axiovision v4.6). DNA was stained with DAPI (0.2 µg/ml, Sigma) for 30 min.
E. histolytica HM-1:IMSS trophozoites were plated on 35-mm plastic culture dishes filled with growth medium at 37°C. After the cells adhered, the medium was replaced with fresh growth medium. The dish was kept inside an incubator (Tempcontrol 37-2 digital, Zeiss, Germany) at 37°C and under 5% CO2 flow system (PeCon GmbH, Erbach, Germany) which was fitted to the Axiovert 200 M fluorescence microscope (Zeiss, Germany). Cells were visualized under a 20× phase contrast objective. The time-lapse images were captured with 1 sec interval for the indicated time, then analyzed and further processed by Axiovision v4.6 software (Zeiss, Germany).
Asynchronously growing E. histolytica HM-1:IMSS cells (1×108 cells) were harvested 48 h after sub-culture, washed with 1× PBS and incubated with 1 µM DNA binding dye- Vybrant DyeCycle Orange for 1 h (Molecular Probes, Invitrogen, USA). Cells were passed through a 40 µm nylon mesh (Becton Dickinson, USA) to remove aggregates and debris before sorting the cells. Vybrant DyeCycle Orange stained cells were excited at 488 nm in a flow cytometer (FACSAria, Becton Dickinson, USA) and emission was measured through a 585/42 Band Pass filter for analyzing the DNA content. On the basis of DNA content analysis (FACSDIVA 6.0 software, Becton Dickinson, USA), cells were demarcated in four different electronic gates and sorted at 4°C. The sorted cells were washed with 1× PBS and resuspended in 2 ml TYI-S-33 medium. Half the cells were inoculated into growth medium and harvested after 3 days. The remaining cells were fixed in 70% ethanol for analysis of the nuclear DNA content and number of nuclei in each cell.
DNA synthesis of E. histolytica is arrested when cells are incubated in serum free media for at least 12 h, followed by re-initiation within 2 h after addition of serum [7],[8]. Endo-replication of the genome in the population can be seen following this synchronization protocol [8]. We have now examined the fates of individual nuclei following this synchronization. Progression of S-phase was monitored by estimating the nuclear DNA content in DAPI stained cells. Nuclear division was estimated from the increase in bi-nucleated cells while cell division was estimated from counting cell numbers. Our data show that after 12 h of serum starvation, E. histolytica nuclei contain heterogeneous amounts of DNA. 1 h after addition of serum the nuclei show a dramatic homogenization and reduction of DNA content (Figure 1A). This early homogenization likely results from nuclear division in polyploid nuclei. An earlier report identified two copies of selected loci using fluorescent in situ hybridization and estimated amoeba to be diploid [16] while other studies have demonstrated that the DNA content can vary several fold in proliferating amoebae so that single cells may contain up to 10n or 12n [7],[10]. We assigned 1n genome content to the lowest nuclear DNA content after homogenization. Compared to nuclei 1 h after the addition of serum (1n–2n), the average DNA content of E. histolytica nuclei increased 2 fold in 80–90% of the nuclei between 2 h and 8 h after addition of serum (Figure 1A). Between 8 h and 10 h after addition of serum, the nuclear DNA content again showed reduction and homogenization to a level similar to that observed at 1 h after serum addition. This suggests completion of chromosome segregation and nuclear division within 10 h. The homogenous euploid nuclear DNA content at 1 h and 10 h after serum addition marks the temporal boundaries of a single nuclear cycle. The nuclear DNA content increased again after 10 h suggesting that another cycle had been initiated. In each of these cycles, 10–20% of the nuclei accumulated greater than 4n DNA content (Figure 1A) suggesting that some nuclei undergo endo-replication without nuclear division. This polyploid population was absent at 1 h and 10 h after serum addition, suggesting that either chromosome segregation occurs more than once in a single nuclear cycle in the polyploid nuclei or the polyploid nuclei can segregate multiple copies of the genome simultaneously.
In order to assess whether chromosome segregation was coupled with nuclear division, we scored the number of microtubular assemblies and bi-nucleated cells in conjunction with changes in DNA content. Earlier studies have demonstrated that microtubules were mostly nuclear [11],[12],[17]. After 12 h of serum starvation, tubulin was dispersed in the nucleus without any obvious structure in most cells. Microtubular assemblies or structures began to appear 2 h after serum addition, continued to increase in number up to 8 h, and then decreased markedly at 10 h (Figure 1B). Bi-nucleated cells were highest (∼20%) at 10 h after serum addition in these cells (Figure 1B). Thus for 20% of the cells nuclear DNA replication was immediately followed by segregation of the chromosomes into daughter nuclei.
Importantly, cell numbers increased gradually rather than in a step-wise fashion coinciding with increase in nuclear number (Figure S1A). While cell density affected the rate of increase in cell number (Figure S1B), the temporal progression of the nuclear division cycle was independent of these factors. This suggests that although nuclear division is coupled to DNA synthesis and chromosome segregation, cell division is random and not obligately linked to nuclear division in these synchronized E. histolytica cells.
During the course of a mitotic cycle, we observed different MT assemblies and structures in the amoeba nucleus. Commonly, anti-Eh β-tubulin antibody showed a diffuse nuclear stain and MT structures were not visible in most nuclei. In some nuclei, a single pole-like body was observed at the center (Figure 2A). Confocal microscopic analyses clearly showed a central ring-like arrangement of MTs with radially disposed short MT fibers (Figure 2A1). It has been shown that Eh γ-tubulin formed a similar ring-like arrangement at the center of E. histolytica nuclei thus defining the pole-like body as ‘MTOC’ in these cells [17]. During progression of the nuclear cycle we identified a variety of MT spindle-like assemblies that likely originated from the MTOC and radial MTs described above. The arrangement of chromosomes on these MT structures suggested that these structures were intermediates formed during genome segregation. Figure 2B–E (corresponding confocal images Figure 2D1 and 2E1) show MT fibres emanating from a single pole that may be an extension of the central pole seen in Figure 2A. Chromosomal DNA appears to have segregated at an early stage after attachment to the central pole and then been subsequently pushed apart at one end of the uni-directionally extending MT fibers (Figure 2C–E). These structures were observed in 25–30% of the spindle intermediates.
We also observed several non-conventional MT structures - a) with DNA at the two ends of bi-directionally extending MT fibers (Figure 2F–H) and b) with DNA bound at one end of several radiating MT fibers that bundled together at the other end (Figure 2I–J and 2I1). Segregation appears to be completed on extended MT structures where the elongated MTs have disappeared from the centre (Figure 2K–L and 2K1). Bipolar spindle like structures were also observed but at a low frequency (3–5% of the MT structures) compared to the monopolar intermediates. Chromosomal DNA was distributed longitudinally over the bipolar spindle or at the two poles (Figure 2M–N and 2M1).
The most frequent structures visible at 2 h–4 h after serum addition were the intermediates shown in Figure 2I–J (50–55% of the MT structures). Based on the different MT structures observed during progression of the nuclear cycle, it is likely that genome segregation occurs either by extension of a radial monopolar assembly (Figure 2C–E) or by lateral separation of MT fibers that are joined at a distal end (Figure 2I and 2J). It is unclear whether the MT structures shown in Figure 2F–H originated from monopolar assemblies or duplication of a MTOC. Instead of chromosomal segregation due to pole-ward contraction of MT fibers, the MT structures in amoeba nuclei suggest either a) separation of chromosomal DNA on the growing end of MTs, b) lateral separation of duplicated chromosomes and MTs, or c) longitudinal movement of chromosomes on elongated MTs.
Importantly, several nuclei with high DNA content were seen that contained two, three or four MTOCs (Figure 3A). Bipolar spindle intermediates likely originated from duplication of the central pole (Figure 3A-a, b). Interestingly, tri-polar, tetra-polar and multi-polar spindles were also seen in several cells (Figure 3B). It is possible that these multi-polar spindles originated from the multiple MTOCs in polyploid nuclei and were used to segregate multiple copies of the genome simultaneously. This could be one of the mechanisms for the segregation of multiple genome copies in polyploid nuclei during a single nuclear cycle.
In summary, E. histolytica uses novel MT structures for chromosome segregation mostly originating from a single pole. Multi-polar spindles are also observed and may account for the rapid generation of DNA content heterogeneity from polyploid cells. Another unique feature is the use of multiple modes of segregation, as suggested by the diverse MT structural intermediates and inter-cellular variation in DNA content.
No significant increase in the number of dividing cells at any time after completion of the nuclear cycle was observed in serum synchronized E. histolytica (data not shown). Therefore, based on earlier reports [18], we provided fresh growth medium to the culture to induce an increase in division events. Using real-time microscopy we identified at least two modes of cytokinesis. Cytoplasmic constriction was initiated at random sites generating two dividing parts of a cell that pulled away from each other, forming a thin cytoplasmic bridge that was eventually severed (Video S1 and Figure S2A). In many dividing cells, the complete scission of the cytoplasmic bridge utilizes the assistance of helper cells. In such cases, the helper cells, migrated through the connecting cytoplasmic bridge and ensured the mechanical rupture of the ‘connector’ (Video S2 and Figure S2B). From five independent experiments, we recorded 40 cell division events by real-time microscopy, after cytoplasmic constriction was initiated. Our results showed that one third (14 out of 40) of E. histolytica cell division events were assisted by helper cells while 45% (18 out of 40) involved independent mechanical severance of the cytoplasmic bridge. We also discovered that 20% (8 out of 40) of the cell division events in E. histolytica were not completed. In these cases, the connecting cytoplasmic bridges became unusually long and thin but were not finally severed. Instead, the cytoplasmic bridge contracted and the two ‘dividing halves’ fused together without division (Video S3).
Thus, besides the heterogeneity in chromosome segregation, and the lack of synchrony between nuclear division and cytokinesis, heterogeneity in modes of cytokinesis and frequent failure of completion of cytokinesis even after initiation, contribute to the genetic heterogeneity that arises in a population of E. histolytica.
In most organisms, the precise location of the cytokinesis apparatus between the two daughter nuclei is observed. We therefore, examined where the cytokinesis furrow was located in amoeba cells relative to locations of the daughter nuclei and how the reorganization of cytoplasmic actin assembly takes place in dividing cells. Longitudinal actin fibers perpendicular to the constriction site along with tiny F-actin patches were seen in cells likely to divide (Figure 4). Importantly, dividing cells may contain one or more nuclei in each ‘daughter’. Most of the events highlighted the randomness of cytokinesis site selection relative to the location of the daughter nuclei (Figure 4). In some cases, the nuclei may move away from their original position after initiation of cell division so that both nuclei are on the same side of the division site and thus ‘anucleate’ cells are generated. Anucleate cells may also form after erratic division of uni-nucleated cells (Figure 5). This inefficient and irregular cell division is likely responsible for the formation of heterogeneous daughter cells with variable numbers of nuclei in a population of E. histolytica. Thus, spatial uncoupling of cell division with nuclear division adds to the genetic heterogeneity in E. histolytica.
In order to determine if the difference in DNA content was a stable property of both euploid and polyploid cells, we separated cells on the basis of their DNA content and followed the DNA content of the progeny. Log phase E. histolytica trophozoites were stained with the vital dye Vybrant orange and sub-populations of live amoebae with varying DNA contents were isolated using a fluorescence activated cell sorter. The fluorescence values representing cellular DNA content profile of unsorted cells showed a broad distribution (Figure 6A, a). Electronic gates (P1–P4) were set (Figure 6A, a) and cells were separated according to differences in DNA content. Figure 6A, b-e shows the cellular DNA content profile of the different sub-populations that were recovered after sorting. The sub-populations P1–P4 were analyzed for multi-nucleated cells and nuclear DNA content and compared with the unsorted cells. Table 1 shows that while the unsorted cells consisted of approximately 92% uni-nucleated cells and 8% bi- and multi-nucleated cells, sorted P1 and P2 cells were mostly uni-nucleated (92–96%) with only a small percentage of bi- and multi-nucleated cells (4–8%). On the other hand, P3 and P4 cells contained a significantly higher number of bi- and multi-nucleated cells (13–20%).
A comparison of the nuclear DNA content in the unsorted cells showed a broad distribution of nuclear DNA content from 1n to 6n while about 15–20% nuclei contained greater than 6n DNA content (Figure 6B, left). P1 cells showed 1n–4n nuclear DNA content, P2 cells showed 4n–14n nuclear DNA content while P3 and P4 cells showed 3n–14n nuclear DNA content. (Figure 6B, right). To investigate whether the isolated sub-populations of P1–P4 cells would retain their DNA content and nuclear number phenotype after growth, the sorted cells were incubated in fresh medium at 37°C. Following an initial lag of 24–36 h (possibly due to the physiological stress induced by cell sorting), the cell numbers subsequently increased in all the tubes. After 72 h all the sorted sub-populations had undergone at least 2 doublings (final cell density ∼1–3×105 cells per ml in each tube). It was observed that P1 cells had a higher percentage of bi-and multi-nucleated cells while P4 cells showed a homogenous population of uni-nucleated cells after growth (Table 1). Changes in the number of bi- and multi-nucleated cells were observed in P2 and P3 cells also. After growth the P2–P4 populations showed a reduction in the number of 10n–14n nuclei (Figure 6C). Therefore E. histolytica is programmed to continuously generate genetic heterogeneity in both nuclear number per cell and DNA content per nucleus, and this program is equally active in the polyploid and the euploid cells.
In this study, we have discovered that genetic heterogeneity in E. histolytica cells results from multiple modes of genome segregation during nuclear cycle, asynchrony between nuclear division and cytokinesis along with variation in nuclear DNA content.
Although recent studies have reported the infrequent occurrence of bipolar MT spindles [11],[19] in E. histolytica cells, metaphase-like equatorial alignment of condensed chromosomes or kinetochores could not be identified in amoeba [12],[14] while anaphase and telophase were identified on the basis of nuclear shape [13]. Indirect immuno-fluorescence showed nuclear microtubular assemblies with fibers radiating from a central region in most E. histolytica cells [17].
This study shows that MT structures in E. histolytica include monopolar, bipolar and multi-polar spindles for the segregation of chromosomal DNA (Figures 2 and 3). On the basis of our findings, we propose a model (Figure 7) to explain possible modes of chromosome partitioning on these different MT structures that were identified during the nuclear cycle in axenically grown E. histolytica cells. It is conceivable that some of the proposed intermediate or observed MT structures are abortive. It is important to note that a large number of proteins known to regulate chromosome segregation and spindle assembly were not encoded in the amoeba genome [20]. Absence of these regulatory functions may lead to the formation of unconventional MT structures.
Cytokinetic processes in E. histolytica (this study) and E. invadens [21] are similar. Both organisms depend upon mechanical rupture of a cytoplasmic bridge that may occur independently or utilizes assistance from helper cells. Motility and the consequent mechanical force driven by actin polymerization in a polymorphic cell may have been one of the earliest modes of cell division [22]. Assistance from helper cells likely results from ‘altruistic’ behaviour of genetically related cells in a clonal population. The cytokinetic process is imprecise and irregular in Entamoeba, unlike the finely regulated cell division in bacteria, yeasts and higher eukaryotes. While endogenous Entamoeba proteins may affect the rate of cell division as deduced from increased multi-nucleation in different mutants [23]–[26], this event appears to be largely controlled by extra-cellular factors. Random selection of cell division sites coupled to a poorly controlled mechanical separation can lead to the formation of anucleate and multi-nucleated daughter cells. Similar cytokinetic events are observed in the social amoeba D. discoideum which shows variations in its mode of cell division. Multi-nucleated Myosin II-null D. discoideum cells, when allowed to adhere on substrate, undergo a cell cycle-uncoupled, inefficient cytofission [27],[28].
The nuclear cycle was clearly defined between initiation of DNA synthesis and completion of nuclear division. That a significant number of nuclei continued accumulation of DNA after doubling their DNA content emphasizes the absence of stringent regulatory mechanisms preventing re-replication that are seen in other eukaryotes. Heterogeneity of MT structures also suggests a lack of strict control of genome segregation mechanisms. It may be noted that even after serum synchronization, the maximal number of nuclei with MT structures and bi-nucleated cells were only around 20%. Treatment of E. histolytica cells with aphidicolin or taxol did not lead to a significant increase of nuclei in different mitotic stages (data not shown). Thus the entire nuclear cycle appears to be poorly controlled, resulting in leaky phenotypes that are apparent in polyploid nuclei. In spite of leaky regulatory mechanisms, S-phase, genome segregation and nuclear division are temporally coupled. Multiple MTOCs and multi-polar MT spindles possibly facilitate a return towards euploidy for polyploid nuclei.
Strict regulation of the cell division cycle has been considered to be the very basis of survival for eukaryotes. Indeed an unregulated cell cycle leads to growth arrest, aneuploidy and tumorigenesis in yeasts and higher eukaryotes. While prokaryotes use overlapping parallel processes of DNA synthesis, segregation and cell division, eukaryotes ensure completion of the preceding stage before initiation of the next, with tight surveillance to ensure the same. Clearly Entamoeba and few other organisms can bypass these surveillance mechanisms and continue their life cycles. How does Entamoeba preserve its genetic composition in the absence of cell cycle regulation? The answer possibly lies in its survival as a parasite, dependant on its surroundings, where plasticity and heterogeneity are favored over precision and homogeneity.
|
10.1371/journal.pbio.1002365 | Interactome Mapping Reveals the Evolutionary History of the Nuclear Pore Complex | The nuclear pore complex (NPC) is responsible for nucleocytoplasmic transport and constitutes a hub for control of gene expression. The components of NPCs from several eukaryotic lineages have been determined, but only the yeast and vertebrate NPCs have been extensively characterized at the quaternary level. Significantly, recent evidence indicates that compositional similarity does not necessarily correspond to homologous architecture between NPCs from different taxa. To address this, we describe the interactome of the trypanosome NPC, a representative, highly divergent eukaryote. We identify numerous new NPC components and report an exhaustive interactome, allowing assignment of trypanosome nucleoporins to discrete NPC substructures. Remarkably, despite retaining similar protein composition, there are exceptional architectural dissimilarities between opisthokont (yeast and vertebrates) and excavate (trypanosomes) NPCs. Whilst elements of the inner core are conserved, numerous peripheral structures are highly divergent, perhaps reflecting requirements to interface with divergent nuclear and cytoplasmic functions. Moreover, the trypanosome NPC has almost complete nucleocytoplasmic symmetry, in contrast to the opisthokont NPC; this may reflect divergence in RNA export processes at the NPC cytoplasmic face, as we find evidence supporting Ran-dependent mRNA export in trypanosomes, similar to protein transport. We propose a model of stepwise acquisition of nucleocytoplasmic mechanistic complexity and demonstrate that detailed dissection of macromolecular complexes provides fuller understanding of evolutionary processes.
| Much of the core architecture of the eukaryotic cell was established over one billion years ago. Significantly, many cellular systems possess lineage-specific features, and architectural and compositional variation of complexes and pathways that are likely keyed to specific functional adaptations. The nuclear pore complex (NPC) contributes to many processes, including nucleocytoplasmic transport, interactions with the nuclear lamina, and mRNA processing. We exploited trypanosome parasites to investigate NPC evolution and conservation at the level of protein–protein interactions and composition. We unambiguously assigned NPC components to specific substructures and found that the NPC structural scaffold is generally conserved, albeit with lineage-specific elements. However, there is significant variation in pore membrane proteins and an absence of critical components involved in mRNA export in fungi and animals (opisthokonts). This is reflected by the completely symmetric localization of all trypanosome nucleoporins, with the exception of the nuclear basket. This architecture is highly distinct from opisthokonts. We also identify features that suggest a Ran-dependent system for mRNA export in trypanosomes, a system that may presage distinct mechanisms of protein and mRNA transport in animals and fungi. Our study highlights that shared composition of macromolecular assemblies does not necessarily equate to shared architecture. Identification of lineage-specific features within the trypanosome NPC significantly advances our understanding of mechanisms of nuclear transport, gene expression, and evolution of the nucleus.
| In order to uncover the origins of eukaryotes, we must understand how their defining organelle, the nucleus, and its delineating nuclear envelope (NE) arose. The NE provides a barrier that defines the nucleoplasm and cytoplasm, and this discrimination represents a major evolutionary transition [1]. The sole mediators of macromolecular exchange between the nucleoplasm and cytoplasm are nuclear pore complexes (NPCs) [2]. Each NPC is a ~50 MDa, cylindrical, and octagonally symmetric structure comprised of nearly 500 proteins, these being multiple copies of ~30 different nucleoporins (Nups) [3–8]. There are three major Nup classes: pore membrane proteins (Poms), core scaffold Nups, and FG-repeat Nups. Poms contain trans-membrane domains (TM) that serve to anchor the NPC to the NE, whilst the core scaffold Nups are major structural components and also interact with the NE and Poms. The core scaffold consists of two inner rings sandwiched between two outer rings and is comprised of three groups of proteins containing only two major folds: α-solenoids and β-propellers, or an N-terminal β-propeller followed by an α-solenoid [9]. Interestingly, vesicle coat proteins, including clathrin/adaptin, COPI, and COPII, share architectural characteristics with the components of the outer ring Nups of the NPC, suggesting a common ancestry between the endomembrane trafficking system and the NPC; this is known as the protocoatomer hypothesis [9–12]. In addition to providing the structural core of the NPC, scaffold Nups provide a platform for anchoring FG-Nups, natively disordered proteins characterized by domains enriched in phenylalanine-glycine (FG) repeats and responsible for the selective permeability barrier to nucleocytoplasmic transport. In animals and fungi, a large subset of FG-Nups have a biased distribution across the NPC, with ~30% predominantly at either the nucleoplasmic or cytoplasmic face of the NPC [6], suggesting that this asymmetry is important for certain aspects of NPC function, despite being apparently dispensable for the basic mechanisms of transport [13].
Ions, metabolites, and proteins <40 kDa can freely diffuse through the NPC between the cytoplasm and nucleoplasm [14,15]. Larger cargos require nuclear localization signals or nuclear export sequences to mediate transport through the NPC, via interactions with soluble transport factors or karyopherins, which themselves interact with the FG-Nups [16,17]. Directionality is provided by a RanGTP/RanGDP gradient, with RanGTP the predominant form in the nucleus and RanGDP in the cytoplasm reviewed in [18]. However, bulk mRNA export is an ATP-dependent and Ran-independent process, unlike protein transport, with directionality provided by a DEAD-box ATP helicase attached to the conserved cytoplasmic Nup82 (yeast) or Nup88 (vertebrates) complex, which remodels ribonucleoproteins (mRNPs) as they exit the nucleus [19–25].
Our current understanding of how nucleocytoplasmic transport works stems from decades of work in yeast and vertebrates, both members of the Opisthokonta, one of five or six major supergroups of the eukaryotic lineage (Fig 1A) [26]. NPC components have been catalogued for yeast, several vertebrates, the plant Arabidopsis thaliana (Archaeplastida) [7,8], and the trypanosome Trypanosoma brucei (Excavata), by us [27]. There is remarkably low sequence similarity between trypanosome and opisthokont Nups, with only five being easily identifiable by sequence alignments [27]. Despite this low sequence similarity, trypanosome NPC components share, to a remarkable level, domain organization and composition with opisthokont Nups. This suggests that most components of the NPC are evolutionarily conserved, albeit with a few exceptions, including the metazoan-specific Nup358 (Ran-binding protein 2), compositional variation in Poms, and the presence or absence of two or three β-propeller proteins in the outer ring of the core scaffold, together with the duplications of Nups in yeast, such as Nup157/170, Nup53/59, or Nup100/116/145N as homologs of the vertebrate Nup155, Nup35, and Nup98, respectively [5,6,27–29]. Indeed, although comparative genomics does not allow full reconstruction of NPC composition for most taxa, data are consistent with overall broad conservation [30]. However, only the yeast NPC has been comprehensively characterized to the architectural level, with partial characterization for vertebrates [3,4,31]. There is, therefore, a significant gap in our understanding of NPC structure and function, as accumulating data suggests significant architectural divergence between different taxa. For example, each vertebrate outer ring is comprised of two reticulated rings, but is a single ring in yeast [4,31,32]. Interestingly, there is a major role for Nup358 in the formation and maintenance of the reticulated cytoplasmic outer ring in metazoa [33]. In trypanosomes, both major components of the nuclear basket, TbNup92 and TbNup110, are highly divergent from the analogous proteins of plants, yeast, vertebrates, and flies [34–37].
We have previously identified and green fluorescent protein (GFP)-tagged 22 T. brucei Nups (TbNups), to which we assigned putative yeast and human orthologs based on secondary structure prediction and molecular weight [27,39]. Opisthokont and plant NPCs contain about 30 proteins [5–7], suggesting that several TbNups had yet to be identified, and so the absence of a complete NPC composition precluded functional predictions. Further, the arrangement of subunits was unknown. Affinity capture/mass spectrometry (MS) interactomics directly addresses these issues by providing high-resolution mapping and exhaustive analysis of quaternary structure and subunit composition. With this strategy, combined with fluorescence and immunoelectron microscopy (iEM), we have characterized the architecture of the trypanosome NPC, uncovering distinct architectural features that provide novel insights into the function and evolution of this central component of eukaryotic cells.
Each described trypanosome Nup was tagged in situ at one allele with GFP [27]. All transgenic parasite lines continued to proliferate normally, indicating that the tag has little impact on cell viability and, likely, NPC function. Tagged cells were expanded, rapidly frozen, and then cryomilled (Methods) [40]. The frozen powder was thawed into various buffers to determine optimum conditions for the isolation of the GFP-tagged Nup together with associating proteins. Complexes were captured using polyclonal anti-GFP antibodies conjugated to magnetic beads. Systematic testing of buffers, detergents, salts, and co-solvents allowed us to affinity purify stable NPC subcomplexes, to preserve interactions between NPC subcomplexes, and also to isolate the entire NPC (Figs 1, 2 and 3; S1 Table; see Fig 3 for a comparison of trypanosome and yeast/human Nup orthologs) [41]. By iteratively repeating these affinity capture purifications, we were able to “walk through” the NPC, robustly characterizing a comprehensive NPC interactome, and ensuring that as full a complement of trypanosome Nups as possible was retrieved. Any new candidate component of the NPC was GFP-tagged, had its location confirmed by fluorescence microscopy, and was subsequently used as an affinity handle for further affinity capture experiments (Figs 1 and 2). We also performed iEM on key members of each subcomplex, including FG-Nups that form interactions with multiple subcomplexes (Fig 4). This has allowed us to map the architecture of the trypanosome NPC.
Previous interactomic analyses in yeast used the nuclear basket components as affinity handles under low stringency conditions to capture essentially the entire NPC [44,45]. We therefore used the nuclear basket component TbNup110 [27,35] as an affinity handle to similarly uncover as full a complement of TbNups as possible. As expected, affinity isolation of TbNup110 under low stringency conditions demonstrated extensive interactions with most of the NPC, recovering most known subunits (Figs 1A and 2) [27,35]. Importantly, we recovered five new nucleoporins, designated TbNup41, TbNup65, TbNup76, TbNup119, and TbNup152 (Fig 1A–1C), which were present in our earlier nuclear envelope proteome, but of insufficient sequence similarity to Nups to warrant inclusion in that study [27]. No additional bona fide TbNups were identified from either these or any of our extensive affinity capture experiments. However, we did isolate the lamina protein NUP-1 [42] and several candidate NE proteins, indicating that the procedure has likely saturated identification of NPC components and, indeed, reached beyond it. These data provide robust confirmation that we have likely identified the full complement of trypanosome Nups (see S1 Fig for complete list of identified proteins).
We first wanted to ask: are the main architectural features that have been defined in opisthokont (yeast and vertebrate) NPCs also conserved in the trypanosome NPC? In opisthokonts, among the most conserved such features are the inner ring of the NPC (Fig 1), which in yeast is comprised of two large α-solenoid proteins, ScNup192 and ScNup188, and two β-α structured paralogs, ScNup157 and ScNup170 [9,46–50]. These four proteins interact with the membrane ring that anchors the NPC to the pore membrane, as well as to the ScNic96 complex [4,50,51]. ScNic96 is an evolutionarily conserved and highly abundant α-solenoid protein, which itself is in a complex with three central channel FG-Nups (ScNup57, 49, and Nsp1) in yeast [52]. This entire inner ring arrangement appears very similar in vertebrates [53,54].
TbNup96 can be readily identified as orthologous to ScNic96 in silico, establishing that this protein is conserved. However, sequence comparisons alone do not fully discern the level of conservation of any other putative inner ring components, or, indeed, if there is an inner ring [27]. Thus, we used affinity capture of TbNup96 in order to “walk out” from this protein to uncover its molecular neighborhood. Affinity isolation of TbNup96 co-purified the two largest α-solenoid proteins in the T. brucei NPC (TbNPC)—TbNup225 and TbNup181—as well as three FG-Nups: TbNup62, TbNup53a, and TbNup53b (Figs 2A and 3). Reciprocal affinity isolates of TbNups62, 53a, and 53b co-purified both TbNups225 and 181 as well as each other (Fig 2A). Interestingly, affinity isolation of TbNup225 co-purified all members of the complex except TbNup181 (Fig 2A). Likewise, affinity isolation of TbNup181 co-purified TbNups96, 62, 53a, and 53b, but not TbNup225 (Fig 2A). These data suggest that TbNup225 and TbNup181 do not interact directly, but rather form two distinct subcomplexes, each containing TbNup96, 62, 53a, and 53b; this is further supported by the affinity capture of Nup96, which co-purifies with both TbNups181 and 225 as well as an untagged form. A similar two-complex architecture is present in two fungi, Saccharomyces cerevisiae and Chaetomium thermophilum [4,55]; in C. thermophilum, the orthologous CtNup192 and CtNup188 compete for the same 90 amino acid binding site on an α-helical motif near the N-terminal of Nic96. The vertebrate orthologs of the Nic96 complex appear similarly organized [56]. Thus, the association of the two large α-solenoid proteins with TbNup96 and three FG-Nups indicates that the composition of this complex represents an extremely conserved module (Fig 3), also definitively assigning TbNup62, 53a, and 53b as orthologs of ScNup57, 49, and Nsp1, with which they share clear domain similarities.
TbNup144 and 119 are composed of an N-terminal β-propeller and a C-terminal α-solenoid (β-α) (Figs 1C and 2A) [27]. TbNup144 is evolutionarily well conserved and orthologous to ScNup157/170 and HsNup155 [27]. In contrast, TbNup119 has weak sequence similarity to ScNup170, based on secondary structure prediction modeling with Phyre2 (www.sbg.bio.ic.ac.uk/phyre2/) [57]. Affinity isolation of TbNup144 reveals an interaction with only the α-solenoid TbNup89 (Fig 2A), whereas TbNup119 co-purified with a large number of TbNups (Figs 2A and 3); thus, it appears that TbNup144 links to the outer ring (see below) through interactions with TbNup89, whilst TbNup119 has extensive connections with the core scaffold of the TbNPC (Fig 3).
We performed post embedding (in resin) iEM gold labeling of the NPC using selected GFP-tagged TbNups as described by Krull et al., 2004 (Figs 4, S2 and S3) [58]. The advantage of post resin embedded labeling on whole cells is superior preservation of NPCs, as they are within their correct cellular environment with no manipulation other than high-pressure freezing. Additionally, there is the benefit of being able to label both externally and internally localized GFP-tagged Nups within the context of the NPC. However, good preservation comes at the expense of signal; because the antigens are embedded in plastic resin, only GFP epitopes exposed on the resin surface are accessible for labeling [59–62].
Using the resulting gold particle distributions, we used our prior methods [3,4] and those of Krull et al., 2004 [58] to provide a preliminary estimate for the position of each protein in the NPC (Methods). Consistent with being a conserved central channel FG-Nup, TbNup62 has a symmetric distribution centered tightly around the median plane of the TbNPC, adjacent to the putative central channel. TbNup119 displayed a very similar distribution to TbNup62, consistent with its assignment as another component of the inner ring core scaffold. By contrast, and in confirmation of the relative accuracy of our iEM methodology, the nuclear basket component TbNup110 has a nucleoplasmic localization centered fully ~40 nm from the median plane of the TbNPC (Fig 4). Taken together, these data strongly support that the entire inner ring structure and composition is highly conserved across the eukaryotes.
The NE is an invariant feature of NPCs and, as such, one would imagine that the membrane anchoring structures of the NPC would be very highly conserved. Remarkably, however, there appears to be an absence of any identifiable orthologs of the opisthokont trans-membrane anchoring Poms (ScPom152/HsGp210, ScNdc1/HsNdc1, ScPom34, and HsPom121) in the trypanosome NPC interactome. Nonetheless, we identified a Nup with a TM domain, but which was intriguingly different from those in opisthokonts. TbNup65, a newly identified TbNup, appears orthologous to ScNup53/59 and HsNup35 (Fig 1) and contains an RRM (RNA recognition motif) domain also found in these opisthokont proteins [63] at residues 81– 153 (Figs 1C and 5A). The Nup35-type RRM is a noncanonical ribonucleoprotein motif that lacks key residues involved in RNA binding, making it identifiable by bioinformatics [63]. However, the most intriguing feature of TbNup65 is the presence of a predicted TM domain at residues 516–535 (Fig 5A). This TM domain is present in all kinetoplastid Nup65 homologs (S4 Fig). The presence of a TM domain in TbNup65 was confirmed by carbonate extraction, where TbNup65 was recovered exclusively in the pellet, behaving identically to another predicted TM protein, Tb927.7.4760, which localizes to both the nuclear rim and the Golgi (S5 Fig). This is distinct from TbNup89, which possess no predicted TM domain and, as expected, was efficiently extracted by carbonate (Fig 5A).
In opisthokonts, ScNup53/HsNup35 connect the pore membrane to the core scaffold of the NPC, a role critical for assembly [65,66]. However, the interaction of ScNup53/59 with the pore membrane is mediated by an amphipathic lipid-packing sensor (ALPS) motif at the C-terminus of each protein, and which associates with membranes [67–69]. Significantly, the ALPS motif and TM domains use different mechanisms of membrane association, as the former does not traverse the membrane [69,70]. TbNup65 interacts with TbNup96 and TbNup225 (Fig 2A), interactions that are conserved with the respective yeast and vertebrate orthologs [55,65,69,71,72]. In yeast, ScNup53 interacts directly with inner ring ScNup170 [4,73]. In vertebrates, Nup35 (ScNup53) interacts with inner ring proteins Nup93 (ScNic96), Nup155 (ScNup157/170), Nup205 (ScNup192), and the pore membrane protein NDC1 [65,71,72]. Thus, while connections between TbNup65 and the NPC appear largely conserved, the mechanism anchoring the NPC to the pore membrane appears to be distinct, and the moieties used to anchor the NPC to the pore membrane (TM domains, ALPS motifs) are interchangeable.
The next question we addressed was: does the level of conservation found in the inner ring component of the core scaffold extend to the outer ring? The outer rings are located on the cytoplasmic and nucleoplasmic faces of the NPC and are dominated by α-solenoids, β-propellers, or the β-α structure [4,6,10,31,58]. The architecture of both the yeast and vertebrate outer rings are comparatively well characterized, permitting more detailed comparisons.
TbNup158 is a clear ortholog of the yeast outer ring component ScNup145/HsNup98-96 [27]. Affinity capture of TbNup158 recovers six additional TbNups: TbSec13, TbNup41, TbNup82, TbNup89, TbNup132, and TbNup152 (Fig 2B). This septameric complex is repeatedly recovered in multiple affinity captures using these proteins as handles, with an additional protein, TbNup109, recovered in the affinity capture of TbNup82, suggesting these proteins may interact directly. Affinity capture using TbNup109-GFP itself confirms TbNup109 as a bona fide member of the same complex, as it co-purified all members of the complex (Figs 2B and 3) as well as being recovered in the entire TbNPC isolation (Fig 1A), suggesting TbNup109 is an easily displaced component. Under more stringent affinity isolation conditions, we find that the complex can be delimited to a module comprised of TbSec13, TbNup89, TbNup132, and TbNup158 (Fig 5B). Localization of a defining member of this complex by iEM, Nup89, shows that it is both axially and radially more distal from the central channel than the inner ring components (Fig 4), consistent with being part of a TbNPC outer ring. Hence, we named this complex the TbNup89 complex, likely representing the outer ring equivalent of the ScNup84 complex and HsNup107-160 complex [27,32,74–79].
The composition of the TbNup89 complex reveals a high degree of architectural conservation of the outer ring complex across eukaryotic evolution (Table 1). However, there are significant differences highlighted in Table 1. The most prominent is the presence of three β-α Nups, TbNup109, 132, and 152, as opposed to just two in opisthokonts [5,6]. Remaining differences revolve around the presence or absence of the small β-propeller proteins Seh1, Sec13, Nup37, and Nup43. Sec13 is present in all characterized versions of this complex, likely through a direct association with orthologs of TbNup158 [76]. Both Nup37 and Nup43 are absent from the S. cerevisiae Nup84 complex, but orthologs of Nup37 are present in other fungi, including Aspergillus nidulans, Schizosaccharomyces pombe, and C. thermophilum [80–83]. This compositional flexibility is also apparent in the absence of a recognizable Seh1 ortholog in the trypanosome NPC, just as in the NPCs of thermophilic fungi [83,84] and in affinity captures of the TbNup89 complex. Rather, TbNup41, the only other β-propeller protein in the TbNup89 complex besides TbSec13, appears to have a distinct ancestry to that of Seh1, as determined by phylogenetic analysis, and is likely orthologous to Nup43 (S6 Fig). Overall, therefore, the outer ring—though carrying many conserved features—has more lineage-specific subunits than the inner ring.
The degree of conservation of the peripheral components of the NPC is much less established. Candidate proteins corresponding to components of the cytoplasmic fibrils, and specifically orthologs of the linker Nup ScNup82/HsNup88 or FG-Nups ScNup159/HsNup214 and ScNup42/hCG1 that are crucial for mRNA export, have never been identified in trypanosomes [6,27,85–93]. Likewise, no apparent orthologs of the nuclear-face-localized FG-Nups ScNup1 or ScNup60 have been identified in trypanosomes [6,27].
We can now assign several trypanosome FG-Nups to specific locations within the NPC, depending on the scaffold Nups with which they stably interact and co-purify. As described above, TbNup53a, TbNup53b, TbNup62, and TbNup158 are all symmetrically disposed FG-Nups, facing both the cytoplasmic and nucleoplasmic faces on the NPC (Figs 2 and 3). However, we could not accurately determine the localization of TbNup64, TbNup75, and TbNup98 by affinity capture alone, as they interact with both inner and outer ring scaffold Nups as well as the nuclear basket (Figs 2C and 3). The FG-Nups TbNups64 and 75 are paralogs, with near-identical amino acid sequences, albeit with several insertions in TbNup75 that are responsible for the size difference between the two. We presume TbNup75 function and localization to be similar to TbNup64, as they interact directly (Fig 2C). To more accurately determine the sublocalization of these FG-Nups, we performed post embedding iEM gold labeling for TbNup64 and TbNup98 (Fig 4). We found that both have a symmetric distribution in the trypanosome NPC, close to the central channel and the NPC’s equator, consistent with their strong interactions with both the inner and outer ring.
Affinity capture of the TbNPC identified TbNup76, a predicted β-propeller protein with a short coiled-coil C-terminal region (Fig 1). This secondary structure is similar to that of ScNup82/HsNup88, the only opisthokont or plant Nup (AtNup88) with this architecture, suggesting that they are orthologs (Fig 1C) [7–9]. This orthology is supported by the observation that capture of tagged Nup76 also yields an untagged copy of itself (Fig 2D), suggestive of the same kind of dimeric architecture found in opisthokonts [4,94].
Affinity isolation of TbNup76-GFP identifies it as part of an NPC subcomplex containing the two largest FG-Nups, TbNup140 and TbNup149 in the TbNPC, which also co-purify with each other and Nup76 (Figs 2D and 3). This complex interacts with some members of the TbNup89 complex, specifically TbNup132 and TbNup158 (Fig 3). The interaction between TbNup76 and the TbNup89 complex suggests that the latter may anchor TbNup76 and its associated FG-Nups. High-density FG repeats (101 in total) comprise 117 kDa of TbNup140, while the N-terminal region contains a 23 kDa predicted coiled-coil [27]. By contrast, TbNup149 is not as FG-rich (18 FGs) and is composed of three near identical repeated domains that comprise the entire protein (S7 Fig). Additionally, the repeated units have putative zinc finger domains, the significance of which is currently under investigation (S7 Fig). Notably, neither TbNup140 nor TbNup149 has structural similarity to either ScNup159, which has an N-terminal β–propeller domain, or ScNup42, the two cytoplasmic FG-Nups of the yeast NPC (or their vertebrate orthologs), suggesting that the organization of the FG-Nups in trypanosomes is likely distinct.
To directly address this, we localized TbNup76, again using post embedding gold labeling iEM (Fig 4). Surprisingly, TbNup76, the putative ortholog of the cytoplasmically facing Nup82 in yeast, exhibits a symmetric localization, suggesting that it is found on both nucleoplasmic and cytoplasmic faces of the trypanosome NPC. By extension, the FG-Nups TbNups140 and 149, which interact with TbNup76, are predicted to localize symmetrically. Together with the apparently symmetric localization of the other Nups tested by iEM, this unexpected result suggests that the only definitively asymmetrically localized components are the nuclear basket proteins TbNup110 and TbNup92 [27,35], while all other components are equally disposed on the nuclear and cytoplasmic halves of the TbNPC. This is highly distinct from opisthokont NPCs, over a quarter of whose Nups are asymmetrically localized to only their nuclear or cytoplasmic faces. This large-scale architectural difference is likely connected to the absence of obvious orthologs of cytoplasmic or nucleoplasmic-biased FG-Nups, i.e., ScNup159/HsNup214 and HsNup153/ScNup1-Nup60.
An absence of clear nucleocytoplasmic asymmetry in the trypanosome NPC is remarkable, especially as NPC asymmetry is crucial for driving opisthokont mRNA export [22,89,95]. In particular, the ATP-dependent DEAD box RNA helicase Dbp5 and the RNA export mediator Gle1, with its cofactor IP6 (inositol hexakisphosphate), associate with the N-terminal β-propeller of cytoplasmic FG-Nup ScNup159/HsNup214, a member of the ScNup82 complex and remodel messenger ribonucleoproteins (mRNPs) exiting the nucleus [19,20,23–25,96–98]. This allows the non-karyopherin RNA export factors (Mex67:Mtr2 in yeast, TAP:p15 in humans) to disengage and recycle back into the nucleus, providing the necessary directionality and energy to RNA export [99–101]. As well as lacking a ScNup159/HsNup214 ortholog, orthologs of Gle1 and Dbp5 are absent from affinity-captured complexes and cannot be identified in the trypanosome genome. (See S8 Fig for phylogenetic analysis. Files are viewable using the free “Archaeopteryx” software.) By contrast, orthologs of other RNA export factors, including ScMex67:Mtr2/HsTAP:p15 and ScGle2/HsRae1, can be readily identified in trypanosomes [27,102,103].
In opisthokonts, Mex67/Mtr2 interacts with numerous Nups and RNA processing factors, including Gle1 and Dbp5 [40]. Therefore, to understand how Mex67 interacts with the NPC in trypanosomes and assess the composition of any potential RNA processing platform, we affinity captured TbMex67 under a variety of stringencies (Fig 6A). Under high stringency conditions, we found TbMex67 co-isolated with TbNup76, TbNup140, and TbNup149, as well as the highly conserved binding partner of Mex67, TbMtr2 [102]. This strongly implies that the TbNup76 complex is part of the mRNA export factor docking platform. Under low stringency conditions, we co-isolated several TbNups and transport factors. Significantly, no potential orthologs of mRNA export factors Dbp5 and Gle1 were identified (Fig 6A). Thus, the absence of these proteins is suggestive of an mRNA export mechanism that is probably different from that in opisthokonts.
Besides TbNups, TbMex67/TbMtr2 forms a complex with Ran and other putative Ran binding proteins (RanBP1 and GAP TbTBC-RootA). It is unclear whether TbMex67/Mtr2 can bind Ran directly or is doing so via these other proteins (Fig 6A). If direct, presumably the interaction would be via the NTF2-like domains of Mex67 and Mtr2. NTF2 binds to and imports Ran-GDP into the nucleus [107–110]. Ran binds NTF2 via a highly conserved phenylalanine (Phe72), called the “switch II” region, which binds a hydrophobic pocket on NTF2 [107]. In opisthokonts, Ran binding to TAP (Mex67) is blocked by a helix preventing access to the equivalent hydrophobic pocket of the NTF2-like domain in Mex67 [106]. Likewise, Ran binding to opisthokont p15 (Mtr2) is prevented by the presence of large hydrophobic residues in the corresponding hydrophobic pocket that obstruct the incoming Phe72 of the Ran switch II region [106,111]. We were able to generate high confidence models of TbNTF2, the TbMex67 NTF2-like domain, and TbMtr2, because of their sequence similarity to their structurally characterized opisthokont orthologs (Methods). Our models support the binding of NTF2 to Ran in trypanosomes, as the hydrophobic Ran binding pocket in TbNTF2 appears to be accessible and conserved (Fig 6B). Our models also suggest that the Ran binding pocket of the NTF2 domains of TbMex67 and TbMtr2 are occluded and, thus, inaccessible to Ran binding, exactly as in opisthokonts [106]. Thus, based on our models, direct Ran GTP-dependent interaction seems unlikely, rather being through RanBP1 and the GAP (TbTBC-RootA).
How can we reconstruct eukaryogenesis and the pathways that lead to and from the prokaryote/eukaryote transition? One potentially valuable approach is to understand the structures and mechanisms operating at the nuclear envelope from key organisms across the eukaryotic lineage. A detailed comparative dissection of the machinery mediating central functions can enable reconstruction of evolutionary history and origins. The data reported here provide the first comprehensive survey of the architecture of the NPC from a highly divergent organism, providing key insights into evolutionary origins of function and mechanism at the nuclear envelope.
Overall, there is a high degree of conservation between the trypanosome, opisthokont, and vascular plant NPCs at the level of subunit composition, although the trypanosome appears most divergent [5,6,8,27]. Rather than primary structure, conservation is at the level of shared structural domains in similar arrangements. Strikingly, the molecular weights of orthologs are very well conserved (S3 Table) and may reflect severe spatial constraints to assembling a cylindrical structure delimiting a ~40 nm channel, containing correctly spaced gating FG repeats and both spanning and stabilizing the ~50 nm pore membrane. The core scaffold (inner and outer rings) of the NPC, comprised of orthologous proteins carrying coatomer-related α-solenoid, β-propellers, and β-α structures, is highly conserved between trypanosomes, vascular plants, animals, and fungi, but with notable differences (Fig 7A) [5,6,8,10,27]. Significant conservation of this NPC substructure was expected, as it is a member of the protocoatomer group of membrane-deforming complexes that mediate membrane trafficking and intraflagellar assembly and transport (reviewed in [11]). This further evidence supports the paradigm that the ancestor of these membrane-deforming complexes arose via a pre-LECA expansion from an ancestral protocoatomer complex [10,11].
Within the core scaffold, the inner ring is the most conserved component of the NPC, with clear orthologs in vascular plants (Fig 7B) [7,8]. This high degree of conservation was unclear until our survey provided robust evidence that the FG-Nups62, 53a, and 53b were the orthologs of ScNsp1/HsNup62, ScNup57/HsNup58, and ScNup49/HsNup54, respectively, and that the inner ring organization of trypanosome, yeast, and vertebrate NPCs are very similar (Fig 7A). By contrast, the trypanosome outer ring TbNup89 complex displays several divergent features, the most significant of which are the absence of the β-propeller protein Seh1 and the possession of three large β-α structure Nups (Nup152, Nup132, and Nup109) rather than just two, as present in all other lineages examined so far (Table 1). Perhaps three β-α structure Nups are a remnant of an earlier, more LECA-like architecture for this complex, lost in other lineages; further detailed structural mapping of the TbNup89 complex as well as analyses from additional divergent taxa may help resolve this possibility.
The high level of conservation of inner ring features extends to TbNup65, the ortholog of ScNup53/HsNup35. TbNup65 interacts with the nuclear membrane via an orthodox TM that is conserved between kinetoplastids (S4 Fig) and represents the sole membrane anchor identified in the trypanosome NPC by our methods. That ALPS and TM domains appear functionally interchangeable suggests that the precise mechanism of anchoring the NPC to the nuclear membrane is unimportant, so long as it has some such mechanism (Fig 7A). This idea is supported by the observation that deletion of all TM proteins from A. nidulans NPCs has no deleterious effects on viability (although the putative ALPS-containing proteins are essential in this context) [81]. The absence of an ortholog to the TM protein ScPom152 in trypanosomes is notable, as orthologs are present in other opisthokonts and plants. Pom152 has a cadherin domain, in common with many membrane receptors and proteins that bridge between two membranes [9]. Thus, while this could reflect lineage-specific loss from trypanosomes, a more attractive interpretation is as an example of neofunctionalisation of a membrane protein into a NPC-specific role, postdating speciation between opisthokonts, plants, and trypanosomes.
Immunoelectron microscopy localization of several Nups, representing key subcomplexes and Nup classes, showed that all were symmetrically disposed between the nuclear and cytoplasmic faces of the NPC. The exception was the nuclear basket analog TbNup110, which confirmed the ability of our approach to reveal asymmetric localizations. Moreover, clear homologs of Nups and accessory transport factors with asymmetric nucleocytoplasmic distributions on the NPC were absent from our affinity captures and from exhaustive informatics screens of the trypanosome genome. Taken together, our data therefore indicates that, with the exception of the nuclear basket, the trypanosome NPC lacks a clear nucleoplasmic- or cytoplasmic-biased localization of Nups, in contrast to opisthokonts (Fig 7A). One source of Nup asymmetry in opisthokonts is from ScNup145/HsNup98-96, which can self-cleave to release an N-terminal fragment (ScNup145N) that localizes preferentially to the nuclear side of the NPC. Intriguingly, ScNup145N facilitates the connection between inner and outer ring complexes via discrete binding motifs for inner ring, central channel, and cytoplasmic Nups [112]. In contrast, TbNup158, the ortholog of this protein in trypanosomes, lacks the catalytic residues required for autoproteolytic cleavage to generate FG-Nup (ScNup145N/HsNup98) and α-solenoid Nup (ScNup145C/HsNup96) fragments [27,113–115]. Thus, FG-Nup symmetry is maintained by ensuring that TbNup158 is incorporated into the TbNPC as a single FG/α-solenoid protein in the symmetrically disposed outer ring complex. In addition, TbNup76, orthologous to ScNup82/HsNup88, is located on both the cytoplasmic and nucleoplasmic faces of the trypanosome NPC, but is part of an exclusively cytoplasmic NPC subcomplex in opisthokonts (Fig 7A). Furthermore, the FG-Nup ScNsp1/HsNup62 is also present in two distinct NPC subcomplexes; the inner ring ScNic96/HsNup93 complex and the cytoplasmic ScNup82/HsNup88 complex [87,116], thus representing another form of Nup asymmetry in opisthokont NPCs. In contrast, none of the potential trypanosome orthologs of ScNsp1/HsNup62 appears to associate with another TbNPC subcomplex, further highlighting the distinct symmetry exhibited by the TbNPC.
The symmetric arrangement in trypanosomes is also consistent with the hypothesis that the basic mechanism of nucleocytoplasmic transport does not require inherent NPC asymmetry [6,13]. However, it is significant that while trypanosomes share a diverse array of FG-Nup “flavors” with opistokhonts, in trypanosomes, this does not correlate strongly with their nucleocytoplasmic arrangement (S4 Table).
The main mRNA export factor Mex67 and its partner Mtr2 are conserved in trypanosomes, consistent with previous observations that karyopherin transport factors are also well conserved [117]. Given this evolutionary conservation of transport factors, there is, a priori, no reason to suspect major differences in transport mechanisms in trypanosomes. However, the cytoplasmically disposed, ATP-powered mRNA export platform formed by the ScNup82/HsNup88 complex, specifically ScNup159/HsNup214 plus the export factors Gle1 and the ATP-dependent helicase Dbp5 in opisthokonts [22,118,119], appears almost entirely lacking in trypanosomes. Therefore, in the absence of this cytoplasmic ATPase assembly, how is mRNA export both powered and provided with directionality in trypanosomes? A possible mechanism is suggested by affinity captures of Mex67, which recovered stoichiometric amounts of the GTPase Ran, RanBP1, and a putative GTPase activating protein, even though neither yeast nor vertebrate Mex67 or Mtr2 bind Ran [40,106]. Previous work has suggested that trypanosome mRNA export may be mechanistically distinct from that in opisthokonts and plants, with a shared platform for transport of rRNA and mRNA [30,120]. Here, our data strongly suggest that mRNA export in trypanosomes is dependent for both directionality and energy on the GTPase Ran, similar to karyopherin-mediated transport (Fig 7B). In opisthokonts, Ran, RanBP1, and a RanGAP are normally involved in an exquisite interplay that promotes hydrolysis of RanGTP to RanGDP, facilitating cargo release into the cytoplasm [121–123], and perhaps an analogous mechanism is involved in trypanosome mRNA export. Trypanosomes have rather unusual mechanisms for controlling gene expression, possibly a reflection of early divergence that places them close to the eukaryotic root [124,125]. Trypanosome protein-coding genes lack introns and are organized into directional polycistronic transcription units (PTUs) comprised of functionally unrelated genes [126,127]. Each gene lacks an individual promoter, with transcription start and stop sites only present for the entire PTU [128]. PTUs are transcribed by RNA polymerase II into long polycistronic transcripts, and the processing of single mRNAs is achieved by trans-splicing and subsequent polyadenylation, with regulation of gene expression therefore relying mainly on mRNA turnover and translation rates [129,130]. This exclusive trans-splicing of protein-coding mRNAs in trypanosomes may remove much of the complexity of mRNA processing, relaxing requirements for extensive chaperoning or quality control during nuclear export, and so accounting for the differences we find between the opisthokont and the kinetoplastid mRNA export machineries. It is appealing to propose that trypanosome may exemplify (or may have reinvented) an ancestral configuration for nucleocytoplasmic transport, whereby all transport factors operated in a Ran-dependent manner, but this remains tentative at this time.
The origin of an NE that defines the nucleoplasm necessitated development of an exchange mechanism with the cytoplasm. Hence, the NPC must, at least in part, embody this major transition in cellular architecture. Despite 1.5 billion years separation, animals, fungi, plants, and trypanosomes all utilize the NPC for nucleocytoplasmic transport, plus mRNA processing and maintenance of the chromatin environment. While the NPC demonstrates significant subunit conservation across eukaryotes, the manner in which the NPC connects with the lamina and mRNA transport is likely highly divergent between these lineages [35,42].
The trypanosome NPC architecture supports our earlier model of NPC evolution, which proposed that the ancestral NPC was an ungated pore, with protocoatomer type subunits stabilizing fenestrations in the protoeukaryotic NE [38]. Conservation of the core scaffold, and the presence of the same folds throughout the scaffold, supports a basic tenet of this model, i.e., that the elaborate architecture of the NPC arose through repeated duplication events from a simple progenitor coating complex. Even the eight-fold symmetry, conserved in trypanosomes [131], suggests a model for a stepwise monomer to dimer to tetramer to octamer transition during evolution. Of significance is that membrane anchoring of protocoatomer systems is promiscuous [11], consistent with divergent NPC membrane tethering described here. Selective gating by FG-Nups was proposed as a more recent acquisition, facilitating more selectivity in import and export [27]. Nevertheless, the high degree of conservation found in the inner ring complex, which contains representatives of all the major elements of the transport machinery (coatomer, karyopherin, FG Nup, membrane association), suggests an intermediate but simpler architecture for a transitional pre-LECA NPC. We propose that, subsequently, a more elaborate architecture evolved, leading to differentiated inner and outer rings and peripheral structures, and providing specific and different functionalities at the nuclear versus cytoplasmic sites. This allowed the development, in particular, of elaborations in mRNP processing and assembly at the NPC's nucleoplasmic face and ATP-dependent export and unloading on the cytoplasmic face. This may also have driven remodeling of FG-Nup positioning, with the trypanosome symmetric arrangement perhaps reflecting that in the LECA NPC, and being consistent with the trypanosomatid lineage as one of the earliest to differentiate following the eukaryogenesis event.
T. brucei procyclic Lister 427 strain cells were cultured in SDM-79, supplemented with 10% fetal bovine serum as previously described [27,132]. Expression of plasmid constructs was maintained using Hygromycin B at 30 μg/ml.
All proteins tagged in this study used the pMOTag4G tagging vectors [133] as previously described [27].
GFP-tagged cell lines were harvested and fixed for 10 mins in a final concentration of 2% paraformaldehyde. Fixed cells were then washed in 1xPBS and visualized as previously described [27].
Trypanosomes were grown to a density of between 2.5 x 107 cells per ml. Parasites were harvested by centrifugation, washed in 1xPBS with protease inhibitors and 10mM dithiothreitol, and flash frozen in liquid nitrogen to preserve protein:protein interactions as close as they were at time of freezing as possible. Cells were cryomilled into a fine grindate in a planetary ball mill (Retsch). For a very detailed protocol, refer to Obado et al., 2015 (in press), Methods in Molecular Biology, or the National Center for Dynamic Interactome Research website (www.NCDIR.org/protocols). Cryomilled cellular materials were resuspended in various extraction buffers (S1 Table) containing a protease inhibitor cocktail without EDTA (Roche), sonicated on ice with a microtip sonicator (Misonix Ultrasonic Processor XL) at Setting 4 (~20W output) for 2 x 1 second to break apart aggregates that may be invisible to the eye, and clarified by centrifugation (20,000 x g) for 10 min at 4°C (Obado et al., 2015 (in press), Methods in Molecular Biology, or www.NCDIR.org/protocols) [41]. Clarified lysates were incubated with magnetic beads conjugated with polyclonal anti-GFP llama antibodies on a rotator for 1 hr at 4°C. The magnetic beads were harvested by magnetization (Dynal) and washed three times with extraction buffer prior to elution with 2% SDS/40 mM Tris pH 8.0. The eluate was reduced in 50 mM DTT and alkylated with 100 mM iodoacetamide prior to downstream analysis (SDS-PAGE followed by protein identification using MS—electrospray ionization (ESI) or MALDI-TOF). Eluates were fractionated on precast Novex 4–12% Bis Tris gels (Life Technology), stained using colloidal Coomassie (GelCode Blue—Thermo) and analyzed by MS [27].
Briefly, protein bands were excised from acrylamide gels and destained using 50% acetonitrile, 40% water, and 10% ammonium bicarbonate (v/v/w). Gel pieces were dried and resuspended in trypsin digestion buffer; 50 mM ammonium bicarbonate, pH 7.5, 10% acetonitrile, and 0.1–2 ug sequence-grade trypsin, depending on protein band intensity. Digestion was carried out at 37°C for 6 h prior to peptide extraction using C18 beads (POROS) in 2% TFA (trifluoroacetic acid) and 5% formamide. Extracted peptides were washed in 0.1% acetic acid (ESI) or 0.1% TFA (MALDI) and analyzed on a LTQ Velos (ESI) (Thermo) or pROTOF (MALDI-TOF) (PerkinElmer).
Newly identified TbNups were analyzed for several secondary structure elements, including β-sheets and α-helices using PSIPRED [134] and Phyre2 [57], natively unfolded regions using Disopred [135], trans-membrane helices using Phobius [136], and coiled-coil regions using COILS [137].
Trypanosomes were cryoprotected with 20% bovine serum albumin (BSA) and applied to a high pressure freezing procedure (EMPACT2, Leica Microsystem System, Wetzlar, Germany). Cells were transferred to a freeze substitution device (EM AFS2, Leica Microsystem System, Wetzlar, Germany), incubated with 0.2% Uranyl acetate in 95% acetone at -90 C°, and embed in Lowicryl HM20 at -35 C°. Ultrathin sections were cut and post-embedding immunostaining was applied. Briefly, sections were blocked with 2% BSA plus 0.1% saponin in Tris buffered saline (TBS; 20 mM Tris-Cl, pH 7.5, 150 mM NaCl) for 30 min. Sections were then incubated in fresh blocking solution containing polyclonal rabbit anti-GFP antibodies (1:150) overnight at 4°C, and washed with TBS the next day. The EM sections were then incubated overnight with secondary goat anti-rabbit antibodies conjugated with 12 nm colloidal gold (1:20) in 0.2% BSA plus 0.1% saponin in TBS and then washed in TBS buffer. An additional wash step using 1 x PBS was performed prior to fixation for 5 min with 2.5% glutaraldehyde. Post fixed grids were washed with water and uranyl acetate (1%), and lead citrate (1%) was applied. The sections were examined in the electron microscope (100CX JEOL, Tokyo, Japan) with the digital imaging system (XR41-C, AMT Imaging, Woburn, Massachusetts). Control experiments were done by following the same procedure, except for the omission of primary antibody and applying just the blocking solution instead.
We selected NPCs sectioned perpendicular to the NE plane with a clearly visible nuclear envelope double membrane. We selected a radius of 300 nm around the estimated center of each NPC as an excision limit and then created an aligned superimposed montage using the resulting excised NPC images [6,36]. See S2 Fig. For the radial position of each Nup (R), we used the method described in [58] and the peak finding algorithm of Alber et al., 2007 [3,4]. For the axial position of each Nup (Z), we essentially used the method described in [3,4]; for both, errors were estimated from the 95% level of the peak finding algorithm.
Powder from cryomilled trypanosomes was resuspended in 0.1 M Na-Carbonate buffer, pH 11 to a ratio of 1:9 (powder:buffer) and then processed as previously described [6].
3D structures were modeled using the program I-TASSER [105,138], which combines fold recognition, where the template is threaded onto similar structures retrieved from the pdb, full length reconstruction of the template involving ab initio modeling of unaligned regions and rigorous high-resolution refinement to generate a final protein model. For our studies, no threading templates from the pdb were specified; instead, we chose to employ the default search criteria on the I-TASSER server for template threading. All models were viewed and figures generated using PyMOL (The PyMOL Molecular Graphics System, Version 1.7.4 Schrödinger, LLC.).
|
10.1371/journal.pntd.0006961 | Modelling the cost-effectiveness of a rapid diagnostic test (IgMFA) for uncomplicated typhoid fever in Cambodia | Typhoid fever is a common cause of fever in Cambodian children but diagnosis and treatment are usually presumptive owing to the lack of quick and accurate tests at an initial consultation. This study aimed to evaluate the cost-effectiveness of using a rapid diagnostic test (RDT) for typhoid fever diagnosis, an immunoglobulin M lateral flow assay (IgMFA), in a remote health centre setting in Cambodia from a healthcare provider perspective. A cost-effectiveness analysis (CEA) with decision analytic modelling was conducted. We constructed a decision tree model comparing the IgMFA versus clinical diagnosis in a hypothetical cohort with 1000 children in each arm. The costs included direct medical costs only. The eligibility was children (≤14 years old) with fever. Time horizon was day seven from the initial consultation. The number of treatment success in typhoid fever cases was the primary health outcome. The number of correctly diagnosed typhoid fever cases (true-positives) was the intermediate health outcome. We obtained the incremental cost effectiveness ratio (ICER), expressed as the difference in costs divided by the difference in the number of treatment success between the two arms. Sensitivity analyses were conducted. The IgMFA detected 5.87 more true-positives than the clinical diagnosis (38.45 versus 32.59) per 1000 children and there were 3.61 more treatment successes (46.78 versus 43.17). The incremental cost of the IgMFA was estimated at $5700; therefore, the ICER to have one additional treatment success was estimated to be $1579. The key drivers for the ICER were the relative sensitivity of IgMFA versus clinical diagnosis, the cost of IgMFA, and the prevalence of typhoid fever or multi-drug resistant strains. The IgMFA was more costly but more effective than the clinical diagnosis in the base-case analysis. An IgMFA could be more cost-effective than the base-case if the sensitivity of IgMFA was higher or cost lower. Decision makers may use a willingness-to-pay threshold that considers the additional cost of hospitalisation for treatment failures.
| Typhoid fever is a common disease among children in Cambodia. It can be fatal or lead to chronic faecal carriage if not treated. In resource-limited settings, typhoid fever is often diagnosed and treated presumptively. This study evaluated the cost-effectiveness of introducing a rapid diagnostic test for typhoid fever in a remote setting in Cambodia. In a hypothetical cohort with 1000 children in each arm we compared the use of a rapid diagnostic test (RDT) with a presumptive clinical diagnosis. In each arm, we calculated the number of true-positive typhoid fever cases detected, treatment success at seven days, and the cost of making a correct diagnosis and providing the correct treatment. The RDT detected 5.87 more true positives, had 3.61 more treatment successful cases, but the total cost was $ 5700 higher per 1000 children. Additional analysis showed that the RDT would be more cost-effective if the sensitivity could be improved or cost lowered.
| Typhoid fever is estimated to cause 21 million new cases per year worldwide [1, 2]. It is a common disease among children in resource-limited settings such as Cambodia [2–5]. In Cambodia, which is classified as a high incidence area for typhoid fever, the distribution of typhoid fever cases is highest in children aged under 15 years [6]. Typhoid fever is a systemic infection with non-specific clinical features that make it difficult to differentiate from other common febrile illnesses [5, 7]. A dry cough, for example, is a common symptom and may lead to confusion with pneumonia [5, 7]. As many as ten to fifteen percent of the patients who have been sick for more than two weeks with typhoid develop severe complications (gastrointestinal haemorrhage, shock or hepatitis) [1, 5, 8]. The case fatality ratio was reported to be 10–30% in the pre-antimicrobial era [1, 9]. Effective antimicrobial treatment should decrease the case fatality ratio to less than 1%. Prompt diagnosis and appropriate antimicrobial drug therapy is needed to avoid severe or fatal disease, relapse and also acute and chronic faecal carriage that may lead to onward transmission of typhoid [9]. Empirical treatment with antimicrobials should be guided by local data since the susceptibility of isolates widely varies among countries and regions [10, 11]. Salmonella enterica serovar Typhi (S. Typhi) isolates at Angkor Hospital for Children (AHC) in Siem Reap, north-west Cambodia are dominated by strains that are multi-drug resistant (MDR) (resistant to chloramphenicol, ampicillin and co-trimoxazole) and with decreased susceptibility to ciprofloxacin [12]. In this setting, oral azithromycin remains an option for the initial empiric treatment of children with suspected uncomplicated typhoid fever. Children with a febrile illness in this setting can have other bacterial illnesses such as community-acquired pneumonia [13, 14]. The common causative organisms among children are Streptococcus pneumoniae, Haemophilus influenzae type b, and following the guidelines of the World Health Organization, they can be treated with amoxicillin [15]. Amoxicillin would be inadequate in this area if the true diagnosis was MDR typhoid [12].
Differentiating typhoid from other causes of fever in children in this area is challenging without rapid and reliable diagnostic tests. The recommended reference standard diagnostic tests for typhoid fever are blood culture or bone marrow culture with sensitivities of 40–80% and 80–95% respectively [5, 7]. Both tests are invasive, technically demanding and are not available in remote health care settings where the majority of uncomplicated typhoid cases present. Additionally, they require several days for a positive result to be confirmed [4, 7]. Low-cost diagnostic test, such as Widal test, is still widely used but lacks sensitivity and specificity [7, 16].
There are a number of commercially available rapid diagnostic tests (RDTs) for typhoid fever [17]. For example, an immunoglobulin M lateral flow assay (IgMFA), which detects immunoglobulin M (IgM) antibodies against the lipopolysaccharide of S. Typhi, has been evaluated in Cambodian and Bangladeshi children [16, 18]. Among other RDTs, this IgMFA has advantages of simplicity to perform, no need for refrigeration, and giving results within 15–30 minutes. It has a sensitivity of 59% and specificity of 98% [18]. The availability of such an RDT in resource-limited settings might be expected to lead to an early and appropriate choice of antimicrobial drug treatment.
The overall costs of typhoid fever have been evaluated in studies in India and other Southeast Asian countries, but a cost-effectiveness analysis of the use of RDTs in resource-limited settings has not been performed [19, 20]. This study aimed to evaluate the effectiveness and costs of using the IgMFA in a typhoid fever endemic country and to inform decision-making on whether to introduce the IgMFA in a remote health centre setting in Cambodia. Cambodia was chosen because it is a high-burden country for typhoid and because of the availability of data to inform model parameterization. In rural Cambodia half of the population is estimated to be aged under 25 years [21] and in children presenting to health facilities with fever, the prevalence of uncomplicated S. Typhi infection is estimated to be between 4.5% and 9.0% [18, 22, 23].
We conducted a cost-effectiveness analysis (CEA) of using an RDT for typhoid fever diagnosis, IgMFA, in a remote health centre setting in Cambodia from a healthcare provider perspective. The comparator was the current standard of care which is presumptive clinical diagnosis without an RDT. The costs that were included were therefore the direct costs of the diagnostic test (including supply, labour and equipment costs) and the costs of treatment. Costs borne by patients including indirect costs due to productivity loss were not included. The effects were natural units: treatment success at 7 days and an intermediate outcome of correctly diagnosed typhoid fever. Cost-utility analysis presenting disability-adjusted life years (DALYs) was not chosen since it would not provide understanding of the net effects of offering a new diagnostic test.
Existing models for malaria RDT and a model for diagnosis of sepsis in low-resource settings were modified to construct a new decision analytic model for typhoid fever [24–26].
Two diagnostic approaches were compared. The new intervention arm employed the IgMFA and the comparator arm was a presumptive clinical diagnosis based on patient’s symptoms. The performance of the RDT was based on studies of the Life Assay Test-It IgMFA (Life Assay Diagnostics, Cape Town, South Africa) in Cambodia [16, 18]. Clinical diagnosis was made based on current clinical practice in AHC and other hospitals [16, 18, 27]. A patient was considered to have typhoid fever if febrile for more than 3 days and one or more of four clinical features was present; presence of abdominal symptoms (constipation, diarrhoea or abdominal pain); body temperature >39 °C; hepatomegaly and/or splenomegaly; or no alternative confirmed diagnosis established.
A decision tree model was constructed with a hypothetical cohort of 1000 children using Microsoft Excel (2016). We hypothesised that the children aged from 0 to 14 years, who visited a health centre in Siem Reap province, Cambodia, with undifferentiated fever, were eligible for the entry in this model. Malarial patients were excluded from the beginning of this model assuming that malaria was tested for and excluded before suspecting typhoid fever. Patients who had complicated symptoms, such as shock, encephalopathy, convulsions, bleeding, deep jaundice or suspected gut perforation were excluded as they would be referred to hospital. The decision tree model is shown in Fig 1.
Following the result of a diagnostic test, two treatment choices were defined based on literature [15, 27–35]. When a test was positive for typhoid fever, azithromycin (250 mg per day) for five days was prescribed. When a test was negative, amoxicillin (1500 mg per day) for five days was prescribed. The average weight of 15 kg was estimated based on data on the age distribution of children attending outpatient appointments in Cambodia [36] (S1 Appendix).
The effect of diagnosis was evaluated as two health outcomes. The number of correctly diagnosed typhoid fever cases (i.e. true-positives) was set as the intermediate health outcome and the number of treatment success for typhoid fever at day seven was set as the primary health outcome. The intermediate model outputs included number and cost of treatment of false-positives and true-negatives. Another choice of health outcome could be the number of correctly diagnosed cases (i.e. both true-positives and true-negatives). However, the main issue for diagnosis and treatment of typhoid fever in remote setting where MDR typhoid is common and where amoxicillin would not be effective is missing true-positives (under-diagnosis and inadequate treatment). Ineffective treatment of the missed case of typhoid fever may lead to the development of complications, hospital admission and mortality. Therefore, in this study, health outcome measures used in the cost-effectiveness analysis focused on the number of true-positives only and the number of treatment success in typhoid fever cases. Health outcomes were calculated using prevalence of typhoid fever, sensitivity and specificity of diagnostic tests derived from studies in Siem Reap province, Cambodia (S2 Appendix).
Several assumptions were made for the decision tree model. In rural areas, 18% of the people seek initial treatment at health centres run by the public sector, followed by private clinics (16%) and pharmacies (8%) [21]. For the purposes of the model we assumed that no child had been treated with antimicrobial drugs before visiting the health centre. A febrile child was assumed to have only one disease and co-infection was not considered. Regarding treatment, we assumed that the health centre workers perfectly adhered to the results of the tests and the treatment protocols in this model, the availabilities of both azithromycin and amoxicillin were 100%, and the patients’ adherence to the treatments was 100%. Although adherence issues are critically important to consider prior to implementation, there were no or limited local data on these parameters. As an important first step, this study aims to explore whether the introduction of IgMFA has the potential to be cost-effective.
We also assumed that there were no azithromycin-resistant strains of S. Typhi in the area. Thus, the MDR strains had the same treatment success probability as that of non-MDR strains when treated with azithromycin (Fig 1 *). On the contrary, we assumed that the MDR strains had 0% treatment effect if treated with amoxicillin. There are no available data for the prevalence of the MDR strains in the community. However, we set the MDR prevalence in the community at 50% based on the assumption that the prevalence of MDR is lower than that in the hospital settings in Siem Reap at 85% [12].
The primary health outcome was measured as the number of children with treatment success at day seven from the start of treatment in typhoid fever cases. The number of children with treatment success at day seven could be a useful measure of diagnosis since it represents the effect of correct diagnosis and subsequent treatment choice and is commonly used for assessing efficacy [32]. It is clinically and biologically plausible to think that after the seventh day patients will seek hospital care if they have no improvement in clinical symptoms.
In all arms, we assumed that no additional diagnostic tests would be performed until the endpoint of this model on day seven. If the diagnostic test for typhoid fever was negative, other possibilities were considered and amoxicillin was used presumptively.
Parameter inputs for the model were obtained from published literature from AHC and other literature from Cambodia or Laos whose study setting is similar to that in Siem Reap, Cambodia [13, 14, 16, 18, 23, 37], and are shown in Table 1. The reference test for both IgMFA and clinical diagnosis was blood culture [18]. Prevalence was estimated from typhoid fever cases confirmed by blood culture test based on local hospital data [18]. Supplementary literature reviews were also used to cross-check the values. The probabilities of treatment success with azithromycin and amoxicillin were calculated from literature of randomised controlled trials [27–35]. For synthesising the effect, the weighted average assuming random effects was calculated using Stata MP 14.1 [38] (S3 Appendix).
Cost parameter inputs are shown in Table 2. No additional data collection was performed for this study. Direct medical costs are presented by the economic costs of diagnosis and treatment. The economic costs include recurrent costs of diagnosis (i.e. unit cost of performing a test) and treatment (i.e. unit cost of antimicrobials and supply costs). Costs of diagnosis include costs of equipment, consumables and staff salary. Starting-up costs, such as costs of training staff were also evaluated. Fixed costs such as facility costs were not evaluated in this study since the health centre facility was assumed to be already present and the facility costs would be the same for both IgMFA and the clinical diagnosis arm. Unit price data were obtained from CHOosing Interventions that are Cost Effective (WHO-CHOICE), International Drug Price Indicator Guide 2015, other literature, websites or expert opinions [18, 42–45]. The costs were derived using a micro-costing method with bottom-up approach [46]. To calculate staff costs and training costs, we assumed the personnel time on the basis of literature on malaria RDT and expert opinion [26, 47]. Regarding the costs of drugs, the median costs were derived from supplier costs of drugs and the shipping (i.e. supply) cost was set at 10%, following the recommendation of International Drug Price Indicator Guide 2015 [43] (S4 Appendix). All costs were presented in the United States (US) dollars in the year of 2016, adjusted using the World Bank purchasing power parities (PPPs) and the World Bank consumer price index (CPI) [48, 49].
Other costs and effects included the number of over-treated patients, cost per child (cost-effectiveness ratio C/E), cost per correctly diagnosed and cost per treatment success in each arm.
The incremental cost-effectiveness ratio (ICER) comparing IgMFA and clinical diagnosis was calculated by measuring the difference in costs and effects and represents the additional cost to gain one additional health outcome. The ICER for both primary and intermediate health outcomes were calculated. The ICER for primary health outcome was calculated as the difference in the total costs between IgMFA and clinical diagnosis, divided by the difference in the number of treatment success between IgMFA and clinical diagnosis. The ICER for the intermediate health outcome was calculated as the difference in the total costs between two arms, divided by the difference in the number of correctly diagnosed typhoid fever cases between the two arms.
For the ICER to gain one additional treatment success, sensitivity analyses were performed. To determine which key parameter drives the results in this model (i.e. assessment of parameter uncertainty), one-way sensitivity analyses were conducted changing one variable and keeping other variables constant. We also conducted two-way sensitivity analyses to evaluate distributions of the ICER, estimated by changes in product profiles of IgMFA. Whether the ICER reaches a certain willingness-to-pay (WTP) threshold was also evaluated in the two-way sensitivity analyses. Probabilistic sensitivity analysis (PSA) was performed to address parameter and model uncertainty.
Regarding test performance, sensitivity and specificity of both IgMFA and clinical diagnosis were changed ranging from a lower to an upper 95% confidence interval (CI) value. Also, we conducted best-case scenario analyses of the sensitivity and specificity for IgMFA at 100% for each, to determine whether the ICER reaches a WTP threshold. For the clinical diagnosis, worst-case scenario analyses with 0% sensitivity and specificity for each were conducted. Concerning treatment effects, the probability of treatment success was changed from a lower to an upper 95% CI in each drug.
Regarding cost parameters, the cost for IgMFA was changed assuming that it would vary from a half to twice the price in the primary data, by applying an existing model in a diagnostic test for sepsis [26]. Cost of azithromycin and amoxicillin was changed from minimum to maximum value in each. Prevalence of typhoid fever was also changed from a lower to an upper 95% CI value. Prevalence of MDR strains was changed from 25% to 90%. Plausible range of the prevalence in MDR strains was assumed based on data in AHC [12].
A PSA was performed with Monte-Carlo simulations by simultaneously varying the variables mentioned above. The assumptions of the distribution of each variable are shown in Table 3. Health outcomes and costs were stochastically generated by 1000 simulations.
A report from five Asian countries showed that the public sector cost of typhoid fever per hospitalised case varied between $0 and $116 (2005) [19]. We assumed that the treatment failure cases would receive additional treatment in a hospital setting after day seven. Therefore, to assess the ICER, the WTP threshold was set at $201 ($116 converted to the year of 2016 using CPI).
The cost, the number of correctly diagnosed typhoid fever cases and the number of treatment success cases in each arm are shown in Table 4. The results of the incremental analyses in CEA are also shown in the same table. In the base-case analysis, the total costs of conducting diagnosis and treatment for 1000 children were $8465 for IgMFA, and $2765 for the clinical diagnosis arm. IgMFA detected 5.87 more true-positives (38.45 true-positive typhoid cases out of 65.17 diseased) per 1000 children than the clinical diagnosis. The cost difference between the two arms was $5700 and the ICER to obtain one additional correct diagnosis of typhoid fever was estimated to be $972. With respect to the primary health outcome, IgMFA had 3.61 more treatment successes than the clinical diagnosis (46.78 versus 43.17). Therefore, the ICER to have one additional treatment success in typhoid fever was estimated to be $1579. We evaluated the impact of starting-up costs on the total cost or to the ICER in CEA. We aimed to compare unit cost of providing a test, and it had a minor impact even if added to the total costs. Thus, the starting-up costs were excluded from the total costs.
Various one-way sensitivity analyses showed that the ICER to gain one additional treatment success in typhoid fever was sensitive to a change in the sensitivity of both IgMFA and the clinical diagnosis, cost of IgMFA and the prevalence of typhoid fever or MDR strains. The ICER showed mild robustness to the change in treatment effect of amoxicillin and staff costs. The ICER was robust to the effect of a change in the specificity of both tests, treatment effect of azithromycin, and the cost of both azithromycin and amoxicillin.
Table 5 shows the effect of a change in the sensitivity of IgMFA on the difference in health outcomes, cost effectiveness ratio and ICER. The results of one-way sensitivity analyses, changing the sensitivity of the clinical diagnosis, prevalence of typhoid fever and MDR strains are also shown in Table 5. When the sensitivity of IgMFA was changed to 42% (lower 95% CI value), the ICER was estimated to be -$1776 since IgMFA resulted in 3.21 less cases of treatment success than the clinical diagnosis. The ICER would decrease to $285 under a situation of a perfect sensitivity with 20.05 more treatment successful cases. The cost of using IgMFA was still $70 higher per treatment success than the clinical diagnosis.
Regarding the effect of the change in the sensitivity of the clinical diagnosis, a 0% of sensitivity showed that the ICER would decrease to $241.
At a typhoid fever prevalence of 4.5% (lower 95% CI value), the ICER was estimated to be $2292, while at a prevalence of 9.0% (upper 95% CI value), the estimation was $1147. With a prevalence of MDR strains at 25%, the ICER increased to $2219. With 90% prevalence of MDR strains, IgMFA was more cost-effective ($1081 per an additional treatment success). The change in the specificity of both tests had a minor effect on the ICER (S5 Appendix). The ICER showed a modest change by a change in the probabilities of treatment success in azithromycin and amoxicillin (S5 Appendix). A tornado diagram in Fig 2 presents the effect of changing cost parameters. The ICER was sensitive to the change in the cost of IgMFA. If the cost of IgMFA was reduced to a half price of the base-case ($1.63), the ICER decreased to $1086, while doubling the price ($6.5) resulted in the ICER reaching $2570. A change in salary had a modest effect on the ICER. When the salary was reduced to the lowest value of $4.77 per hour, the ICER decreased to $1314, while the ICER increased to $1804 at the highest salary of $11.85. On the contrary, little effect was seen on the cost-effectiveness of IgMFA when the price of azithromycin or amoxicillin was changed (Fig 2).
Two-way sensitivity analyses were conducted to identify a combination of test characteristics which would fall below a WTP threshold of $201. From the results of the one-way sensitivity analyses, two-way sensitivity analyses were conducted in combinations of parameters as follows: sensitivity of IgMFA and cost of IgMFA, sensitivity of IgMFA and prevalence (both typhoid fever and MDR strains), and cost of IgMFA and prevalence (both typhoid fever and MDR strains). When the sensitivity of IgMFA approached to 100% and the cost of IgMFA was $1.63, the ICER dropped below $201. Distributions of the ICER are shown in Fig 3. When the sensitivity of IgMFA was 100% and the prevalence of MDR strains was 90% (maximum assumed value), the ICER reached $195 (Fig 3). No combination of the sensitivity of IgMFA and the prevalence of typhoid fever, the cost of IgMFA and the prevalence of typhoid fever or MDR strains reached the ICER below $201 (S6 Appendix).
Fig 4 shows the results of the probabilistic sensitivity analysis of the cost-effectiveness of replacing clinical diagnosis by IgMFA in terms of a cost-effectiveness plane. Each dot represents a pair of incremental effect and incremental cost, calculated by a combination of random values for parameters, which are assumed to be distributed as per Table 3. The horizontal axis measures incremental number of successfully treated cases when clinical diagnosis was replaced by IgMFA. The vertical axis measures incremental costs when clinical diagnosis was replaced by IgMFA. In all 1000 simulations the cost of IgMFA was higher than that of clinical diagnosis. Most cost-effect pairs lay in the north-east quadrant, suggesting that IgMFA resulted in a larger number of treatment success but was more costly than clinical diagnosis in those pairs. Some cost-effect pairs also lay in the north-west quadrant, meaning the use of IgMFA was less effective and more costly to gain treatment success.
A cost-effectiveness acceptability curve was generated to show the probabilities of diagnostic tests being considered as cost-effective, according to a WTP threshold by decision-makers to gain one additional treatment success (Fig 5). For instance, if a decision-maker in Cambodia was willing to pay $1000 per one additional treatment success of typhoid fever, there was a 30.5% probability that the use of IgMFA being cost-effective. However, the probability that the use of IgMFA being cost-effective would increase only to approximately 75%, even a decision-maker was willing to pay more than $15000 for an additional treatment success of typhoid fever case. If the WTP threshold was assumed to be $200, the probability of that IgMFA being cost-effective was 0.2%.
The particular IgMFA studied, with a sensitivity of 59% and cost of $3.25, was estimated to be more effective but more costly ($972 per one additional correct diagnosis of typhoid fever; $1579 per one additional treatment success in typhoid fever cases) than the clinical diagnosis in the base-case analysis. One-way sensitivity analyses highlighted that the sensitivity and the cost of IgMFA, the prevalence of typhoid fever and MDR strains are the key drivers for the ICER. The PSA suggests that the probability of IgMFA being cost-effective was 0.2% when WTP threshold to gain one additional treatment success was $201. However, two-way sensitivity analyses showed that in some situations under improved sensitivity and cost in IgMFA, the ICER could decrease below $201 (i.e. cost-effective).
There were several limitations in this study. Cost data were derived from literature reviews and websites. A bottom-up approach for costing was used, which is more precise than the step-down approach but more time-consuming and difficult to implement when detailed data are not available [46, 50]. Aggregated cost data were not available for a step-down approach. The study did not consider healthcare worker adherence to the test results or availability of antimicrobial drugs. In studies using malaria RDTs, adherence to test results has not been perfect [24, 51]. Azithromycin availability may be low in health centre settings and affect the number of day seven treatment successes. Treatment effects derived from trials, some of which included adults, and which were not conducted in Cambodia, may not reflect the true situation. Adherence of patients to treatments is a further variable to be assessed.
In this study, we assumed that malaria cases would have been excluded because in most tropical countries, malaria RDTs are widely used and a positive result would usually result in the patient being treated for malaria with no further testing. However, co-infection is possible and difficult to diagnose clinically. If we had not excluded malaria cases from our analysis, we would expect the IgMFA to increase the correct identification of malaria cases co-infected with typhoid. However, there are no data or reason to suggest that this would be any different from malaria uninfected cases. Therefore, this would not affect the results of the cost-effectiveness analysis. Similarly, although our analysis assumed no pre-treatment with antibiotics, we do not expect that including pre-treatment would affect the results of this analysis.
This study applied CEA for a health outcome measured as the number of correctly diagnosed cases and the number of treatment success at day seven. These health outcomes might not capture the overall effects of introducing IgMFA. There is an uncertainty in relationship between diagnosis and health outcomes and a correct diagnosis may not always translate into the overall benefit [52]. Also, this study might be underestimating the benefit of IgMFA because we are not accounting for the potential benefit in terms of the development of antimicrobial resistance through the reduction in the overuse of antibiotics in patients falsely identified as having typhoid. However, this is complicated by the possible shift of prescribed antibiotic classes, which was reported after malaria RDT [53]. Careful interpretation is necessary for a CEA of a diagnostic test and to obtain the overall effect, calculation of DALYs is necessary. Not only DALYs of true-positives and false-negatives (i.e. patients who have typhoid fever) but also DALYs of false-positives and true-negatives (i.e. patients who have other diseases) are required.
If false-positives were treated with azithromycin, the treatment effect for other diseases has to be also considered. Azithromycin will treat respiratory infections, but not urinary tract infections or other serious invasive bacterial infections including meningitis or sepsis due to other Gram-negative bacteria. The new drug may also change DALYs in other diseases. Under-treatment for typhoid fever may impose other issues, such as increased number of children with complicated disease or relapse and prolonged faecal shedding of the organisms leading to enhanced transmission. These effects may not be captured by a DALY of a patient in a static model but need dynamic modelling.
The probability of treatment success following a clinical diagnosis in this model was derived from data in a hospital setting and a clinical diagnosis by doctor with the benefit of additional blood tests (elevation of liver enzymes, low or normal white cell count, and serum glutamic pyruvic transaminase). The accuracy of clinical diagnosis would be lower available in a remote health centre setting with a basic healthcare worker. Thus, the difference in effect between IgMFA and the clinical diagnosis could be higher, and the IgMFA more cost-effective than the base-case estimation.
Introducing the IgMFA would lead to a small increase in the number of true typhoid fever cases detected, and a small increase in the number of treatment successes but with a high incremental cost ($1579 per an additional treatment success). Sensitivity analyses did not alter the result that the use of IgMFA was more costly than the presumptive management. The number of children successfully treated by replacing clinical diagnosis with IgMFA depends on the sensitivities of IgMFA and clinical diagnosis.
For the RDT to be cost-effective, a more accurate test is needed. For a cost less than $1.65, and a sensitivity close to 100% with a prevalence of MDR strains of 90%, the IgMFA can be cost-effective. Decision-maker may use a WTP threshold also considering the additional cost incurred when a treatment failure arises.
|
10.1371/journal.pcbi.1006011 | The self-organization of plant microtubules inside the cell volume yields their cortical localization, stable alignment, and sensitivity to external cues | Many cell functions rely on the ability of microtubules to self-organize as complex networks. In plants, cortical microtubules are essential to determine cell shape as they guide the deposition of cellulose microfibrils, and thus control mechanical anisotropy of the cell wall. Here we analyze how, in turn, cell shape may influence microtubule behavior. Building upon previous models that confined microtubules to the cell surface, we introduce an agent model of microtubules enclosed in a three-dimensional volume. We show that the microtubule network has spontaneous aligned configurations that could explain many experimental observations without resorting to specific regulation. In particular, we find that the preferred cortical localization of microtubules emerges from directional persistence of the microtubules, and their interactions with each other and with the stiff wall. We also identify microtubule parameters that seem relatively insensitive to cell shape, such as length or number. In contrast, microtubule array anisotropy depends on local curvature of the cell surface and global orientation follows robustly the longest axis of the cell. Lastly, we find that geometric cues may be overcome, as the network is capable of reorienting toward weak external directional cues. Altogether our simulations show that the microtubule network is a good transducer of weak external polarity, while at the same time, easily reaching stable global configurations.
| Plants exhibit an astonishing diversity in architecture and morphology. A key to such diversity is the ability of their cells to coordinate and grow to reach a broad spectrum of sizes and shapes. Cell growth in plants is guided by the microtubule cytoskeleton. Here, we seek to understand how microtubules self-organize close to the cell surface. We build upon previous two-dimensional models and we consider microtubules as lines growing in three dimensions, accounting for interactions between microtubules or between microtubules and the cell surface. We show that microtubule arrays are able to adapt to various cell shapes and to reorient in response to external signals. Altogether, our results help to understand how the microtubule cytoskeleton contributes to the diversity of plant shapes and to how these shapes adapt to environment.
| Despite their amazing diversity in shapes, biological organisms share some common structural components at the cellular level. Among those, one of the best conserved proteins across eukaryotes, tubulin, assembles into protofilaments, which in turn form 25 nm nanotubes known as microtubules, usually made of 13 protofilaments. The network of microtubules is highly labile and can reshape itself in a matter of minutes. In plants, microtubules form superstructures before (the preprophase band), during (the spindle) and after (the phragmoplast) cell division. Plant microtubules also form dense and organized arrays at the periphery of the cell during interphase [1] and these arrays are known as cortical microtubules (CMTs). The behavior of CMTs has been studied extensively at the molecular level [2]. One of their main functions is to guide the trajectory of the transmembrane cellulose synthase complex and thus to bias the orientation of cellulose microfibrils in the wall. This in turn impacts the mechanical anisotropy of the cell wall and controls growth direction [3–6]. This function explains why most mutants impaired in microtubule-associated proteins exhibit strong morphological defects [7].
Whereas this provides a clear picture on how microtubules impact cell shape, in turn how cell shape impacts microtubule behavior has been less explored. There is evidence that microtubule orientation depends on cell shape [8–10], with microtubules being mostly transverse to the longest axis, but this might require specific regulation because microtubules orient along the longer axis of the confining domain in vitro [11]. There is also evidence that shape-derived mechanical stress can bias cellulose deposition, possibly through microtubule orientation towards the direction of maximal tension, both at the tissue and single cell scales [8, 12–21]. The molecular mechanism behind remains largely unknown. Finally, cortical microtubules orientation may change in response to signals such as blue light or hormones, see for instance [17, 22] and may be oriented by the hydrodynamic forces due to cytoplasmic streaming [23, 24]. Here we use modeling to explore the relative contributions of cell geometry and external directional cues in the final microtubule organization.
The molecular basis for microtubule dynamics is rather well established. Consistent with the absence of centrosome in land plants, microtubule nucleation is dispersed in plant cells, as it occurs at the cell cortex [25], along existing microtubules during branching events [25–27], and at the nuclear envelope [28]. As they grow, microtubules form stiff and polar structures. They can alternate growth, pause and shrinking at the so-called plus end [29], whereas they mainly shrink or pause at the minus end [30]. The combination of an average shrinkage at the minus end and dynamic instability at the plus end leads to an overall displacement of the microtubule, also called hybrid treadmilling [30], with a dominant contribution of short treadmilling microtubules in the final microtubule organization [31]. The growth of microtubules in persistent directions is the main cause for microtubule encounters.
When one microtubule encounters another microtubule, different outcomes can be observed [9, 32]: if the encounter angle is shallow, zippering can occur, i.e. the growing microtubule bends and continues its polymerization along the encountered microtubule, which leads to the creation of microtubule bundles; if the encounter angle is steep, crossover can occur, i.e. a microtubule polymerizing without deviating its trajectory and crossing over the encountered microtubule; or alternatively catastrophe is triggered, i.e. a rapid plus end shrinkage after contact with the encountered microtubule. Such selective pruning of microtubules may explain how microtubules can form parallel arrays from initially random orientations, and conversely change the net orientation of their arrays over time, through a phase of randomisation [33, 34]. Selective pruning has indeed an essential ingredient of most models for microtubule dynamics [35–44]. The presence or absence of different microtubules associated proteins (MAPs) can modulate the stability of microtubules or their capacity to form bundles and to self-organize. For instance, the microtubule severing protein Katanin accounts for most of the pruning events at crossover sites [45].
The microtubule network is a typical example of a self-organizing system, where properties of individual elements and their interactions induce specific and sometimes counter-intuitive global properties. To predict how regulation at the level of each microtubule can give rise to specific global outcomes, one can resort to computational models. Modeling approaches have been developed, simplifying microtubule interactions by restricting them to the plasma membrane, i.e. a simpler 2D space [40, 46]. In those agent-based models, several microtubule properties were coded and interactions between CMTs, based on these properties, were simulated. The outcome is an emergent network, whose characteristics can be analyzed. For instance, increased microtubule severing was predicted to generate a larger number of free microtubules, more amenable to bundle into aligned arrays [9, 42] and this was observed in experiments [47].
So far, most of the microtubule models have been implemented in a 2D space or with microtubules confined to the surface of the cell. A major outcome of such models was to demonstrate that global orientations of the network can spontaneously emerge from the interactions between microtubules [48]. Many combinations of parameters and behaviors have been studied: instability at the plus end [35, 37], role of zippering [9, 36, 38, 39], nucleation modes [39, 42], and severing [44]. Beyond their differences, a global orientation emerges in most of these combinations suggesting that converging toward a global orientation is a robust feature of microtubule networks. Conversely, the diverse combinations of microtubule properties provide different scenarios for the fine-tuning of the network structure and stability of this emerging behavior.
Some aspects of cell geometry were related to microtubule behavior in certain simulations. Simulations showed how different directional biases in nucleation can induce an ordering of the array toward directions that are correlated to cell geometry [35, 43]. Further, branching nucleation rules can elicit handedness of the global direction of microtubule arrays, provided that the branching is biased toward one direction [41]. Other studies used a simulation space where borders, analog to cell edges, induce more or fewer catastrophe events or are more or less permissive toward microtubule growth [10, 33, 35, 41]. Most studies concluded that a global orientation of microtubules can be correlated to cell face orientations.
The contribution of the third dimension to microtubule behavior has started to be investigated. Computational models for animal systems have focused on 3D considerations but the nucleation hypotheses are too different from that in plants to be transposed directly [49, 50]. Fully 3D models suited for plants are still lacking: almost all existing studies have confined microtubules to surfaces embedded in 3D [10, 11, 41, 51]. In [35], a 2D model was extended into a full 3D model but it did not include cell boundaries, which yielded microtubules distributed over the whole simulated domain, in contrast with the cortical localization of microtubules in planta.
In this paper, we explore the influence of 3D cell shape on the basic properties of a dynamic microtubule network. We do not assume that microtubules are confined to the cell surface; rather, we simulate a closed volume where microtubules are more or less free to grow in all directions. Anchoring to the membrane is not imposed by the model and instead becomes a variable in the model. Using this framework, we investigate to what extent microtubule interaction with the membrane can influence microtubule dynamics. Our study also addresses the relative contributions of cell shape, microtubule interactions, and external directional cues in network organization.
Following previous studies, we modelled microtubules as a set of line segments that nucleate, grow, shrink, and interact with each other and with the cell surface represented as a triangular mesh (see Methods for details). Nucleation of the minus end occurs randomly at the surface. Growth occurs from the plus end with a small directional noise that is related to the persistence length of microtubules (Fig 1A). Shrinkage starts randomly at the nucleation site (minus end) and then continues at constant velocity.
A microtubule that encounters a preexisting microtubule either changes direction to that of the preexisting microtubule if the encounter angle is shallow (a process known as zippering, see Fig 1A), and otherwise starts shrinkage from the plus end (“head-on” collision). We considered two types of interaction with the cell surface: strong anchoring, whereby microtubules remain on the surface, as in previous studies [10, 11, 41, 51], and weak anchoring, whereby microtubules are prevented from leaving the cell interior. More specifically, in the case of weak anchoring, the interaction between a growing microtubule and the nearby surface is similar to the interaction between two microtubules: the microtubule encountering the surface at a steep angle starts shrinking, and otherwise starts growing tangentially to the surface (Fig 1A); a microtubule may leave the surface because of the directional noise.
We considered three base shapes: cube, “square” (flattened cube), and “long” (elongated small cube), of dimensions in the order of 10 μm, typical of plant cells. These shapes were smoothed so that the maximal curvature corresponded to a radius of either ∼1 μm (“sharp”) or ∼5 μm (“smooth”), which corresponds to typical radii of curvature measured in root epidermis [10], see Fig 1B. We also considered an ellipsoidal shape when investigating the effects of an external cue. Although these shapes are not fully realistic, they make it easier to disentangle the geometrical parameters influencing microtubule dynamics.
A typical simulation with weak anchoring is given in S1 Video. Corresponding snapshots are shown in Fig 1C and 1D with various 3D and 2D views. A few first observations can be made: the microtubules tend to bundle; a well-defined local orientation appears; microtubules appear to be mostly close to the cortex.
We first considered the effect of the anchoring of the microtubules to the membrane. There are proteins that have been shown to be associated with both microtubules and a plasma membrane component in plants [52]. For instance, CELLULOSE SYNTHASE INTERACTIVE PROTEIN 1 (CSI1) interacts with both CMTs and the cellulose synthase (CESA) complex [6, 53] and CSI1 was also proposed to stabilize the microtubule network [54]. Yet, the influence of CMT-CESA interactions on the microtubule network is still poorly understood. Thus, we took advantage of the 3D nature of our model to study the impact of the anchoring rule to the membrane on the global parameters of the microtubule network.
We investigated the microtubule dynamics when the anchoring to the membrane is weak. As a reference case, we also considered strong anchoring, whereby microtubule are constrained to grow on the membrane, as if putative anchoring proteins were highly concentrated. As expected, strong interactions led to a surface-localized cortical zone with microtubules trajectories embedded in the plane parallel to the mesh (Fig 2A).
In the case of weak anchoring, microtubules grow in all directions, but occasionally, as they encounter the membrane, their direction may be transiently tangent to the membrane. The typical distance between microtubules and membrane ranges from 50 to 250 nm (Fig 2A) according to the shape. Even if weak anchoring allows microtubules to grow through all the cell volume, we find that such weak interaction with the membrane is enough to elicit the existence of cortical microtubules. Therefore, the three-dimensional nature of our model helps us demonstrate that strong anchoring is not required for the presence of large populations of cortical microtubules in plant cells: the directional persistence, together microtubule growth mode, can cause such sub-cellular localization.
As the microtubules tend to stay at close to the membrane, they also bundle, with the proportion of tubulin in bundles varying from 30% to 75% (Fig 2B). The ability for the microtubule network to generate a spontaneous bundled structure is consistent with previous models constraining microtubules to the surface. Strikingly, this effect is also present in the case of weak anchoring.
Independently of the encounter rule, weak anchoring strength increases the total number and the size of microtubules (by about 20%, in length and in number) when compared to strong anchoring. Weak anchoring also yields less bundling (reduction of about 20%). A likely explanation is that a weak anchoring to the membrane allows microtubules to “escape” inside the cell, thus diminishing the encounter probability. Consequently, microtubules weakly bound to the membrane have more space to grow and are less subject to shrinkage-induced collisions or bundling with other microtubules (Fig 2B).
Next, we used our model to determine the consequences of changing cell shape on global properties of the network. We simulated the network in three main sharp shapes represented on Fig 1A. The results from the simulations indicate that these shapes only have a small effect on the number of microtubules (less than 10% difference), on the length of microtubules (5 to 15% difference) and on the proportion of bundles (less than 15% difference). Overall, elongated cells have more microtubules, more bundles, and longer microtubules, while cubes show the lower values (Fig 3). A likely explanation is that microtubules are more likely to follow the long axis in the long shape (see below) so that they are less affected by cell boundaries.
We also analyzed the effect of cell shape on the microtubule array anisotropy, averaged over the cell (see Methods); this is quantified with an order parameter with values between 0 and 1. As microtubule array anisotropy is skewed towards low values (found inside the cell) we used a non-parametric test based on ranks for statistical comparisons. Consistently with its effect on bundling (see above), we found that anchoring strength slightly affects the anisotropy of the microtubule arrays. Weak anchoring decreases the anisotropy of the network by about 10% (Fig 4C).
The type of cell shape did not appear to influence the global anisotropy of the microtubule network (Fig 4A). This is interesting as it suggests that different cell types with various shapes do not require differential regulation of the network in order to maintain the anisotropic properties of their CMT arrays.
However, smooth and sharp shapes differ significantly in anisotropy of the microtubule network (by 10%, p < 0.001); more curved cell edges in sharp shapes correspond to a lower anisotropy (Fig 4B). Larger surface curvature induce more “heads-on” collisions when microtubules grow nearby, leading to a smaller microtubule density, as they become more distant from the surface (Fig 2A); this would lead to more spatial variations in the orientation of microtubules and hence lower anisotropy.
In order to determine how the shape of the cell influences the global orientation of the network, we measured the distribution of orientations along two opposite faces of the cells. We chose two arbitrary faces for cube shape, the largest faces for square shape, and two of the largest faces for the long shape; accordingly, these two opposite faces are among the two largest faces in area for the shape of interest. An angle of 0° corresponds to the long axis in the case of the elongated cell. Angles of 90° or -90° correspond to directions perpendicular to that axis.
First, we observed that in the case of square shaped faces, most of the microtubules align along the cell face diagonal, i.e. the longest path (Fig 5). This occurred whatever the anchoring strength and the shape of faces at the side. Second, we observed a strong correlation between the axis of the cell and microtubule orientation, and this correlation is the highest for the elongated cell: The microtubule distribution is always either maximal or minimal at an angle of 0. This effect is higher in the case of strong anchoring, whereas in the case of weak anchoring, secondary peaks show that the diagonals are also overrepresented. These results indicate that the microtubule network is able to read the longest axis of the cell and orient toward that axis, by default. Interestingly, cortical microtubules become longitudinal in hypocotyl cells when growth stops [5, 55, 56], suggesting that they may adopt their configuration by default in that situation.
Next, we investigated the robustness of microtubule network. Microtubule arrays entirely reorganize during cell division [57]. Light and hormones can also completely reorient the microtubule network within minutes [33, 34, 58], suggesting that the constraints on microtubules array must not be too strong to allow such rapid reorganization in vivo. We tested whether our model provides such adaptability, using the case of external, directional cues (Fig 6 and S2 Video).
We investigated the effect of an external cue, assuming that, as microtubules grow in the cell, their growth direction is biased toward the direction of cue, with a specific weight. We used an ellipsoidal cell with a circumferential cue, which is orthoradial to the long axis of the cell. Such a cue could, for instance, be related to cell polarity (vector along the long axis) or could be a proxy of mechanical tension [19]. In the absence of the cue, microtubules are on average parallel to the main axis of the cell. Strikingly, our simulations indicate that even a very low bias (0.1%) could disrupt the main orientation of the network. When the weight reaches 1% or more, microtubules massively reorient toward the direction of the cue. Interestingly, the transition between the longitudinal and circumferential orientation occurs when the bias is comparable to the directional noise that is used to account for the persistence length of microtubules. This reinforces the idea that fluctuations of microtubules and the interactions between them lead to alignment with the long axis of the cell and that the small cue is sufficient to counteract these fluctuations and influence the orientation of microtubules. Accordingly, we found that anisotropy increases from no cue to a weight of about 1% and then seems to saturate (S1 Fig).
These results indicate that, despite an apparent robust organization, the microtubule network remains extremely sensitive to directional cues. As such, it is capable of reading slight directional cues and generating an ordered array in that direction. This ability to read directional cues is probably linked to the self-enhancement of microtubule orientation through their interactions: As more microtubules orient toward a direction, they prevent growth of microtubules in the perpendicular direction.
Our simulations indicate that the network should exhibit two behaviors concerning its orientation. When no external directional cue is present, the network orients toward the main axis of the cell, and generates a polarity that is a direct reading of the global shape. When a directional cue is present, the network reorients so as to emphasize the direction of the cue.
The impact of external cues on the microtubule network has been extensively characterized in experiments. However, because real cells and tissues are never really devoid of external cues, the behavior of the microtubule network by default in a plant cell remains an open question. Our work provides some clues to address this question, by taking into account the 3D shape of the cell and by using a minimal set of parameters for microtubule behavior. For instance, the model with weak anchoring does not require a specific rule for the crossing of microtubules (unlike when microtubules are confined to the cell surface), because microtubules naturally cross each other if they are farther than their diameter. We found that microtubule directional persistence largely determines the subcellular localization and orientation of the microtubule network in various cell shapes. We also identified parameters that seemed relatively insensitive to cell shape, such as microtubule length or number. Last, we found that microtubule dynamics yields at the same time stable orientation and high sensitivity to directional cues, even when such cues go against the default orientation. Altogether, this provides a conceptual framework to dissect the exact contribution of microtubule regulators to the microtubule network organization, in relation to 3D cell shape.
Based on TEM images where microtubules are often seen very close to the plasma membrane, it is assumed that anchoring of microtubules to the plasma membrane is relatively strong. However it remains unknown how such anchoring would be mediated [52]. Many biochemical studies have been performed to extract proteins that would link the plasma membrane to cortical microtubules, and so far the only published candidate is a phospholipase [59], for which no follow-up results have been obtained to the best of our knowledge. Other links have been put forward, such as CLASP [10] or CELLULOSE SYNTHASE INTERACTIVE PROTEIN 1 (CSI1) [53, 60], but they might represent rather indirect regulators of the link between microtubules and plasma membrane. Does this mean that microtubule could be cortical without any anchoring module? Our results suggest that in most of the measured cell shapes, microtubules do not need to be strongly connected to the membrane to remain cortical. This prediction implies that modulating anchoring would affect self-organizing properties tangentially to the cell surface, rather than modulating the density of microtubules inside the cell. This would allow molecular regulators to modify the microtubule organization without directly affecting the rate of cellulose deposition.
In our simulations, we observed that anisotropy varied according to curvature of face-face contacts of the cell, which is an emerging property of the weak anchoring mode. Cells that have sharper edges exhibit decreased array anisotropy compared to smoother cells. This result is interesting, as, for instance, epidermal cells possess edges with different curvatures [10]. Similarly, in the L1 layer of the shoot apical meristem, the outermost wall has a stronger curvature than all the walls separating the cell from its neighbors. Pavement cells also display a broad range of curvature values. Based on our results, higher anisotropy would be produced in L1 cells without the need for specific regulation.
Such results are difficult to account for in models with microtubules constrained to surfaces embedded in 3D, in which for instance additional hypotheses linking curvature and catastrophe rate are implemented [10]. Nevertheless, our results do not preclude microtubule associated proteins from additionally modulating microtubule dynamics according to curvature or to other membrane-localized cues [52]; we only account for the default behavior of microtubule networks according to surface curvature.
Similarly, changes in the curvature of the epidermal wall can occur through changes in internal pressure, which in return influence the anisotropy of the network [61]. An increase in the microtubule organization could be a result of an increased pressure, that would increase the curvature of the epidermal cell. Further experimental work is required to investigate the role of curvature on microtubule behavior and its relation to mechanical stress in the epidermis.
Our simulations show that the cell aspect ratio has an important impact on the global orientation of the network. The predicted default behavior of microtubules is their alignment parallel to the long axis of a cell, due to the directional persistence of microtubules. This is in agreement with previous models with microtubules confined to surfaces embedded in 3D [11, 51], where the default orientation is longitudinal for long cylinders.
This default state was observed with microtubules polymerizing in vitro inside elongated 3D chambers [11]. In slowly growing cells of the hypocotyl, microtubules are oriented along the long axis of the cell, whereas microtubules are circumferential in rapidly elongating cells [55, 56]. Our model suggests that directional cues are needed to avoid this default orientation is growing cells. At the boundary between the shoot apical meristem and the primordia, cell division leads to cell shapes that are elongated along the axis of the boundary; our model predicts that microtubules will be oriented along the same direction by default, amplifying their response to mechanical stress [19].
In this study, we show that the microtubule network is oriented by default along the longest axis of the cell. However, microtubules in plants often show supracellular orientation, independently of cell shape, a behavior that has been ascribed to tissue-level signals, notably mechanical stress [18–20]. Moreover, It has been demonstrated that inside the cell, microtubules orientation is coupled to polarity markers such as proteins from the PIN FORMED and RHO OF PLANTS families [62, 63]. Simulations have assessed how localized membrane heterogeneity could result in a biased orientation of the microtubule network [10, 41]. In this study, we show that a weak directional cue influencing microtubule growth rapidly modifies the orientation of the network towards the direction of the cue. This cue could be due to mechanical stress, hormone gradients [22], or to cytoplasmic streaming [23, 24], for instance. As such, the network behaves as a sensor translating an external directional information into a structural polarity inside the cell. The coexistence of a default orientation and a strong ability to reorient could shed a new light on orientation changes in cells. Changes in microtubule orientation need not be always related to specific regulation but may also be related to the arrest of signals and the return to the default state. This could be occurring in the shift from transverse to longitudinal orientation in hypocotyls responding to light or hormones [33, 34, 58].
The shape of the cell has little influence on mean length, number of microtubules or bundles proportions. In addition, anisotropy of the network is not highly correlated to changes in global cell shape. This prediction of a robust network suggests that plant cells do not need specific regulations to compensate for their great variations in shapes. Accordingly, the microtubule network appears as a good polarity system, with a default orientation and a high sensitivity to directional cues. It was recently shown that, for global polarity to emerge in a tissue, an important requirement is the existence of internal cellular polarity [64]. In this work we show that the microtubule network is suited for such a requirement.
Overall, microtubules and associated proteins form a complex self-organizing system that is difficult to comprehend without resorting to models. The results obtained here demonstrate that our three-dimensional model provides a framework to test hypotheses on the regulation of the microtubule cytoskeleton in plant cells.
The model given here is only the beginning to a more complete analysis. We have not yet incorporated microtubule severing [44, 58], microtubule branching [26, 27], nucleation at the nuclear envelope [28], or the possible effects of connections between cortical microtubules and cellulose fibrils outside of the cell as mediated by the cellulose synthase complex [65]. Severing in particular has been shown to be key to microtubule reorientation [44] following mechanical signals [66]. We also have not included limiting levels of tubulin [51], which could affect overall microtubule number. Altogether, we expect our model to help progress in understanding how microtubule self-organization integrates directional cues with three-dimensional cell shape and how microtubule-associated proteins modulate this integration.
The dynamical microtubule model was implemented in C++. The code is available from https://gitlab.com/slcu/teamHJ/vincent/microtubule_simulations. Simulations were performed on Intel/AMD desktop computers running Debian and Ubuntu operating systems.
Microtubules were coded as 3D multi-segment vectors of constant length. A ring of tubulin of the length of a dimer is represented as a unit vector in the simulation. Microtubule growth in the model occurs by adding one vector element at the plus end of the microtubule, at the position of the end of the last vector. In plants, microtubules are considered to be mostly static, their growth and shrinkage are the result of treadmilling processes. To code for microtubule directional persistence (which relates physically to bending stiffness), the direction at which a new vector is added to an existing microtubule changes by a small random amount. Microtubule shrinkage occurs at the minus end by removing the first vector from the list. The main model parameters are shown in Table 1.
Typically, a cell has a width of several micrometers, and we take the unit of length as 8nm, the height of a ring of tubulin. Considering a measured speed of growth at plus ends of 3-5 μm/min [67], a simulation time step is approximately 0.1 to 0.2 s. A typical simulation of 10000 time steps thus represents 15 to 30 min of real time.
The microtubules are nucleated on the cell surface [25–27] at a constant rate. The default value is np = 4.7 ⋅ 10−7 per time step per unit surface, corresponding to 1 to 2 nucleation events per time step, or 5 to 20 per second. Once nucleated, the microtubules do not immediately shrink. At each of the time steps that follow nucleation, a microtubule has a probability ns to begin shrinkage. Once a microtubule has started to shrink, one vector is removed from the minus end at each time step.
The cell contour is described with a triangular mesh. Each vertex is endowed with the information of the vector normal to the surface, which is used during the simulation to calculate a local approximation of the tangential plane. It is possible to add other informations at the vertex level that can be read during the simulation and serve as extrinsic input. Inputs can be scalars or tensors. The distance from any point in space to the membrane is calculated using the nearest point on the surface. At this point, the membrane is approximated by the plane perpendicular to the normal of the mesh. The distance between a tubulin ring (unit vector) and the membrane is calculated as the shortest distance between the endpoint of the vector and this plane. A collection of standard cell shapes was generated for our simulations (Figs 1 and 6). The main shapes were constructed starting from square parallelepipeds of dimensions 8.8μm×8.8μm×8.8μm (cube), 9μm×9μm×4.7μm (“square”), and 4.8μm×4.8μm×15.6μm (“long”), which are comparable to typical plant cell dimensions. Then the meshes were smoothed so that the minimal radius of curvature was 1.3μm and 4.7μm for sharp and smooth shapes, respectively, which roughly spans the range 0.5-5μm measured in plant roots [10]. The ellipsoid shape is an ellipsoid of revolution around the long axis; the short and long axis have dimensions of 10.3 μm and 16.8 μm, respectively.
The simulation progress is made through discrete timesteps. At each timestep new vectors are added and vectors are removed from the simulation space according to the rules specified in the previous subsections (Nucleation and minus end behavior; Plus end growth). Collision tests are performed so as to implement the different growth or shrinkage rules. In order to increase the simulation speed, space is divided into subelements and the vectors are identified according to the subelement to which they belong, which diminishes the number of particles involved in the collision test (locality-sensitive hashing [73]).
To visualize microtubule density and orientations an image is created using a matrix of resolution (rx, ry, rz). rz is larger than rx = ry by typically a ratio of 10 to 1, which mimics the anisotropic resolution of a confocal microscope. The simulation space is then screened. When a vector is located inside a cube of the matrix, the value of this cube is incremented by one, and the immediate neighbouring cubes are incremented by a lower number (typically 0.3). At the end of the process, a stack is formed where microtubules appear as blurred intensity signals. One can either visualize each sub-image from the stack by moving along the z axis, or create a projection that sums the matrix along the z axis.
We used the standard nematic order parameter to quantify anisotropy. The space is subdivided into cubes of arbitrary size, typically segmenting the structure into circa 27 cubes (S2 Fig). Segmenting the structure into 216 smaller pieces gives similar trends, with globally higher anisotropy values. All tubulin ring directions are extracted as a 3 × N matrix, D. We then compute the square symmetric (3 × 3) matrix M = DT ⋅ D/N, where T stands for the transpose. M is diagonalised, yielding three eigenvalues λi, i ∈ 1, 2, 3. Local anisotropy, A, of the microtubule array in each cube of space is defined as
A=32(λ1−λm)2+(λ2−λm)2+(λ3−λm)2λ12+λ22+λ32;λm=λ1+λ2+λ33. (3)
The value of A is such that 0 ≤ A ≤ 1. Anisotropy value is then computed as the average of the local value A over the whole cell.
The simulations were run 5 times for each parameter value or shape considered; to avoid artificial correlations between data, only the last snapshot was considered for further statistical analysis. As data for the three default values of np were statistically identical, they were pooled together for Figs 1–5. The plots were produced with the boxplot function of R: the boxes extend between the first and third quartiles, the segments in the box indicate the medians, and the whiskers are representative of extreme values.
|
10.1371/journal.pcbi.1005704 | Host population structure impedes reversion to drug sensitivity after discontinuation of treatment | Intense use of antibiotics for the treatment of diseases such as tuberculosis, malaria, Staphylococcus aureus or gonorrhea has led to rapidly increasing population levels of drug resistance. This has generally necessitated a switch to new drugs and the discontinuation of older ones, after which resistance often only declines slowly or even persists indefinitely. These long-term effects are usually ascribed to low fitness costs of resistance in absence of the drug. Here we show that structure in the host population, in particular heterogeneity in number of contacts, also plays an important role in the reversion dynamics. Host contact structure acts both during the phase of intense treatment, leading to non-random distributions of the resistant strain among the infected population, and after the discontinuation of the drug, by affecting the competition dynamics resulting in a mitigation of fitness advantages. As a consequence, we observe both a lower rate of reversion and a lower probability that reversion to sensitivity on the population level occurs after treatment is stopped. Our simulations show that the impact of heterogeneity in the host structure is maximal in the biologically most plausible parameter range, namely when fitness costs of resistance are small.
| The rising levels of drug resistance in many human infections are cause of great concern for public health. There is a repeating pattern of introduction of new drugs, rise of resistance to these drugs, and phasing out ineffective drugs once resistance has become common. With a decreasing rate of drug discovery it is important to study the dynamics of reversion back to sensitivity for drugs that are no longer in use in the host population. While it is known that fitness cost of resistance plays an important role in this reversion process, this study is the first to show that structure in the host population also heavily impacts the reversion dynamics.
| The emergence of resistance of infectious pathogens to antimicrobial drugs is a growing concern for public health. The control of many infectious diseases such as tuberculosis, malaria, HIV, or gonorrhea show a recurring historical pattern of the introduction of new potent drugs that initially control the infection efficiently, followed by the subsequent evolution of resistance once the drugs are used widely [1–3]. Over time, a combination of various factors, such as decreases in efficacy, adverse side effects, difficulties in administration, and overall economic costs may lead to changes in treatment guidelines. What generally follows is the successive introduction of novel drugs and the subsequent phasing out of old ones.
Such a pattern of drug use has important consequences for the evolutionary dynamics of resistance. During phases of intense use, the emergence and subsequent spread of resistant strains is driven by the large selective advantage of the resistant over the sensitive strains in presence of treatment [4]. This has been the case in many diseases, such as tuberculosis [5], malaria [6], Staphylococcus aureus [7], gonorrhea [8, 9] and HIV [10]. If drug use is discontinued due to changes in treatment guidelines or other factors we generally observe either of two scenarios: One, the resistant strain may decrease in frequency as a result of the change in selective pressure and the population reverts back to high levels of drug sensitivity. For example, in Malawi the clinical efficacy of the antimalarial chloroquine had fallen to low levels because of a high frequency of resistance. But efficacy increased again during the 12 years after the cessation of its use in 1993 from less than 50% to 99% in 2005 [11, 12]. Two, the resistant strain may continue to persist for prolonged periods of time despite substantial reductions in drug use. Examples include the resistance to several antibiotics in Neisseria gonorrhoeae [13], streptomycin resistance in tuberculosis [14, 15], sulphonamide and trimethroprim resistance in Escherichia coli [16, 17], as well as the persistence of vancomycin resistant enterococci in pigs after the ban of avoparcin in 1995 [18, 19].
Multiple factors have been proposed that might impede reversion back to sensitivity. Firstly, if the reduction in drug use is not substantial enough or only affects a subpopulation of hosts, then the residual use of drugs in the population can sustain the selection for resistance. Secondly, the absolute differences in fitness between resistant and sensitive strains in absence of treatment are generally considerably smaller than in presence of treatment [4]. This asymmetry in fitness differences between the absence and presence of treatment can further be exacerbated by compensatory mutations that alleviate fitness costs associated with resistance mutations [20–22]. Finally, reversion in vitro is impeded by genetic interactions between resistance and compensatory mutations. Such interactions can obstruct reversion back to sensitivity, because reverting only the resistance mutation without reverting the compensatory mutation or vice versa is associated with a fitness decline in the absence of drugs [23]. However, the obstruction of reversion back to sensitivity due to such fitness valleys is only expected to be relevant in situations where the sensitive strain has to re-emerge de novo from the compensated resistant strain by mutation. As changes in treatment guidelines typically occur much before the resistant strain has fixed in the pathogen population across hosts, we expect that the crossing of fitness valleys is likely not relevant for the reversion back to sensitivity in epidemiological scenarios. For malaria in Malawi, for example, the reversion is believed to have occurred through a re-expansion of the susceptible parasite in the population and not through de novo back mutation [24].
In the context of antibiotic resistance, Johnsen et al. [25] listed further effects that may contribute to the persistence of resistance in absence of drugs. These include selection of other beneficial traits genetically linked to the resistance gene, the role of reacquisition of resistance through horizontal gene transfer and mechanisms preventing plasmid loss. While all of these factors plausibly contribute to obstructing or slowing the reversion process [26], it is difficult to conclusively demonstrate which factors are at play for any particular resistant pathogen population or even to demonstrate that together they are sufficient to explain a slow or absent reversion.
Common to all these factors is the pathogen-centric view with only little attention given to the role of the host. In particular, host contact structure may be another important factor modulating the evolutionary dynamics of resistance. A large body of theory has shown that contact structure profoundly effects epidemiological dynamics [27–32]. Furthermore, Lieberman et al. [33] showed that host structure can affect evolutionary dynamics under a Moran process [34], by modulating the relative importance of selection and random drift. Similarly, both theory and experimentation have established that spatial structure, a specific form of contact structure, influences the evolution of virulence [35–39].
Another element that is absent in the pathogen-centric view is the between-host transmission of resistance. Studies suggest that a resistant strain with only a small fitness deficit in the absence of treatment impacts the overall course of an epidemic in a manner largely independent of its probability of de novo emergence [40–42]. Hence, transmission of resistance cannot be neglected and it is particularly important to understand how resistant pathogen strains compete with the wild type throughout the course of an epidemic.
The epidemiological dynamics of wild-type and resistant strains can be seen as a special case of two distinct pathogen strains that are simultaneously spreading in the same host population. Such cases have been studied theoretically for self-limiting dynamics (e.g. SIR-type models), where infected hosts are removed from the population upon recovery, ultimately leading to a depletion of available hosts. Under these dynamics it is possible for two pathogens to both cause an epidemic when spreading sequentially in the same host population, even if infection with the first pathogen confers immunity towards the second [43]. Further studies generalized this finding to cases where the second strain starts to spread simultaneously [44] or with only a small delay [45].
The successful spread of the second pathogen depends on the residual network of susceptible hosts that remains after the first pathogen has spread. In host populations with heterogeneous contact distributions, a host’s likelihood of infection increases with the number of its contacts [32]. This leads to a specific structure of infected hosts in the population and, as a consequence, the residual network left for the second pathogen to spread on does not represent a uniform sub-sample of the entire network [46]. Bansal et al. [46] showed that for random networks with fixed average number of contacts an increase in contact heterogeneity leads to a decrease in the epidemic size of the second pathogen.
The concept of residual networks is less clear for non-self-limiting dynamics (e.g. SIS models), where an infected host returns to a susceptible state after recovery. In such models, the mere advantage of spreading more rapidly is less evident as competition arises through co-existence in a continuous epidemic. We have recently shown that in the case of a continuous epidemic heterogeneity in contact structure impedes the invasion of a second fitter pathogen when starting from a single individual in the population [47].
To be able to understand the evolution of resistance in the context of treatment, it is crucial to take into account the two mechanisms of how resistance increases on a population level: either through de novo emergence in treated patients or through transmission [48]. A first step was made by Hébert-Dufresne et al. [49] who extended models of self-limiting dynamics to incorporate treatment and treatment failure leading to de novo emergence of resistance. They showed that even small changes in the ratio of the strains’ fitnesses can drastically affect the total epidemic size.
Here, we propose an epidemiological modelling framework with explicit host contact structure to study non-self-limiting dynamics. This framework allows studying for an endemic disease both the process of resistance emergence during treatment, as well as the subsequent competition dynamics between resistant and sensitive strains after treatment is discontinued. We first describe the general dynamics of a non-self-limiting susceptible-infected-susceptible (SIS) model on a heterogeneous contact network prior to, during and after treatment. Then, we assess the likelihood of reversion to a drug sensitive wild-type population after stopping treatment. Finally, we investigate how the spread of resistance is influenced by the network structure and the colonization of the network by the resistant strain during treatment. Our simulations confirm that contact heterogeneity lowers the probability of reversion back to sensitivity, even when a substantial fraction of the pathogen population is sensitive at the time point when treatment is stopped. Our study reveals that large fitness differences between sensitive and resistant strains result in a non-trivial distribution of sensitive and resistant infections over the network, which in turn influences the reversion dynamics once treatment is stopped. Importantly, the modulating effects of host contact structure on the probability of reversion is strongest in the biologically relevant case when fitness differences between sensitive and resistant strains in absence of treatment are small.
We investigated the spreading and competition dynamics of sensitive (wild-type) and resistant infectious disease strains using a model of disease spread in a host population with heterogeneous contact structure in presence and absence of treatment (see Methods).
Fig 1 illustrates the distinct phases of the spread and competition of the strains before, during and after treatment.
Initially, the sensitive strain spreads in a fully susceptible host population and reaches an endemic state. Treatment is then initiated on a population level. As a consequence of treatment, new infections with the sensitive strain begin to decline in the host population. Within an individual, however, treatment selects for resistant strains. These can be generated de novo within a treated individual and then be transmitted in the host population. Resistant strains are assumed not to be affected by treatment. Hence during treatment the fraction of resistant infections increases in the host population, while the fraction of sensitive infections further declines. Once the fraction of resistant infections has risen to a critical level, treatment is stopped in our simulations. From this point onwards, there is no further generation of de novo resistance and the dynamics are governed by the competition of wild-type and resistant strains in absence of treatment. We assume that wild-type and resistant strains can coexist on the level of the host population but not within a single individual. This leads to a competitive exclusion between the strains and therefore competition for the hosts on the between-host level. This assumption completely excludes co-infections and any resulting within-host competition, which allows us to study the impact of the host contact structure in an isolated manner and obtain a clear outcome, i.e. extinction of either of the two strains.
In the long run either reversion to the wild-type or fixation of the resistant strain is observed in the host population. We define the ‘probability of reversion’ back to the wild-type strain, Prev, as the fraction of simulations in which the resistant strain goes extinct after stopping treatment.
The probability of reversion back to the wild type decreases with increasing relative fitness of the resistant strain both for high and low critical fractions of resistance, fr (Fig 2a and 2b for fr = 0.5 and fr = 0.1).
In general, the probability of reversion as a function of the relative fitness has a sigmoidal shape, with an inflection point at a relative fitness of sA = 1. For low enough relative fitness of the resistant strain (sA < 0.97), reversion to the wild type happens almost certainly (Prev ≈ 1). Equivalently, for high enough relative fitness (sA > 1.03) the resistant strain almost certainly goes to fixation (Prev ≈ 0) (see S1 Fig). The transition from almost certain reversion to the wild type, to almost certain fixation of the resistant strain is less steep for simulated host contact networks than for a fully mixed population and depends on the variance in degree of the network. An increase in variance makes the transition more smooth, i.e. results in a decrease in the probability of reversion for sA < 1, and an increase in probability of reversion for sA > 1. As a reference, we report the probability of reversion in the case of a Moran process on a random host population with homogeneous degree, P rev * = 1 - s A - f r N 1 - s A - N (solid black line in Fig 2a and 2b; ref. [50]). Note, that pathogen competition is not expected to directly match this type of process, as the population of infected hosts is a dynamically changing sub-sample of the entire population: In contrast to the Moran process, transmission does not occur within this sub-sample, but exclusively between the infecteds and the rest of the population. In the limit of large transmission rates, however, the population of infected hosts extends to the entire host population and it can be shown that we recover the Moran process (see S2 Text). In the terminology of a Moran process the previous observation reads: increasing variance in the degree distribution of the host network decreases the effects of selection and hence favours random drift.
The treatment halt on the population level is implemented by not providing treatment for any newly infected individuals, but allowing the individuals currently on treatment to finish their protocol. We refer to this scenario as gradual treatment halt. In a more instantaneous scenario, treatment is stopped in all patients at the same time, leading to a discontinuation of the treatment in infected individuals. We refer to this scenario as immediate treatment halt. A gradual treatment halt generally decreases the probability of reversion compared to an immediate treatment halt (Fig 2c and 2d). This is expected, since a gradual treatment halt will continue to disfavour the wild-type compared to the resistant strain, because some individuals remain on treatment beyond the end of the treatment phase and therefore have a reduced transmission rate.
The range of relative fitness spanning the transition from almost certain reversion to almost certain extinction narrows with increasing population size (Fig 2e) leading to a step-like transition in the limit of large system size. To assess the impact of connection density we report the probability of reversion for systems with fixed size and zero variance but different mean in the degree distribution (Fig 2f). We compensate for changes in the connection density by adapting the transmission rate such that the epidemic threshold is kept constant, i.e. all systems depict the same basic reproductive ratio of R0 = 3. Tuning the connection density in this manner has no impact on the reversion probability for relative fitness values that lead to almost certain reversion (sA < 0.97). For relative fitness values closer to one, an increased density in contacts leads to a reduction of the probability of reversion.
To further disentangle the effect of network heterogeneity from the effects of relative fitness, we isolated the contribution of variance in degree by comparing the probability of reversion in host networks with non-zero variance to networks with zero variance. Here we consider the case of a gradual treatment halt (see S2 Fig for the scenario with an immediate treatment halt). An increased variance in degree decreases the probability of reversion most strongly for small fitness differences between the wild-type and the resistant strains in absence of treatment (Fig 3a). Because the costs of resistance are generally small for many pathogen-drug combinations, the biologically most relevant parameter range is the range where the effects of contact heterogeneity are expected to be largest. The range of relative fitness values for which variance in degree affects the reversion probability straitens and shifts closer to sA = 1 with increasing system size (Fig 3d). Increasing the contact density generally reduces the effect of degree variance on the reversion probability (Fig 3e).
Network heterogeneity can influence which strain takes over in two ways: directly, by modulating the competition dynamics in the post-treatment phase; or indirectly, by influencing the positioning of the wild-type and resistant strains as a result of the competition during the treatment phase. To assess the direct impact of variance, we randomized the distribution of wild-type and resistant strains among the infected individuals at the end of the treatment phase and again compared the probability of reversion for networks with non-zero variance to networks with zero variance (Fig 3b). The reduction in probability of reversion persists when randomizing within infected individuals of the networks, indicating that host population structure directly modulates the competition dynamics during the post treatment phase. To assess the indirect impact, we compared the reversion probabilities from these randomised distributions (Fig 3c) to their non-randomised counterparts (Fig 3a). Generally, shuffling the distribution of the resistant and wild-type strains has little effect on the probability of reversion (Fig 3c). Only for excessively small fitness differences (sA > 0.995) does the distribution of strains among the infected individuals at the end of the treatment phase additionally favour the wild-type strain. We note that this slightly beneficial effect on the reversion probability increases with increasing relative fitness of the resistant strain. While the distribution of resistant and wild-type strains at the end of treatment shows a small effect on the probability of reversion, contact heterogeneity predominantly modulates the competition dynamics directly during the post-treatment phase.
To get a better understanding of the underlying processes that influence the competition after treatment halt, we investigate the time dynamics of the relative prevalence of the wild-type strain during the post treatment phase. Fig 3f shows mean (dotted lines) and standard deviation (borders of colored areas) of the relative prevalence of the wild-type strain during the post treatment phase after an immediate treatment halt. The simulated mean relative prevalence increasingly deviates from the pair approximation (solid lines) with increasing variance in degree.
If exclusively simulations that revert back to the wild type are considered (see S3 Fig), eventual reversion occurs faster with increasing degree variance. This indicates an increase of reversion rate with increasing degree variance and is consistent with results from an analytical approach using a pairwise approximation model (S4 Fig and S1 Text for further details on the two-strain pairwise approximation approach).
If all simulations are considered, increasing variance leads to slower reversion, inverting the trend observed both in the analytical solution and when considering reverting simulations only. This inversion of the effect of degree variance on the reversion rate can be understood when considering the standard deviation of the relative prevalence (boarder of colored areas in Fig 3f): Increased degree variance leads to an increased standard deviation which, in turn, increases the chance for the wild-type strain to go extinct during the post-treatment phase despite its fitness advantage. As a consequence we observe both lower probability of reversion and a lower mean relative prevalence in networks with high degree variance. Hence, even with an initial abundance of 90% of the wild-type strain, stochastic effects play an important role in the process of reversion. Contact heterogeneity in the host population increases the magnitude of these stochastic effects resulting in a prolonged phase of co-existence of the competing strains.
To assess the indirect effect of placement of wild-type and resistant strains in the network at the end of the treatment phase, we counted the number of contacts between pairs of individuals of various infection status (Fig 4a).
Randomizing the strains among infected individuals keeps the overall number of susceptible/infected pairs constant, but destroys any non-random pattern of occupancy. Thus, a change in the fraction of a specific type of pair in the non-randomized occupancies could explain the differences in outcome between the randomized and non-randomized cases. The fraction of wild type/wild type and resistant/resistant pairs is higher in the non-randomized occupancies than in the randomized occupancies, while the fraction of wild type/resistant pairs is lower in the non-randomized occupancies (Fig 4a). This indicates a tendency for the resistant strains to aggregate during the treatment phase. Additionally, the fraction of susceptible/resistant pairs is also lower in the non-randomized occupancies (Fig 4a). From the point of view of susceptible individuals, the wild-type strain therefore acts as an insulator from the resistant strain, leading to an advantage of the wild-type strain and thus a higher probability of reversion. During the treatment phase, the wild-type strain is heavily disfavored, i.e. sA ≫ sP, such that a susceptible individual in contact with a wild-type infected is less likely to contract the diseases than in the case of a susceptible/resistant pair. A higher density of susceptible individuals is thus expected in regions of the host network that are dominated by the wild-type strain. Interestingly, this effect diminishes for networks with higher variance in degree.
For host networks with non-zero variance in degree, a strain gains an advantage from both an increased exposure to susceptible hosts, as well as from occupying individuals with a high degree. Infected individuals generally have a higher degree than randomly chosen individuals in the network (Fig 4b). Additionally, individuals infected with the resistant strain have an even higher degree on average. During the treatment phase, the fitness advantage of the resistant strain thus allows it to occupy nodes with high degree in the network, leading to an advantage independent of the positioning relative to the wild-type infected and susceptible hosts.
The spread of the resistant type over the network during the treatment phase thus generates two effects that influence the reversion. Firstly, there is a difference between resistant and wild-type strains in terms of their likelihood to be connected to a susceptible neighbor and thus their potential to infect a given neighbour. Secondly, the nodes occupied by resistant and wild-type strains differ in their degree and thus the resistant type has more neighbours to spread to. The first effect favours the wild type, while the second effect favours the resistant during reversion. With increasing variance in contact structure the first effect is weakened (Fig 4a) and the second strengthened (Fig 4b), resulting in an overall lower probability of reversion for high variance.
The difference in mean degree between wild-type and resistant infecteds is consistent over a vast range of treatment coverage (Fig 4c) and robust with regard to the fraction of resistants at treatment halt (Fig 4d). Also, the effect persists across a wide range of de novo rates of resistance, and breaks down only at rates that are unrealistically high.
Our simulations reveal the following key results: i. Increasing variance in contact structure lowers the probability of reversion. ii. Stochastic effects dominate the competition phase after treatment halt even when both strains represent a substantial fraction of the population. iii. The distribution of infected individuals at the end of the treatment phase is highly specific and influences the reversion dynamics.
Given the inherent simplifications of a random network with a heterogeneous distribution of contacts, we caution against taking our quantitative results as representative for the magnitude of effects to be expected in real contact networks. We expect, however, that in real systems, qualitatively the mentioned effects to be present.
From an evolutionary perspective, i. suggests that treatment in a homogeneous host contact structure would lead to a stronger selection of resistant strains with increased transmissibility as compared to treatment in a more heterogeneous host population. It has been repeatedly hypothesised that an increase in virulence correlates with an increase in transmissibilty [51–54]. Under this assumption, our findings extend conclusions from earlier studies suggesting that more homogeneous contact patterns [46] and increased global connectivity in spatially structured populations [36–38] enhance the evolution of virulence: homogeneity in the host population structure fosters the co-selection of resistance and virulence under treatment.
What is responsible for the effect of variance on reversion probability? Previously it has been shown that the fitter mutant has a disadvantage when invading from a single individual into a resident population of the wild type [47]. This effect may, at least in part, be due to stochastic effects in small populations [55]. In our simulations, however, the fitter variant (i.e. the wild type) is present at high frequencies of 50% or even 90% where stochastic effects due to small population size are negligible. Another potential explanation would be that the absolute prevalence at treatment stop affects the probability of reversion and that variance in the host degree distribution merely acts on the absolute prevalence. We tested this possibility but found no evidence that changes in the order of 5% to 10% in total prevalence at treatment halt had any effect on the probability of reversion (see S3 Text). The randomizations of the network occupancy at the end of the treatment phase show that contact heterogeneity in the host structure does not only impact the configuration of resistant and wild-type infecteds at the end of treatment, but also shapes the competition dynamics during the post treatment phase (Fig 3). In fact, for a large range of frequencies of the fitter strain, most of the observed effect of variance in degree on the probability of reversion is due to the increase in stochastic effects during this competition phase. Recent works [56, 57] have addressed the prolonged co-existence of wild-type and resistant strains observed in many real-world diseases. The shift towards random drift, i.e. mitigation of selection pressures resulting from fitness differences when degree heterogeneity increases, is an additional factor favoring a prolonged co-existence. It is important to note that the stochastic nature of the reversion process does not primarily result from effects of small population size, but rather the small fitness differences between the strains, a property that is found in real systems. Thus ii. is in contrast to what has been reported in the case of self-limiting dynamics, where the outcome of the simultaneous spread of two pathogens is largely determined by the initial proportions of the pathogens [44]. This is also reflected by the decent quality of predictions from analytical models [46, 58] for self-limiting dynamics.
The specific distribution of infecteds at the end of treatment, i.e. result iii., can be characterized by two distinct features in the occupancy patterns, both affecting reversion (Fig 4): First, we observe that individuals infected with the resistant strain tend to be aggregated at the end of the treatment phase. This implies that at the beginning of the reversion phase the resistant infected individuals have a lower per contact probability to be connected to a susceptible individual thus favouring reversion to wild type. This observation is surprising, given that the underlying model for the contact structure is a random graph providing no topological support for aggregation. We thus hypothesize this aggregation tendency to potentially have a much bigger impact on structures that provide non-random topologies.
Second, it has previously been shown that the sub-population of infected individuals tends to have a higher average degree [32]. We find that that the same holds true for the resistant sub-population among the infecteds: At the end of treatment, individuals infected with the resistant strain tend to have a higher mean number of contacts than those infected with the wild-type strain. The higher number of contacts leads to an increased chance of transmission for the resistant strain and thus lowers the probability of reversion.
The first effect promotes reversion while the second effect obstructs it. However, both share a common tendency with increasing degree variance in the networks: (a) the tendency of resistant strain to aggregate decreases, reducing the positive effect on the probability of reversion and (b) the difference in mean contact number of resistant and wild-type infecteds increases, further favoring the resistant strain.
Combining the above described effects leads to two main conclusions: First, an increase in heterogeneity in the host network diminishes the probability of reversion back to the wild type. Second, the impact of heterogeneity is most pronounced for small fitness differences between the resistant and the wild-type strain in the absence of treatment.
Our findings consider a competition process of two strains under a SIS-dynamics with a first-come-first-serve exclusion on the level of a single host. It goes without saying, that the observed competition and its outcome are consequences of this particular model choice. The SIS-type model—the simplest mathematical model of an endemic disease [68]—allows to study the effect of host heterogeneity on a continual pathogen evolution. It is, however, not suitable to investigate the case of successive single-wave outbreaks in a population that is partially immune. The effect of heterogeneity in the host structure on pathogen evolution in the case of single-wave dynamics has been covered by several studies [43–46, 49]. They make use of SIR-type models which are more suitable for this type of outbreaks. SIR-type models would require additional modeling of demographic changes in the host population for an endemic state to be possible.
The first-come-first-serve exclusion is an extreme case of interaction amongst the strains as it excludes both the possibility of a simultaneous infection of a single host by both strains (co-infection) and a displacement of one strain by the other in a currently infected host (super-infection). Conversely, a complete absence of interaction will lead to simultaneous but independent epidemics of both strains. In-between those two extremes, strain interactions can occur in a multitude of forms and will depend on the particular real world disease.
The low fitness cost of resistance mutations is likely one of the key factors contributing to the observation that reversion to wild type is often slow or even absent. Interestingly, the network effects we observe here are most pronounced exactly at these small fitness differences. The reason why fitness costs are often small is that the potentially larger direct costs of resistance conferring mutations are often alleviated by compensatory mutations [20–22]. Reversion from a genotype that carries both resistance and compensatory mutations may require crossing a fitness valley, which obstructs the reemergence of sensitivity, as has been shown in in vitro studies [23]. This argument, however, only applies when the wild-type strain has become extinct. While extinction of the wild type might occur in individual patients, it is unlikely to occur on the epidemiological level. As soon as there is even a small fraction of wild type in the population, the outcome of the dynamics is a matter of competition, and does not require de novo reemergence of the sensitive wild type. Although the mechanism is entirely different, treatment on the population level with heterogeneous host contact patterns leads to a similar phenomenon: the path to resistance is easier than the way back. The network structure of the host population obstructs the reversion back to the wild type. The small fitness difference between resistant and sensitive strains in absence of treatment not only result in a low reversion rate, but, together with heterogeneity in contact number, the small fitness difference additionally results in a reduced probability of reversion.
Individual variations in contribution to disease transmission were shown to be present across the bard of infectious diseases [59, 60]. This is particularly important for sexually transmitted diseases where individual variations are thought to occur through differences in patterns of sexual-partner acquisition [61–65] thus leading to heterogeneous contact structures.
Given this generally heterogeneous nature of host transmission patterns and the maximal impact of the here described network effects in biologically relevant fitness differences, we expect our findings to be of general relevance. It would be interesting to test on the basis of simulations on realistic real world networks the strength of the here described effects. The other option is to explore by simulations what type of networks would show these effects most strongly. Both directions we decided, for reasons of scope, to leave for further studies.
We modelled heterogeneous contact structure in the population using random networks with degrees distributed according to a discretized gamma distribution. This allows us to keep the mean fixed but tune the variance in number of contacts per individual. For a given mean μ and variance σ2 the scale θ and shape k parameter of the gamma distribution is given by
θ = σ 2 μ (1) k = μ 2 σ 2 (2)
From the generated distribution we then draw for each node in the network a value and round it to define its degree. We then use a stub connecting algorithm to generate a contact network [66].
The SIS dynamics on the network are simulated using the Gillespie next reaction method [67]. In brief, starting from the first infected individual in the network we draw the duration of its infection from an exponential distribution with recovery rate parameter γ and the times to infection of all its neighbors from an exponential distribution with transmission rate parameter β. Then, we record the time of recovery as well as those time points of infection that occurred prior to recovery of the individual in a queue of events. The algorithm then proceeds to the next event in the queue. In case this event is an infection and the node to infect is susceptible, the above procedure is repeated. Note, that the condition to infect only susceptible nodes implies that there is no super-infection. In case the event is a recovery, the status of the infected node is reset to susceptible.
During treatment, infected individuals either do or do not receive treatment with a probability c, the treatment coverage. Treatment has two consequences: Firstly, during the time period in which an individual receives treatment, the infection rate is reduced by a factor 1 − e, where e is the efficacy of treatment in preventing transmission. Secondly, treatment can change a wild-type infection into a resistant one. A time point for such an event is determined according to an exponential distribution with a rate of de novo emergence r. In case this event occurs prior to recovery, it replaces the recovery event in the queue, the status of the node is changed from wild-type to resistant, and a new recovery event for this node is generated according to the parameters of the resistant strain. All transmission events after the de novo emergence are discarded and replaced by new transmission events generated according to the transmission rate of the resistant strain βres.
The treatment phase ends when the fraction of resistant infections among the infecteds reaches a value fr. We implement the end of treatment in the population considering two scenarios: In the gradual treatment halt scenario, patients who are already on treatment continue therapy, but no newly infected individuals receive treatment. In the immediate treatment halt scenario, treatment is stopped in all individuals simultaneously. The immediate treatment halt necessitates that all transmission events are discarded at the start of treatment and are re-generated with updated transmission parameters. The gradual treatment halt implies that the transmission rate parameter is altered in newly infecteds only.
The simulations proceed though 3 distinct phases: The initial phase, the treatment phase and the post treatment phase. At the end of each phase the host contact networks, along with its epidemic state is saved.
In the initial phase, we first infect a randomly chosen individual in the population with the wild-type strain. This individual then infects neighbors at a rate βwt over the duration of its infection, and recovers at rate γ. The initial spreading phase ends when the frequency of wild-type infected individuals reaches a quasi-steady state after a sufficiently long burn-in phase. In case the infection dies out before quasi steady state, the simulation is restarted. Once the quasi steady state is reached, we halt the simulation and store the network and its epidemic state. Subsequently, the treatment phase starts. Wild-type infected individuals that receive treatment infect at a rate βwt(1 − e), individuals infected with the resistant strain transmit at a rate βres, and recover with the same rate, γ, as wild-type infecteds. We assume that the cost of resistance is small relative to the effect of treatment on the wild-type strain and thus approximate the average relative fitness of resistant versus wild-type strains during the treatment phase as a function of the treatment only: sP = βres/(βwt(c(1 − e) + (1 − c))) ≈ 1/(1 − ce). After treatment is stopped, the post treatment phase starts. In this phase the simulations continue until either of the two strains completely disappears from the population. Here the fitness of the resistant relative to the wild-type strain is simply given by the ratio of their transmission rates, sA = βres/βwt.
Unless otherwise specified, we used a network size of N = 2000 and a mean degree of μ = 4 at varying levels of variance, σ2 = 0, 1, 2, 4, 6, 10, 16, 24. For each variance level we generate 1000 networks. The transmission and recovery rate are chosen such that the resulting basic reproductive ratio in the case of zero variance is: R 0 = ( μ + σ 2 μ - 1 ) β wt / γ = 3 [68]. The initial phase is run twice on each generated network. The treatment phase is then run three independent times for all saved states of the initial phase and for all sets of treatment parameter. We choose treatment efficacy, e = 0.5, and complete treatment coverage, c = 1, throughout. The de novo emergence rate is, if not specified otherwise, r = 0.0001. The critical levels at which treatment is halted are fr = 0.1 and 0.5. The post treatment phase is then run on each output of the treatment phase and for each transmission rate of the resistant strain. In this manner we end up with at least 5000 simulations of the reversion dynamics, this for each parameter combination.
For the resistant strain a set of transmission rates is chosen such that the relative fitness of the resistant strain SA = βres/βwt ranges from 0.975 to 1.025.
To test the effect of network occupancy on the probability of reversion to wild type we shuffled the status of wild-type versus resistant infected individuals at the end of the treatment phase, therefore keeping fr constant. When shuffling at the end of the treatment phase, we first discard all future infection events. Then we shuffle the status within all infecteds, with the status of an individual being a pathogen strain and a time to recover. Finally, we redraw infection events for all neighbours of the infected individuals.
EndemicPy, the software package used in this study is freely available on GitHub [69].
|
10.1371/journal.ppat.1003860 | Single Cell Stochastic Regulation of Pilus Phase Variation by an Attenuation-like Mechanism | The molecular triggers leading to virulence of a number of human-adapted commensal bacteria such as Streptococcus gallolyticus are largely unknown. This opportunistic pathogen is responsible for endocarditis in the elderly and associated with colorectal cancer. Colonization of damaged host tissues with exposed collagen, such as cardiac valves and pre-cancerous polyps, is mediated by appendages referred to as Pil1 pili. Populations of S. gallolyticus are heterogeneous with the majority of cells weakly piliated while a smaller fraction is hyper piliated. We provide genetic evidences that heterogeneous pil1 expression depends on a phase variation mechanism involving addition/deletion of GCAGA repeats that modifies the length of an upstream leader peptide. Synthesis of longer leader peptides potentiates the transcription of the pil1 genes through ribosome-induced destabilization of a premature stem-loop transcription terminator. This study describes, at the molecular level, a new regulatory mechanism combining phase variation in a leader peptide-encoding gene and transcription attenuation. This simple and robust mechanism controls a stochastic heterogeneous pilus expression, which is important for evading the host immune system while ensuring optimal tissue colonization.
| Streptococcus gallolyticus (formely known as S. bovis biotype I) is an emerging cause of septicemia and endocarditis in the elderly. Intriguingly, epidemiological studies revealed a strong association, up to 65%, between endocarditis due to S. gallolyticus and colorectal malignancies. Whether S. gallolyticus infection is a cause or a consequence of colon cancer remains to be investigated. We previously showed that colonization of damaged cardiac valves with exposed collagen is mediated by the Pil1 pilus in S. gallolyticus. In the present work, we report that Pil1 is heterogeneously expressed at the single cell level, giving rise to two distinct bacterial subpopulations, a majority of weakly piliated cells and a minority of hyper-piliated cells. We have characterized, at the molecular level, a novel regulatory mechanism responsible for Pil1 heterogeneous expression combining phase variation in the leader peptide and transcriptional attenuation. Pili are highly immunogenic proteins proposed as vaccine candidate in pathogenic streptococci whose expression involves a fitness cost due to the selective pressure of host immune responses. Hence, this robust and simple system mitigates susceptibility to immune defenses while ensuring optimal colonization of host tissues.
| Streptococcus gallolyticus, formerly known as Streptococcus bovis biotype I, is present asymptomatically in the gastrointestinal tract of 2.5–15% of the human population [1]. However, this commensal bacterium can become a pathogen responsible for infective endocarditis in the elderly. Intriguingly, epidemiological studies pointed out a strong association, up to 65%, between endocarditis due to S. gallolyticus and colorectal malignancies [1]–[3]. Whether S. gallolyticus presence is a cause or a consequence of colon cancer development remains unknown [4]. Genome analysis of S. gallolyticus UCN34, a strain isolated from a patient suffering from infective endocarditis and colon cancer, revealed the existence of three pilus loci named pil1, pil2, and pil3 [5]. Pili are long filamentous structures extending from the bacterial surface, composed of covalently linked pilin subunits, which play key roles in adhesion and colonization of host tissues. Each pilus locus encodes two structural LPXTG proteins and one sortase C, an enzyme which covalently links pilin subunits during assembly of the pilus filament. The Pil1 locus of strain UCN34 is composed of three genes encoding a major pilin, PilB (Gallo2178), a collagen-binding adhesin, PilA (Gallo2179), and a sortase C (Gallo2177). In previous studies, PilA was shown to bind to collagen type I, the major component of cardiac valves, and to collagen type IV, enriched in basal lamina of pre-cancerous polyps [6], [7]. PilA constitutes the major collagen-binding protein in S. gallolyticus, conferring adhesive properties to the pilus, and is involved in the development of infective endocarditis in a rat experimental model [7].
Using immunogold electron microscopy, we previously noted that expression of Pil1 pilus in S. gallolyticus strain UCN34 was heterogeneous with less than half of the bacteria expressing detectable pili [7]. Similar observations have been reported in other piliated gram-positive bacteria, such as Corynebacterium renale and Corynebacterium pilosum, Streptococcus pneumoniae, Streptococcus pyogenes, and Enterococcus faecalis [8]–[16]. Regulation of pilus genes expression occurs primarily at the transcriptional level and most pilus regulatory genes are located immediately upstream of the divergently transcribed pilus loci [17]. The corresponding regulators belong either to the AtxA/Mga superfamily (e.g. EbpR in E. faecalis), including the RALP-family of RofA-like regulators (RlrA in S. pneumoniae, Nra in S. pyogenes and RogB in S. agalactiae) or are members of the AraC/XylS- family (MsmR in S. pyogenes and Ape1 in S. agalactiae).
A bistability mechanism was proposed recently to explain the heterogeneous expression of the PI-1 pilus in S. pneumoniae [12], [18]. In this bacterium, the transcription of PI-1 locus is tightly controlled by RlrA, which also activates its own expression. The bistable expression of the PI-1 genes is mediated by this positive-feedback loop [19]. It was therefore proposed that cells in which rlrA expression is autoactivated display high intracellular concentrations of RlrA with the concomitant high-level expression of the pilus genes. Return to the low expressing pilus state is mediated by the pilus structural component RrgA acting as a negative regulator of RlrA [12]. External environmental factors could also influence the heterogeneous expression of pili. For example, in S. pyogenes M49, the percentage of cells expressing the FCT-3 pilus increased with lower temperatures, where it is significantly higher at the temperature of external surfaces such as skin (47% at 30°C) compared to the body temperature (20% at 37°C) [14]. In E. faecalis strain OG1RF, expression of the pilus operon ebpABC increased in presence of bicarbonate [15] or serum [20], with more cells producing pili, and the extent of piliation in bacteria recovered from rat endocarditis vegetations was even higher. Although the mechanisms explaining the various regulatory features of pilus expression are largely unknown, these observations highlight the fact that bacteria respond to physiological conditions by altering their tissue adhesive capacity.
In this study, we identified a new regulatory mechanism that controls heterogeneous expression of Pil1 in S. gallolyticus UCN34. We showed that the heterogeneity of pilus Pil1 expression depends on a phase variation mechanism between short tandem repeats in a leader peptide-encoding gene located at the 5′ end of the pilus gene cluster transcript. Some of these rearrangements provoke sequence frameshift leading to synthesis of longer leader peptides that potentiate transcription of the downstream pil1 genes through ribosome-induced destabilization of a stem-loop transcription terminator. Pili are highly immunogenic protein polymers required for bacterial adhesion. Therefore, this stochastic mechanism of pilus expression constitutes a robust and simple system used by this organism to evade the host immune response and, when needed, to ensure optimal colonization of host tissues.
Examination of immunogold scanning electron micrographs at lower magnification revealed a heterogeneous expression of Pil1 pilus in the S. gallolyticus strain UCN34, with less than 50% of the cells in the population displaying pili on their surface when labeled with a specific antibody against the major pilin PilB. This result contrasts with the homogeneous labeling of cells when the pil1 operon is constitutively expressed in the heterologous host Lactococcus lactis NZ9000 (Fig. 1A). Consistently, two distinct subpopulations of cells with low (Pil1low) and high (Pil1high) Pil1 pilus levels were observed by immunofluorescence analyses with anti-PilB antibody whereas, again, no heterogeneity was seen in a lactococcal strain expressing pil1 under the control of a constitutive promoter (Fig. 1B). Quantification of these data by flow cytometry analyses showed that Pil1low cells account for 67% and Pil1high cells for 28% of the total population; the remaining 5% being Pil1 negative (Fig. 1C and 1D). Similar results were obtained using antibodies against the pilus-associated adhesin PilA for immunodetection and flow cytometry (data not shown).
The pil1 locus of S. gallolyticus strain UCN34 consists of three genes encoding the collagen-binding adhesin PilA (Gallo2179), the major pilin PilB (Gallo2178), and the sortase C enzyme (Gallo2177) responsible for the covalent polymerization of PilA and PilB (Fig. 2A) [5], [7]. Upstream and divergent from pil1 lies a gene, gallo2180, encoding a putative transcriptional regulator belonging to the TetR family (Fig. 2A). However, this regulatory gene is not always associated to the pil1 locus in S. gallolyticus isolates. It is present in the genomes of all sequenced S. gallolyticus strains, even in those that do not carry the pil1 locus, and in some closely related non-pathogenic species that lack this operon, such as Streptococcus macedonicus. Quantitative RT-PCR did not reveal any changes in the transcript levels of gallo2180 in our set of clinical strains displaying different levels of pil1 [7]. Finally, overexpression or mutational inactivation of gallo2180 in strain UCN34 did not alter the pattern of Pil1 expression but strongly altered bacterial cell morphology (data not shown). Taken together, these results demonstrate that gallo2180 gene product does not control pil1 transcription.
A closer examination into the intergenic region located between gallo2180 and pilA revealed the presence of 22 GCAGA repeats (110 nt) followed by a putative stem-loop structure (ΔG = −18.7 kcal/mol) (Fig. 2B, 2D and 2E). It is located entirely within a predicted open reading frame (ORF) specifying a putative short 38 amino acid peptide, with a recognizable ribosome binding side immediately upstream of an ATG. This remarkable GCAGA-containing sequence is conserved in the published genomes of S. gallolyticus UCN34, ATCC43143, and BAA-20690 and in all isolates of our collection expressing Pil1 pilus (data not shown). However, these strains differed in the number of GCAGA repeats and in the profile of Pil1 expression (Fig. S3).
Determination of the transcription start sites of the pil1 operon in S. gallolyticus strain UCN34 by primer extension analysis using a primer within pilA showed two bands (Fig. 2C), one positioned upstream the GCAGA repeats (position -390 bp from ATGpilA) and the other within the downstream stem-loop structure (position -130 bp from ATGpilA) (Fig. 2D). The first transcription start site is preceded by a canonical -35 (TTGTGT) and -10 Pribnow (AATAAT) boxes, suggesting that it represents the site of initiation of transcription, whereas the second initiation site does not (Fig. 2D). We therefore hypothesized that the second signal corresponded to the pause of reverse transcriptase during elongation at this secondary RNA structure or a processed mRNA. Consistently, fusions between various DNA fragments from this region and a lacZ reporter gene confirmed that only the canonical promoter displayed a detectable activity (Fig. S1). Overall these results indicate that pil1 transcription starts 400 nucleotides upstream from the first gene pilA.
To determine if the heterogeneous expression of Pil1 pilus is controlled by the GCAGA-containing sequence, the pil1 genes (pilA, pilB and srtC) were deleted in S. gallolyticus UCN34 (Δpil1) and reintroduced in trans using two different expression plasmids. In plasmid pTCVerm-Ptet-pil1, the pil1 operon genes were transcribed from the constitutive promoter Ptet fused immediately upstream the pilA gene, whereas in pTCVerm-Ppil1-pil1 it was transcribed from the UCN34 pil1 promoter region, i.e. the 518-pb intergenic region located between gallo2180 and pilA and containing the GCAGA repeats. Flow cytometry analyses with anti-PilB antibody showed that the Δpil1 mutant complemented with the plasmid pTCVerm-Ptet-pil1 displayed a strong and homogeneous signal (Fig. 3C), whereas the complementation with pTCVerm-Ppil1-pil1 restored heterogeneity of pil1 expression, with 76% of Pil1low cells and 16% of Pil1high cells (Fig. 3D). As expected, the Δpil1 mutant was negative for pil1 expression (Fig. 1C and Fig. 3B). These observations were confirmed by immunofluorescence using anti-PilB antibody (Fig. 3C–D). Thus, this GCAGA-containing pil1 promoter region controls the heterogeneous expression of pil1 genes in S. gallolyticus.
To decipher the molecular mechanism underlying the heterogeneous expression of Pil1 pilus, UCN34 Pil1+ variants, highly enriched in Pil1high cells, were separated from the wild type (WT) population by immunoscreening. Briefly, a colony-blot was performed on isolated colonies grown on TH plates with anti-PilB antibody. As shown in Fig. 4A, approximately 10% of the WT UCN34 colonies displayed a clear ‘Pil1+ phenotype’ with an approximately 10-fold increase of PilB level (referred from now as Pil1+var). Larger immunolabeling screening showed that 5 to 10% of clones with a Pil1+ phenotype were always obtained from UCN34 WT; similarly, 5 to 10% of the clones from a Pil1+ variant recovered the WT pil1 expression level (Fig. 4B).
One variant named Pil1+var23 was further characterized phenotypically by immunogold electron microscopy, flow cytometry, Western blotting, and transcriptional analyses (Fig. 4C–F). As expected, Pil1+var23 displayed a high and homogeneous Pil1 expression profile. Interestingly, quantitative RT-PCR analyses indicated a strong increase of pil1 operon transcription in Pil1+var23 compared to the WT UCN34 strain (Fig. 4F). However, as shown in Fig. 4F, this increase was only observed for transcripts located downstream from the putative stem-loop structure (Fig. 2E). Several other Pil1+ variants were analyzed and displayed similar phenotypic traits (data not shown). Of note, pilB transcripts are 3 to 4 fold more abundant than pilA specific transcripts (Fig. 4F). This might be due to differences in mRNA stability along the pilus operon and/or to additional regulatory mechanism controlling the level of pilB transcript.
Direct sequencing of the pil1 promoter region of UCN34 Pil1+ variants revealed modifications of the number of GCAGA repeats compared to the UCN34 WT strain, i.e. 21 or 23 repeats (respectively Pil1+var21 and Pil1+var23) instead of 22 repeats. Reversion from a Pil1+ to a WT phenotype was always accompanied by either a switch to the original 22 GCAGA repeats, or surprisingly acquisition of an in-frame number of repeats, i.e. 19 or 25. These WT-like populations displaying 19 or 25 repeats were able, in turn, to generate new Pil1+ variants displaying out-of-frame number (17, 18, 20, 21, 23, 24 or 26) of GCAGA repeats within the coding sequence of the putative peptide in the 5′ end of the transcript (Fig. 4G). From herein, we will refer to this peptide as leader peptide as defined by Molhoj and Degan [21]. In summary, WT-like expression of Pil1 pilus, with two distinct subpopulations of cells observed in flow cytometry (2/3 of Pil1low, 1/3 of Pil1high), was always correlated with alterations to a number of GCAGA repeats that was in-frame with the putative leader peptide (22+3n repeats). Otherwise, in the Pil1+ variants with an out-of-frame number or repeats (e.g. Pil1+var23), the population displayed a single homogeneous peak highly enriched in Pil1high cells, as revealed by flow cytometry (Fig. 4G). This regulation of Pil1 expression due to addition or deletion of GCAGA repeats is reminiscent of a phase variation phenomenon.
The putative peptide encoded in the 518-bp pil1 promoter region of Pil1+ variants with 21 or 23 in-frame GCAGA repeats terminates immediately upstream (4 bp) or within the hairpin structure, respectively. Strikingly, in UCN34 WT (22 repeats), this peptide ends at a stop codon located 49 bp upstream of the transcription terminator (Fig. 2D). We therefore hypothesized that translation of this putative leader peptide encoded by the region containing the GCAGA repeats was involved in pil1 transcription. To test this hypothesis, we carried out quantitative RT-PCR analyses of the pil1 operon genes with total RNA extracted from UCN34 WT strain and Pil1+var23 cultivated in the presence of chloramphenicol to uncouple transcription and translation. Addition of chloramphenicol to Pil1+ variants decreased the transcription of the pil1 operon to a level similar to that observed in the WT strain (Fig. 5). Addition of chloramphenicol to UCN34 WT strain had no significant effect on the transcription of the pil1 locus or of the control gene tanA (Fig. 5).
The above-described results led us to propose a model for the regulation of expression of pilus genes (Fig. 6A) where a regulatory leader peptide encoded in the 5′ end of the transcript and whose length varies upon phase variation (addition/deletion of GCAGA repeats), controls the switch of pilus transcription through translation-mediated anti-termination at the stem-loop structure acting as a premature transcription terminator upstream the pil1 operon.
Flow cytometry analysis revealed that UCN34 WT population consists of two distinct subpopulations with 67% of the cells displaying low level of Pil1 pilus (Pil1low) and 27% a high expression level of Pil1 (Pil1high) (Fig. 1B). According to our model, Pil1low cells should possess 22 GCAGA repeats leading to synthesis of a short leader peptide. Thus, most transcripts initiated at the promoter Ppil1 end at this hairpin structure, and the downstream pil1 operon is transcribed at a low level (Pil1low phenotype). In contrast, in Pil1high cells, addition or deletion of a GCAGA repeat (e.g. 21 or 23, repeats) leads to a frameshift associated with synthesis of a longer leader peptide whose stop codon is located within or close to the transcription terminator. As the transcript-bound ribosome covers around 30 nucleotides it prevents the stem-loop formation and the downstream pil1 operon is transcribed at a high level (Pil1high phenotype). Thus, translation of this regulatory leader peptide at the 5′ end of the mRNA up-regulates pil1 transcription by preventing the formation of the transcription terminator.
To further test the regulation model proposed in Fig. 6A, three mutations in the sequence encoding the putative leader peptide and the stem-loop structure were introduced in S. gallolyticus UCN34 chromosome: deletion of one strand of the stem-loop structure shown in Fig. 2E (Δterm); alterations of the RBS and ATG of the regulatory leader peptide (ATG* mutant where the GGAG of the RBS was replaced by CCTC and the ATG translational initiation codon by ACC); and addition of two stop codons at the end of the repeats (3 STOPs) to block translation in the three reading frames (Fig. S2). For each mutant generated in strain UCN34, we also selected a clone that reverted to the WT genotype (bWT) following homologous recombination. These bWT strains (standing for back to the WT) should display the WT phenotype and are isogenic to their mutant counterparts, i.e. they should possess the same secondary mutations, if any that may have occurred during their engineering. As shown in Fig. 6B, the Pil1 expression profile of each bWT strain was always superimposable to the WT UCN34. Flow cytometry analyses and immunolabelling were carried out to quantify the level of Pil1 using anti-PilB antibody (Fig. 6B, 6C and data not shown).
As predicted by our model, the S. gallolyticus UCN34Δterm mutant strain showed a high and homogeneous Pil1 expression by flow cytometry compared to the heterogeneous expression of the parental strain UCN34 (Fig. 6B). In addition, the immunoscreening of 384 individual clones clearly demonstrated that 100% of the cells in the Δterm population were trapped in the Pil1high configuration and highly expressed pil1 (Pil1+locked mutant); in contrast, 10% of the cells in the Pil1+ variants returned to a “WT pil1 expression” by a phase variation mechanism (Fig. 6C). Quantitative RT-PCR analyses showed a 11-fold increased pilB transcription in the Δterm mutant compared to the parental UCN34 strain (Fig. 6D). These results are consistent with our proposal that the stem-loop structure upstream of the pil1 genes acts as a premature transcription terminator. To further substantiate this finding, we have introduced the transcriptional terminator (TT) containing the run of 6T residues in the 3′ region downstream from the constitutive promoter Ptet in the beta-galactosidase reporter vector pTCV-lac (annotated Ptet-TT-lacZ ). As shown in Fig. S1, addition of the transcriptional terminator led to the reduction of lacZ reporter transcription (white/light blue colonies) compared to the control plasmid Ptet-lacZ . Interestingly, as predicted for E. coli intrinsic terminators, deletion of the 6T residues in the 3′ region (annotated Ptet-TTdelT-lacZ ) resulted in inactivation of the transcriptional terminator (strong blue colonies).
In the ATG* mutant, the ATG initiator codon of the regulatory leader peptide was replaced by ACC and its ribosome binding site drastically modified (GGAG to CCTC) (Fig. S2). Mutational inactivation of the leader peptide translation start signal led to a low and homogeneous pil1 expression as shown by flow cytometry (Fig. 6B) and immunolabeling screening (not shown). Since the leader peptide is not translated, there is no opening of the transcription terminator and the heterogeneous expression is lost. As expected, the ATG* mutant displayed a significantly reduced amount of pil1 transcription by quantitative RT-PCR compared to the parental strain UCN34. The remaining pil1 expression, referred to as Pil1low, is probably due to a leak of the transcription terminator.
Finally, in the 3 STOPs mutant (Fig. S2) where only a short leader peptide was translated irrespective of the number of repeats (22, 18 or 10 repeats), a single peak was observed by flow cytometry (Fig. 6B, Fig. S6). Unexpectedly, the Pil1 expression level in these mutants is slightly higher than in the UCN34 WT Pil1low cells subpopulation, a feature likely due to mutation-associated disturbance (e.g. increase mRNA stability).
Taken together, these results demonstrate that the translation of this regulatory leader peptide controls pil1 expression.
We previously showed that Pil1 was necessary and sufficient for binding to collagen [7]. Here, using isogenic mutants of S. gallolyticus UCN34, we showed a perfect correlation between Pil1 synthesis and the ability to bind to type I collagen (Fig. S5). These results confirmed that Pil1 is the major collagen-binding determinant of S. gallolyticus.
We previously showed that specific antibodies directed against S. gallolyticus pili, including Pil1, were present in the Dutch and American populations [22]. In addition, when raising polyclonal antibodies in rabbit and mice against the WT bacteria UCN34, a substantial proportion of the antibodies were directed against Pil1 pilins (data not shown). Therefore, although beneficial for colonization of host tissues, a high expression of Pil1 could be detrimental for evading the host immune response. To test experimentally the possible role of heterogeneous pilus expression in immune escape, we first assessed the survival of S. gallolyticus UCN34 WT, Δpil1 and Δterm (Pil1+locked) strains in human whole blood from three different donors. As shown in Fig. 7, the Pil1+locked mutant was killed more efficiently than the Δpil1 mutant, most probably by neutrophils.
We also infected the human monocyte-macrophage cell line THP-1 with UCN34 WT, Δpil1 and Δterm mutants in the presence or absence of purified antibodies against Pil1. S. gallolyticus UCN34, Δpil1 and Δterm (Pil1+locked) were poorly phagocytosed in the cell medium alone (≈2–5%). Addition of purified anti-Pil1 antibodies led to a 10-fold increase in the uptake of Pil1+locked mutant via opsonophagocytosis but did not modify that of Δpil1 (Fig. 8A–B). Furthermore, flow-cytometry analysis of the intracellular UCN34 bacteria revealed that the bacteria phagocytosed by THP-1 macrophages in the presence of opsonizing antibodies corresponded mainly to the Pil1high cells subpopulation (Fig. 8C).
Several recent publications have identified pili as playing a key role in the virulence of many gram-positive bacteria [17], [23]. Pili have been implicated in many facets of the infectious process such as adhesion/colonization of host tissues, translocation of epithelial barriers, and modulation of the innate immune responses. Their surface localization and high immunogenicity makes them attractive targets for the development of vaccines against important gram-positive pathogens such as Group B Streptococcus [24]. However, although the expression analysis of these important antigens is of obvious interest, very little is known about the regulation of pilus expression and how it is modulated by different environmental conditions.
Here, using single cell level analysis, we have shown that Pil1 pilus expression is heterogeneous in S. gallolyticus strain UCN34, with the co-existence of two subpopulations, a majority of weakly piliated (Pil1low) cells and a minority of hyper piliated (Pil1high) cells. We further demonstrated that this phenotypic variability depends on changes in the pil1 promoter region encoding a leader peptide made of a variable number of repeats and a downstream transcription terminator. These two cis-acting elements control transcription of the downstream pil1 genes. Sequence analysis of variants expressing high level of Pil1 pilus (referred to as Pil1+var), revealed out-of-frame addition/deletion of repeats compared to the parental UCN34 sequence; in contrast, clones displaying biphasic phenotype never exhibited sequence variation or in-frame addition/deletion of repeats. These results strongly suggested a phase variation mechanism involving simple sequence repeats (SSR) within the leader peptide-encoding gene. Accordingly, we postulate that translation of this leader peptide, whose length varies upon phase variation (addition/deletion of GCAGA repeats), controls the switch of pilus transcription through translation-mediated anti-termination at the stem-loop structure acting as a premature transcription terminator upstream the pil1 genes (Fig. 6A).
The phenotypically distinct cell subpopulations of UCN34, Pil1low and Pil1high, are present in variable ratios in other six S. gallolyticus isolates from our laboratory collection (Fig. S3). Sequencing of pil1 promoter region of these strains confirmed the presence of GCAGA tandem repeats, but in different number. Similarly, the transcription terminator identified in UCN34 was found in all but one isolate (strain 2472) where another stem-loop structure is present at the same position. Isolates in which the regulatory leader peptide stops at least 28 nucleotides before the terminator (strains 2471, 2477, and 2479) displayed the same heterogeneous Pil1 expression profile as UCN34 (2/3 Pil1low cells, 1/3 Pil1high cells). In contrast, isolates in which the leader peptide terminates immediately upstream (strain 2475) or extends through the stem-loop structure (strains 2470 and 2472) displayed a high proportion of Pil1high cells. Thus, the Pil1 expression profiles of these isolates could be predicted by our regulation model based on the sequence analysis of their pil1 promoter regions. It is noteworthy that we could not find any external factors (temperature, pH, oxygen concentration, glucose, bicarbonate, serum, plasma, intestinal juice) modulating the two-third/one-third Pil1low/Pil1high ratio in strain UCN34 (unpublished data). This result is consistent with a phase variation mechanism usually considered as a stochastic or random event and not as a responsive process.
We expressed pil1 in Lactococcus lactis on a plasmid under the control of the promoter Ppil1 and we also observed heterogeneity in Pil1 expression, but the two populations were not as clearly separated as for S. gallolyticus (Fig. S7). In addition, we found a 23-repeat variant expressing pil1 homogeneously at a high level. So it appears that the regulatory mechanism of pil1 remains functional in this closely related species.
Sequence analysis revealed the presence of GCAGA repeats in the promoter region of another pilus locus (pil3) in the strain UCN34. Pil3 pilus expression is also heterogeneous, with two distinct subpopulations of cells observed by flow cytometry, Pil3high and Pil3low (Fig. S4). Strikingly, its promoter region also includes an ORF with 13 GCAGA tandem repeats encoding a putative regulatory leader peptide and a stem-loop transcription terminator (ΔG = −28.20 kcal/mol) located 33 bp downstream. Highly expressing Pil3+ variants all possessed out-of-frame addition of GCAGA repeats and encoded a longer regulatory leader peptide whose stop codon was located within or downstream of the hairpin structure. Similarly to Pil1, Pil3 expression is regulated by phase variation in S. gallolyticus.
Transcription attenuation is a commonly used regulation strategy characterized by the presence of an attenuator sequence involved in formation of mRNA stem-loops that prevent transcription of downstream genes. The best characterized transcription attenuation system is that of the Escherichia coli tryptophan (trp) operon [25]. At low intracellular tryptophan concentrations, the ribosome stalls during the translation of a tryptophan-containing leader peptide at a location that impedes the formation of a transcription terminator, thus enabling transcription of the trp operon. Attenuation of the trp operon is made possible by the fact that the rate of translation influences RNA structure, which in turn influences continuation of transcription. Translation therefore interferes with transcription, making this an example of translation-mediated transcription attenuation [26]. In the case of pil1 operon, it is the length of the leader peptide varying upon addition/deletion of repeats that controls the transcription attenuation process.
Commonly described mechanisms of phase variation include gene inversion, gene conversion, epigenetic modifications and slipped-strand mispairing [27], [28]. Short sequence repeat (SSR) tracts, that undergo slipped-strand mispairing, are a major mechanism of stochastic switching of genes expression in bacterial commensals and pathogens such as Haemophilus influenzae, Escherichia coli, Salmonella enterica, Campylobacter jejuni, Neisseria gonorrhoeae and Helicobater pylori [29]–[36]. In a SSR tract, a short unit of several nucleotides is repeated multiple times in the DNA sequence. During DNA replication, the process of recA-independent slipped-strand mispairing (or recA-dependent unequal crossing-over) can change the number of repeated units at a particular locus; this, in turn, can cause a shift in the reading frame, resulting in an out-of-frame translation, or alter critical spacing in the promoter region, which can, in both cases, “switch off” genes expression [27]. Mutations of this type are frequent and reversible, leading to rapid, stochastic on-off switching of expression of the gene and associated phenotype [27], [37]. Such populations may thus be “primed” for rapid adaptation [38]. SSRs have a high prevalence among surface associated proteins, such as pili and adhesins, probably due to their direct interactions with host structures and as targets of the immune responses [37]. In S. gallolyticus strain UCN34, the phase variation rate of the pil1 locus, as defined by Eisenstein, is estimated to be 5.10−3/cell/generation [39], which is compatible with the rates (10−3 to 10−4/cell/generation) commonly reported, such as for opa genes in Neisseria gonorrheae [40]. This high mutation frequency is a characteristic of slipped-strand mispairing mechanism [27], [37].
Expression of highly immunogenic pili proteins involves a fitness cost due to the selective pressure of host immune responses. Hence, heterogeneous pili expression allows a subset of bacteria to be phenotypically pre-adapted to take advantage of particular environments and/or to avoid adverse conditions. The highly piliated subpopulation could allow colonization of specific host tissues exposing collagen, while the weakly expressing subpopulation is more prone to dissemination because it is less adherent to host tissues and may escape the host immune response. The hypothesis of a facilitated immune escape through phase variation has been demonstrated for major surface exposed structures such as LPS and PorA in Neisseria meningitidis [41], [42]. Consistently, we showed here that the Pil1+locked mutant (Δterm) survived less in human whole blood than the Δpil1 mutant. This result indicates that neutrophils constituting the first line of innate host defenses better recognized bacteria that highly express Pil1, and killed them efficiently. We also showed that the non-piliated Δpil1 mutant better escape to opsonophagocytosis by THP-1 human macrophages than the Pil1+locked mutant (Δterm). Furthermore, we have shown that among the heterogeneous Pil1 expressing WT bacteria, the bacteria that are phagocytosed by THP-1 are those expressing higher Pil1 levels i.e. the Pil1high cells subpopulation.
This study describes the first example of pilus regulation through phase variation in a gram-positive pathogen. At the molecular level, it occurs through an original mechanism that combines phase variation in a leader peptide-encoding gene and transcription attenuation. Phase variation stands in contrast to other classical regulatory mechanisms, involving stand-alone or two-component regulators that tend to drive the entire bacterial population into an alternative expression status. This relatively simple yet robust regulation mechanism ensures a stochastic heterogeneous pilus expression at the bacterial surface, a feature important for evading the host immune system and to ensure optimal colonization versus dissemination in host tissues.
Blood collection from human healthy volunteers was supplied by the ICAReB Plateform at the Institut Pasteur (Paris, France) in accordance to the guidelines of the agreement between Institut Pasteur and the Etablissement Français du Sang.
The human monocytic cell line THP-1 was cultured in RPMI-1640 GlutaMAX™ medium (Gibco reference 61870-010) supplemented with 10% FCS in a 10% CO2 atmosphere at 37°C. Phorbol 12-myristate 13-acetate (PMA) was used at 50 ng/mL to induce THP-1 monocytes to differentiate into macrophages.
Bacterial strains, plasmids and oligonucleotide primers are listed in Table S1. S. gallolyticus strains were grown at 37°C in Todd-Hewitt (TH) broth in standing filled flasks. L. lactis strain NZ9000 [43] was grown in M17 medium supplemented with 1% glucose (M17G). Heterologous expression of pil1 in L. lactis strain was described previously [7]. Tetracycline and erythromycin were used at 10 µg/mL.
For scanning electron microscopy (SEM), bacteria were collected after overnight growth, fixed, and stained with rabbit anti-PilB IgG followed by anti-rabbit secondary antibody conjugated to 10 nm colloidal gold as previously described [44].
S. gallolyticus or L. lactis recombinant strains were grown overnight in 10 ml of TH (OD600≈2) or M17G supplemented with erythromycin (OD600≈4) respectively. Bacteria were washed twice in phosphate buffered saline (PBS) before fixation in PBS containing 3% paraformaldehyde for 15 min at RT. Fixed bacteria were washed twice with PBS, blocked with PBS-BSA 3% for 30 min, and incubated for 45 min with rabbit or mouse primary antibodies diluted in PBS-BSA 0.5% at the following dilutions: anti-PilB (1/800) and anti-PilA (1/500). After three washings with PBS, samples were incubated for 30 min with secondary DyLight 488-conjugated goat anti-rabbit or mice immunoglobulin diluted in PBS-BSA 0.5% (1/300 dilution; Thermo Scientific Pierce) and Hoecht 33342 (1/1,000). Coverslips were mounted with 4 µl of Fluoromount-G mounting medium (SouthernBiotech). Microscopic observations were done with a Nikon Eclipse Ni-U and images acquired with a Nikon Digital Camera DS-U3.
To analyze Pil1 pilus expression by flow cytometry, 500 µl of overnight cultures were collected and washed twice in PBS. Bacterial pellet was resuspended in PBS-BSA 3% blocking solution for 30 min, and then incubated with rabbit anti-PilB (1/800) or anti-PilA (1/500) antisera diluted in PBS-BSA 0.5% for 45 min on ice. After three washes with PBS, samples were incubated with secondary DyLight 488-conjugated goat anti-rabbit or anti-mouse immunoglobulin (Thermo Scientific Pierce) diluted in PBS-BSA 0.5% for 30 min on ice. Cells were washed before fixation in PBS containing 1% paraformaldehyde for 20 min. Samples were acquired on a MACSQuant Analyzer apparatus (Miltenyi Biotec) and data were analyzed using FlowJo software.
Total RNA was used as template for primer extension reaction using a radiolabeled specific primer complementary to a sequence located downstream the putative pil1 promoter region, CD37 (Table S1), as previously described [45]. The corresponding Sanger DNA sequencing reactions (GATC) were carried out by using the same primer and a PCR-amplified fragment containing the pil1 upstream region (primer pair CD35/CD37) with the Sequenase PCR product sequencing kit (USB).
Transcriptional fusions of different promoter regions with a spoVG-lacZ reporter gene have been performed using pTCVlac using EcoRI/BamHI restriction sites, and the various plasmids were introduced into Streptococcus agalactiae NEM316. Overnight cultures were diluted 1 in 100 in 10 ml of fresh TH broth with erythromycin and grown for 4–5 h at 37°C. The β-galactosidase activities were measured using the Beta-Glo assay system (Promega, WI) according to the manufacturer's recommendations on exponentially growing bacteria (OD600nm = 0.7). The assay consisted of mixing an equal volume (50 µl) of a bacterial suspension in PBS (adjusted to OD600nm of 1) with the Beta-Glo Reagent containing a luciferin-galactoside substrate (6-O-β-galactopyranosyl-luciferin). After 1 h incubation in the dark, the light produced was measured in a luminometer Lumat LB 9507 tube (Berthold Technologies). The results are expressed as relative light unit (RLU) per OD600nm, and are representative of at least three independent experiments.
Bacteria were grown in TH medium at 37°C and harvested for protein analysis during late exponential phase of culture. Cell wall extracts were prepared as previously described (24). For analysis of PilB expression by Western immunoblotting, cell wall proteins were boiled in Laemmli sample buffer, separated by SDS-PAGE on 4–12% Tris-Acetate Criterion XT gradient gels and transferred to nitrocellulose membrane (Hybond-C, Amersham). PilB was detected using specific polyclonal antibodies and horseradish peroxidase (HRP)-coupled anti-rabbit secondary antibodies (Zymed) and the Western pico chemiluminescence kit (Pierce). Image capture and analysis were done on GeneGnome imaging system (Syngene).
For colony blots, 100 µl of a 10-6 dilution of overnight culture was spread on TH agar plates, incubated for 24 h at 37°C and transferred onto nitrocellulose (Hybond-C, Amersham). The membrane was dried in the microwave (3 min at 300 W) and soaked in blocking solution of PBS with 3% skimmed milk for 30 min. Immuno-detection was performed as described above.
For immunolabeling screening, 384 isolated colonies of each strain grown on agar plates were inoculated in 100 µl of TH medium in four 96-well plates. After 8 h culture at 37°C, the 96-well plates were gently shaken for 15 s and each plate was replicated on TH agar plates with a robot Rotor HDA (Singer Instruments). Following overnight incubation at 37°C, the four plates were combined on a single TH agar plate containing 384 clones. This master plate was duplicated and transferred onto nitrocellulose membrane (Hybond-C, Amersham) for immuno-detection.
For sequencing of short regions, genomic DNA was prepared with the DNeasy Blood & Tissue kit (Qiagen), and sequencing reactions were performed on 10 µl of DNA matrix with the BigDye Terminator v3.1 Cycle Sequencing kit (Applied Biosystems). Genomic DNA of variants (Pil1+var23; 23 GCAGA repeats) and mutant strains (Δterm, ATG*) were extracted using MasterPure Gram Positive DNA Purification Kit (Epicentre, Illumina) and genomes were sequenced by Next-Generation Sequencing (NGS) technique to confirm the absence of disturbing secondary mutations. The samples were sent to the Institut Pasteur platform PF1 for sequencing.
Total RNA (15 µg) were extracted and treated as described previously [46]. Quantitative RT-PCR analyses were performed as previously described [46] with gene-specific primers (Table S1). For treatment with chloramphenicol, the antibiotic was added to 20 ml of culture in exponential phase (OD600nm = 0.3) to 8 µg/mL, a concentration that inhibits bacterial growth. The mock control consists of the same volume of culture with the diluent (ethanol). After 30 min at 37°C, the cultures were harvested for RNA extraction.
Experimental details concerning the construction of the S. gallolyticus UCN34 Δpil1 mutant deleted for the three genes constituting the pil1 operon (gallo2179-2178-2177) is the subject of a manuscript describing a method to construct isogenic mutants in this non-naturally transformable species (Danne et al., in preparation). The same technique has been used to construct Δterm, ATG* and 3 STOPs mutants (for primers, see Table S1).
Human monocytic THP-1 cells were seeded into 24-well plates at 5. 105 cells per well in RPMI-1640 GlutaMAX supplemented with 10% FCS and PMA (50 ng/mL) and incubated for 48 h prior to phagocytosis. This assay was performed on exponentially grown bacteria (OD≈0.4) washed once in PBS and diluted in RPMI-1640 GlutaMAX medium. When indicated, purified antibodies against the major pilin (Gallo2178 or PilB) were added to the bacteria (dilution 1/200) for 15 min at room temperature before addition to the host cells. Macrophages were infected at a multiplicity of infection (m.o.i.) of 10 bacteria per cell and incubated for 1 h to allow phagocytosis. Extracellular bacteria were killed with 50 µg/mL gentamicin for 1 h at 37°C. Viable intracellular bacteria (cfu) were determined by macrophage lysis in 1 ml of cold water. The percentage of phagocytosis was calculated as follows (cfu on plate count/cfu in original inoculum ×100). Assays were performed in triplicate and were repeated at least three times. Similar experiments were performed on coverslips and used for indirect immunofluorescence microscopy. After phagocytosis, the infected THP-1 cells were fixed in PBS containing 4% paraformaldehyde for 15 min at room temperature. Non-specific binding sites were blocked with PBS BSA 3% for 30 min at room temperature. Extracellular bacteria were labeled using a rabbit polyclonal antibody directed against S. gallolyticus UCN34 and revealed with AlexaFluor 594-conjugated goat anti-rabbit immunoglobulin (1/200 dilution in PBS-BSA 1%, Molecular Probes, Invitrogen). Next, the macrophages were permeabilized with PBS-Triton X-100 (0.2%) for 5 min at room temperature. Bacteria were labeled using a rabbit polyclonal antibody directed against S. gallolyticus UCN34 and revealed with AlexaFluor 488- conjugated goat anti-rabbit immunoglobulin (1/200 dilution in PBS-BSA 1%, Molecular Probes, Invitrogen). The THP-1 macrophages were labeled with Cy3-conjugated phalloidin (1/300 dilution) to visualize the actin network and with Hoeschst 33342 (10 µg/mL) to detect the nuclei. Using this differential staining technique, intracellular bacteria will appear only labeled in green whereas extracellular bacteria will appear both in green and red ( = yellow). Images are acquired with Apotome Zeiss and Axiovision 4.6 software. This system provides an optical slice view reconstructed from fluorescent samples, using a series of “grid projection” (or “structured illumination”) acquisitions.
|
10.1371/journal.ppat.1002363 | Inhibition of IL-10 Production by Maternal Antibodies against Group B Streptococcus GAPDH Confers Immunity to Offspring by Favoring Neutrophil Recruitment | Group B Streptococcus (GBS) is the leading cause of neonatal pneumonia, septicemia, and meningitis. We have previously shown that in adult mice GBS glycolytic enzyme glyceraldehyde-3-phosphate dehydrogenase (GAPDH) is an extracellular virulence factor that induces production of the immunosuppressive cytokine interleukin-10 (IL-10) by the host early upon bacterial infection. Here, we investigate whether immunity to neonatal GBS infection could be achieved through maternal vaccination against bacterial GAPDH. Female BALB/c mice were immunized with rGAPDH and the progeny was infected with a lethal inoculum of GBS strains. Neonatal mice born from mothers immunized with rGAPDH were protected against infection with GBS strains, including the ST-17 highly virulent clone. A similar protective effect was observed in newborns passively immunized with anti-rGAPDH IgG antibodies, or F(ab')2 fragments, indicating that protection achieved with rGAPDH vaccination is independent of opsonophagocytic killing of bacteria. Protection against lethal GBS infection through rGAPDH maternal vaccination was due to neutralization of IL-10 production soon after infection. Consequently, IL-10 deficient (IL-10−/−) mice pups were as resistant to GBS infection as pups born from vaccinated mothers. We observed that protection was correlated with increased neutrophil trafficking to infected organs. Thus, anti-rGAPDH or anti-IL-10R treatment of mice pups before GBS infection resulted in increased neutrophil numbers and lower bacterial load in infected organs, as compared to newborn mice treated with the respective control antibodies. We showed that mothers immunized with rGAPDH produce neutralizing antibodies that are sufficient to decrease IL-10 production and induce neutrophil recruitment into infected tissues in newborn mice. These results uncover a novel mechanism for GBS virulence in a neonatal host that could be neutralized by vaccination or immunotherapy. As GBS GAPDH is a structurally conserved enzyme that is metabolically essential for bacterial growth in media containing glucose as the sole carbon source (i.e., the blood), this protein constitutes a powerful candidate for the development of a human vaccine against this pathogen.
| Streptococcus agalactiae (Group B streptococcus, GBS) is the leading infectious cause of morbidity and mortality among neonates. However, there is still no satisfactory explanation of why neonates are so susceptible to GBS infections. Intrapartum antibiotic prophylaxis (IAP) was implemented in many countries but led to the emergence of antibiotic-resistant GBS strains. Therefore, maternal vaccination represents an attractive alternative to IAP. Here, we show that the high susceptibility of newborn mice to GBS infections is associated with their propensity to produce elevated amounts of immunosuppressive cytokine IL-10. We also demonstrate that IL-10 impairs neutrophil recruitment into infected organs thus preventing bacterial clearance. We identified extracellular GAPDH as the GBS factor that induces the high IL-10 production detected early upon neonatal infection. We show that maternal vaccination with recombinant GAPDH confers robust protective immunity against lethal infection with a GBS hyper-virulent strain in mice offspring. This protection can also be obtained either by antibody neutralization of GBS GAPDH or by blocking IL-10 binding to its receptor. As GBS GAPDH is an essential protein for bacterial growth, it is present in all GBS strains and thus constitutes an appropriate target antigen for a global effective vaccine against this pathogen.
| Streptococcus agalactiae, also named Group B Streptococcus (GBS), is a Gram-positive encapsulated commensal bacterium of the human intestine that colonizes the vagina of up to 30% of healthy women. This bacterium is the leading cause of neonatal pneumonia, septicemia, and meningitis [1], [2], [3], [4]. Neonatal GBS infections are acquired through maternal transmission and may result in early-onset disease (EOD), which occurs within the first week of life, or in late-onset disease (LOD), that occurs after the first week and accounts for most meningitis cases and deaths [3], [5], [6]. Despite early antimicrobial treatment and improvement in neonatal intensive care, up to 10% of neonatal invasive GBS infections are lethal and 25 to 35% of surviving infants with meningitis experience permanent neurological sequelae [3]. Because recommendations for intrapartum antibiotic prophylaxis (IAP) for mothers in labor at risk for GBS infection have been widely implemented in many countries, the incidence of EOD has declined to <1/1,000 births, but the incidence of LOD has slowly increased in the last decade [7]. An unexpected burden of case fatalities among children aged less than 90 days caused by GBS infection was recently reported in different European countries [8], [9], [10]. Moreover, recent reports described the emergence of antibiotic-resistant GBS strains likely caused by the widespread use of IAP [11], [12].
Maternal vaccination is the best alternative to IAP to deal with GBS neonatal infections. Vaccines to prevent GBS disease have been initially developed by coupling capsular polysaccharide (CPS) antigens to immunogenic protein carriers. Glycoconjugate vaccines against nine GBS serotypes have been shown to be immunogenic in animals, but the existence of distinct epitope-specific capsular serotypes has hampered the development of a global GBS vaccine [5], [13]. Moreover, glycoconjugated vaccines directed against the ten known serotypes of GBS would not protect against infections by nontypeable GBS isolates that are increasingly being reported [14], [15], [16], [17].
The sequencing of numerous GBS genomes has accelerated advances in vaccine development and new protein antigens have been revealed using reverse vaccinology [5], [18], [19]. To avoid the selection of mutants that escape immune recognition, the ideal human GBS vaccine should be directed against structurally conserved antigens that are essential for GBS virulence and/or growth, but none of the hitherto described candidate antigens fulfills these requisites.
The causes for the neonatal susceptibility to GBS infections are still poorly understood. Newborn immune system is not completely developed at birth, and undergoes an age-dependent maturation until fully developed. Thus, invasive infections in the first days of life pose serious threats for the newborn due to accentuated deficiencies in both innate and adaptive arms of the immune responses. Cases of early-onset GBS sepsis are usually characterized by an unexpectedly low number of neutrophils in infected tissues [20], [21], [22], [23]. This is commonly explained by the reduced neutrophil chemotaxis and impaired granulocyte maturation observed in neonates [24], [25], [26]. Of interest, high concentration of plasma and cord blood IL-10 in preterm neonates evaluated for sepsis was associated with mortality and is considered as an early indicator of prognosis [27], [28].
We have previously shown that the essential housekeeping enzyme glyceraldehyde-3-phosphate dehydrogenase (GAPDH) also acts as a GBS extracellular virulence factor that induces rapid production of interleukin-10 (IL-10) by the host [29]. Adult C57BL/6 mice (resistant to GBS infection) infected with a GBS mutant strain that over-express GAPDH (oeGAPDH) had increased bacterial colonization compared to mice infected with wild-type (WT) GBS. Increased bacterial burden in oeGAPDH infected C57BL/6 mice was accompanied by elevated serum levels of IL-10. Consequently, acquired susceptibility of C57BL/6 mice to oeGAPDH infection was completely reverted in IL-10-deficient animals [29]. This suggested that an exacerbated production of IL-10 during GBS infection might facilitate pathogen immune evasion. We demonstrate here that maternal immunization with rGAPDH confers protection against GBS infection in neonatal mice by abrogating the early IL-10 production detected upon the bacterial challenge. We also demonstrate that blocking GAPDH-induced early IL-10 production restores the recruitment of neutrophils in infected organs, which is essential for pathogen elimination and host protection against GBS infection.
Since GBS GAPDH is a structurally conserved enzyme that is metabolically essential for bacterial growth in blood, it constitutes an attractive target for the development of a human vaccine.
GAPDH, a key enzyme of the glycolytic pathway, is structurally conserved in all 8 published GBS genomes (identity>99.8%). Anti-rGAPDH IgG antibodies purified from sera of rGAPDH immunized mice or rabbits were thus used to demonstrate the presence of GAPDH in culture supernatants of ten unrelated GBS clinical isolates (Figure 1A) belonging to different serotypes and/or MLSTypes (Table S1). GBS GAPDH displays 44.7, 45.8 and 44.0% amino acid identity with rabbit, mice, and human GAPDH, respectively (Figure 1B). However, western blot and ELISA analysis revealed that rabbits and mice antibodies directed against GBS rGAPDH do not react with human, mouse, or rabbit GAPDH (Figure 2A and 2B). To favor the production of antibodies recognizing linear buried epitopes, mice were immunized with heat-denaturated rGAPDH (ΔT_rGAPDH). Anti-ΔT_rGAPDH antibodies purified from the sera of these animals did not show any cross-reactivity against mouse (self cross-reactivity) or human GAPDH when analyzed by western blot and ELISA (Figure 2C and 2E). These results are consistent with the fact that the longest identical stretches observed between eukaryotic and prokaryotic GADPH sequences are only 10-aminoacid long (Figure 1B).
To test whether maternal immunization with rGAPDH conferred protection to the offspring against GBS infection, female BALB/c mice were immunized with rGAPDH in alum adjuvant. Control mice were sham-immunized with the adjuvant alone.
Pups born from sham-immunized or rGAPDH-immunized females were infected intraperitoneally (i.p.) 48 h after birth with 5×106 colony-forming units (CFU) of serotype III virulent GBS strain NEM316. All but one mouse born from rGAPDH-immunized mothers survived the infection (95% survival) whereas 22 out of 27 infected pups succumbed to GBS challenge in the control group (18.5% survival) (Figure 3A). Most of the cases of GBS meningitis and LOD are caused by a serotype III hyper virulent clone, defined by multilocus sequence typing as ST-17 [30], [31], [32]. To better assess the effectiveness of maternal vaccination with rGAPDH, pups born from sham- or rGAPDH-immunized progenitors were i.p. infected 48 h after birth with 106 CFU of BM110, a serotype III GBS hyper virulent strain ST-17 (Table S1). All ST-17 GBS challenged neonates born from sham-immunized mothers died whereas mortality rate dropped to 21.4% in neonates born from rGAPDH-immunized mothers (Figure 3B). The protective effect conferred by rGAPDH maternal immunization was also observed in neonate mice infected by the subcutaneous (s.c.) route with 2.5×104 BM110 CFU. As shown in Figure 3C, none of the mice born from sham-immunized mothers survived this infectious challenge whereas only 23% of mice born from rGAPDH-immunized mothers died. Altogether, these results show that maternal vaccination with rGAPDH protected the offspring against GBS infections, including those caused by the hyper virulent strain BM110.
Pups born from rGAPDH-immunized mothers presented increased serum titers of anti-rGAPDH IgG antibodies when compared with those born from sham-immunized mothers (Figure S1). To evaluate the importance of these maternal antibodies in the newborn protection against GBS infection, neonatal mice were passively immunized with purified anti-rGAPDH IgG antibodies 12 h prior to GBS challenge. The passive antibody transfer conferred protection against infection caused by the virulent NEM316 or hyper virulent BM110 strains (Figure 4A and 4B). Anti-rGAPDH IgG antibodies conferred a similar protection to neonate mice infected by the s.c. route (data not shown). As described by others [33], we observed that GAPDH is present at the cell surface of GBS strains (Figure S2) and, it was therefore conceivable that protection conferred by anti-rGAPDH antibodies could be due to an enhanced opsonophagocytosis-mediated killing of GBS. However, anti-rGAPDH IgG antibodies did not enhanced in vitro phagocytosis or complement-mediated killing of GBS BM110 cells (Figure 4C). This indicated that protection conferred by anti-rGAPDH antibodies was not mediated by these mechanisms. Furthermore, complete protection against GBS infection was observed in neonate mice treated with purified anti-rGAPDH F(ab')2 fragments 12 h before i.p. infection with BM110 strain. In contrast, all pups that received the same amount of control F(ab')2 fragments died within the first 3 days upon the infectious challenge (Figure 4D). Altogether, these results demonstrate that enhanced opsonophagocytic killing or complement activation did not mediate the observed protective effect of anti-rGAPDH antibodies.
We have previously described a rise in IL-10 serum levels in adult mice treated with rGAPDH [29]. As shown in Figure S3, a similar increase in serum IL-10 levels was detected in newborn mice 1 h after i.p. injection of rGAPDH. Inactivation of rGAPDH enzymatic activity did not reduce this effect (Figure S3). This result indicates that IL-10 production induced by GBS GAPDH is independent of the dehydrogenase activity. We have also described that adult mice infected with GBS oeGAPDH mutant strain presented higher serum IL-10 levels than counterparts infected with WT GBS [29]. Thus, we also quantified the levels of serum IL-10 in mice pups at early times after GBS infection. As shown in Figure 5, infection of newborn mice with GBS WT strain NEM316 resulted in a rapid increase of serum IL-10 concentration. Maternal rGAPDH vaccination or treatment with anti-rGAPDH F(ab')2 fragments completely abrogated the elevated amount of IL-10 found in the sera of infected pups born from sham-immunized mothers or treated with control F(ab')2 (Figure 5A and 5B). Altogether, these results strongly suggest that the elevated IL-10 serum levels detected upon infection were due to GBS GAPDH.
The results presented above indicate that newborn susceptibility to GBS infection is most probably associated with early IL-10 production induced by GAPDH. To confirm this hypothesis, IL-10 deficient (IL-10−/−) pups and WT controls were infected with this bacterium. In agreement with our hypothesis, IL-10−/− pups were more resistant (78%) to GBS infection compared to WT controls (10%) (Figure 6A). To demonstrate further the essential role of IL-10 in neonatal susceptibility to GBS, newborn mice were treated with anti-IL-10 receptor (IL-10R) mAb 12 h before NEM316 or BM110 GBS challenge. As expected, most pups treated with anti-IL-10R mAb survived (86% or 82%, respectively) while all control pups died (Figure 6B and 6C). No additional protection was observed when newborn mice were treated simultaneously with anti-IL10R mAb and anti-rGAPDH IgG (Figure 7). Altogether, these results indicate that protection achieved using anti-GAPDH antibodies is due to inhibition of host IL-10 production. Several studies have shown that GBS can survive for prolonged periods within the phagolysosome of macrophages [34], [35], [36], [37]. Interestingly, we observed that the simultaneous addition of anti-rGAPDH IgG's, or anti-IL10R mAb, and GBS cells to bone marrow-derived macrophages (BMMφ) cultures inhibits the bacterial survival (Table 1). This result, combined with those shown in Figure S3, indicates that the inability of the macrophages to kill the intracellular GBS is due to IL-10 production induced by GBS GAPDH.
A plausible explanation for the observed protection induced by maternal vaccination with rGAPDH could be that GAPDH-mediated early IL-10 production, elicited upon the GBS challenge in newborns, inhibits the initiation of a host-protective inflammatory response. Neutrophil recruitment is an early event associated with protection against bacterial infection in neonates. Moreover, lack of neutrophil recruitment into infected organs has already been associated with neonatal susceptibility against GBS infections [21], [22], [23], [38]. To confirm that neutrophil recruitment into infected organs is an essential event for newborn protection against GBS infection, neutrophils of newborn mice were depleted by treatment with anti-Ly6G (1A8 clone) monoclonal antibodies (Figure S4). We observed that either blocking IL-10 signaling with neutralizing antibodies or anti-rGAPDH antibody treatment was not sufficient to protect neutropenic pups infected with a lethal inoculum of GBS (Figure 8).
In addition, we assessed the numbers and frequency of neutrophils in the liver and lungs of GBS NEM316-infected pups previously treated with anti-rGAPDH IgG or anti-IL10R mAb. The frequency and the total number of neutrophils quantified 18 h after GBS challenge in the analyzed organs of infected pups treated with control IgG (or isotype control) was as low as with the non-infected pups (data not shown). Treatment with either anti-rGAPDH IgG or anti-IL-10R mAb prior to GBS infection significantly increased the neutrophil recruitment in organs (Figure 9A and 9B). Consistently, the organs of pups treated with anti-IL10R mAb or anti-rGAPDH IgG contained significantly less bacteria than those of untreated pups (Figure 9C). These results indicate that an efficient neutrophil recruitment into infected organs is crucial for neonatal protection against GBS infection whereas impaired neutrophil recruitment facilitates GBS colonization. Moreover, no bacterial colonization was detected three weeks after GBS infection in the brain or in any other organ of pups born from rGAPDH-immunized mothers and infected with NEM316 or with the ST-17 hyper-virulent strain BM110 (data not shown). These results indicate that rGAPDH-maternal vaccination is also effective in preventing LOD.
Newborns are highly susceptible to infectious disease and deficiencies of key components of the complement cascade combined with the inability to produce high amounts of antibodies against T-independent antigens greatly impairs their ability to respond to encapsulated bacteria [39]. Absence of previous interactions with environmental microbes also implies that no immunological memory exists against specific antigens, which means that the acquired immune protection of newborns relies mainly on antibodies passively transferred from their mothers [40]. In addition, previous reports indicate that neonatal innate immune cells are less efficient in producing Th1-type inflammatory cytokines, but more competent in producing the immunosuppressive cytokine IL-10 upon Toll-like receptor (TLR) engagement by microbial products [41], [42], [43], [44]. Moreover, newborn mice leukocytes are highly committed to produce increased amounts of IL-10 [44], [45], as also shown in human neonates [41], [43], [46]. As report in this work, high levels of serum IL-10 could be detected in the sera of GBS infected newborn mice early (1 h and 4 h) upon infection. This result is in agreement with a previous report by Cusumano et al. [47], showing that elevated levels of plasma IL-10 were detected in newborn mice 24 h and 48 h upon GBS challenge. These authors suggested a host protective role for IL-10 in the outcome of neonatal GBS sepsis as pre-treatment of newborn mice with recombinant IL-10 improved their survival upon a lethal s.c. GBS challenge [47]. Nevertheless, they also showed that a therapeutic administration of this cytokine (24 h after the bacterial challenge) did not improve survival. This would limit its use in human therapies because neonatal GBS infection is usually acquired before or during labor [3], [5], [6].
Our results revealed that blocking IL-10 signaling through anti-IL-10R mAb administration was sufficient to confer protection against a bacterial challenge using either the s.c. or i.p. routes. Thus, they contrastingly indicate that IL-10 has a deleterious effect in the newborn host resistance to GBS infection. The increased resistance of IL-10-deficient neonates to GBS infection reported here constitutes further support for a deleterious effect of IL-10 in host resistance to GBS. As shown here, the GBS GAPDH induced host IL-10 production detected early after bacterial infection.
IL-10 is produced by multiple cell types and inhibits leukocyte activation, pro-inflammatory cytokine production and down-regulates the expression of anti-microbial molecules on activated phagocytes [48], [49], [50], [51]. IL-10 also inhibits production of CC and CXC chemokines by activated monocytes [52], [53], [54]. Since these chemokines are implicated in the recruitment of leukocytes during inflammation, IL-10 production indirectly inhibits leukocyte trafficking to inflamed tissues [50], [55], [56].
IL-10 production was already associated with host susceptibility against different pathogens [24], [45], [57], [58], [59], [60], [61]. We show here that treatment of pups with either anti-IL10R mAb or anti-rGAPDH IgG prior to the GBS challenge increased the neutrophil recruitment in liver and lungs that is triggered upon infection. Neutrophil recruitment is a crucial event in the host effector immune response to GBS [21], [62], [63] and, consequently, lack of neutrophil infiltration in infected sites has been reported in cases of severe early-onset GBS sepsis [21], [22], [23], [38]. Thus, neutralization of GAPDH, and hence blockade of the induced IL-10 production, allowed an effective immune response at an early stage of infection that prevented death of pups. Moreover, pups protected by maternal immunization with rGAPDH presented no GBS CFU in the brain, lungs, and liver 3 weeks upon the infectious challenge. This indicates that protection achieved by this vaccination strategy might prevent LOD.
The recruitment of neutrophils into infected tissues is very important to restrain bacterial replication. Thus, qualitative and quantitative deficiencies in the neutrophils of newborns may explain the observed susceptibility to GBS infections. Indeed, newborn neutrophils have reduced adhesion capabilities due to reduced expression of adhesion molecules [24], [25] and they produced a limited number of microbicidal molecules. Moreover, the number of these cells is also reduced when compared to adults due to insufficiencies on neonatal granulocyte lineage development [26]. As a consequence, the intra-cellular and extracellular killing of pathogens is greatly impaired in neonates [64]. However, our results indicate that, despite these serious functional defects, the neutrophils of neonates can control the GBS infections as long as the inhibitory effect of IL-10 is blocked. Importantly, IL-10 blockade with a specific mAb did not significantly decreased the elevated serum TNF-α levels detected upon GBS infection in neonate mice [47]. This further suggests that impairment of neutrophil recruitment rather than inhibition of pro-inflammatory cytokines could be the prominent effect of IL-10 produced in the course of neonatal GBS infection.
Neonatal sepsis is a pathological condition associated with elevated levels of pro-inflammatory cytokines, including IL-1β, TNF-α, and IL-6. However, it has been previously described that cord blood or plasma IL-10 concentration is significantly increased in neonatal sepsis, constituting an early indicator of prognosis [27], [28]. Of interest, it was also reported that high IL-10 levels are found in children at initial phases of fulminant septic shock [65], [66]. This indicates that early IL-10 production, instead of being a physiological attempt to counterbalance the elevated levels of pro-inflammatory cytokines, could be a predisposing factor for disease. Our results are in accordance with this hypothesis and they provide the first evidence that the lack of neutrophil recruitment in infected organs combined with elevated cord blood IL-10 concentration may account for neonatal susceptibility to GBS infection. Hence, the discovery of GAPDH as an extracellular virulence factor of GBS that induces an early IL-10 production by the infected host could be a significant contribution to our understanding of the pathology of neonatal infections.
GAPDH is a promising candidate for a human GBS vaccine because it is an essential metabolic enzyme that also plays a critical role in virulence. Our results show that maternal vaccination with rGAPDH protects the offspring against GBS lethal infection, including those caused by the hyper virulent ST-17 clone, which is responsible for most cases of neonatal meningitis [31], [32], [38]. As a consequence, maternal rGAPDH vaccination might efficiently protect against both EOD and LOD [5], [67], [68]. We demonstrated that passive immunization of neonates with GAPDH-specific IgG antibodies is sufficient to confer protection against GBS infection. Importantly, rGAPDH maternal vaccination prevents the early production of IL-10 in GBS infected pups and similar protective effect was obtained when GAPDH-specific antibody F(ab')2 fragments were used instead of whole IgG. These results indicate that neutralization of GAPDH-mediated IL-10 production, rather than complement activation or bacterial opsonophagocytosis, accounts for the observed protection. The extracellular GAPDH was detected at the bacterial surface and in culture supernatants of GBS isolates, which suggests that neutralization of its biological activity by antibody binding should not be sterically impaired by surface capsular polysaccharides. Recently, Margarit et al. showed that pili proteins could be used as a human vaccine to prevent GBS infections but, due to sequence variability, a combination of 3 antigens was required to confer protection against 94% of contemporary GBS strains [19]. It is likely that under selective pressure this vaccine will select GBS variants expressing new pili antigens, as shown for Neisseria gonorrhoeae [69], [70], [71], [72], [73]. In contrast, since it is an essential and highly conserved metabolic enzyme, GAPDH is unlikely to accumulate rapidly escape mutations or rearrangements under such a selective immune pressure.
Taken together, our results demonstrate that extracellular GAPDH confers a selective advantage to GBS for survival in the infected host. In particular, GBS GAPDH acts on the host immune system to elicit IL-10 production thereby favoring bacterial colonization and survival. As we demonstrated that GBS GAPDH was still able to induce host IL-10 production upon exposure to an oxidative agent, this mechanism may still operate within the highly oxidative environment resulting from the host inflammatory response. Our data highlight the critical role played by this immunosuppressive cytokine in determining susceptibility to GBS infection at an early time after birth. Our results also show that GBS-associated pathology can be counteracted either by rGAPDH vaccination or IL-10 neutralization. In the future, it will be essential to explore the use of either strategy to induce protection towards other human neonatal pathogens.
Relevant characteristics of the GBS strains used in this study are summarized in Table S1. Escherichia coli BL21 (DE3) strain (Novagen) and the pET28a plasmid (Novagen) were used for production of recombinant GAPDH (rGAPDH) as described previously [29]. GBS was grown in Todd-Hewitt broth or agar (Difco Laboratories) containing 0.001 mg/mL of colistin sulphate and 0.5 µg/mL of oxalinic acid (Streptococcus Selective Supplement, Oxoid) and E. coli was cultured on Luria-Bertani medium. Bacteria were grown at 37°C.
Male and female BALB/c mice (6-8 weeks old) were purchased from Charles River. IL-10-deficient BALB/c (IL-10−/−) mice were kindly provided by Dr. A. O'Garra (National Institute for Medical Research, London, U.K.). New Zealand White rabbits were purchased from Charles River. Animals were kept at the animal facilities of the Institute Abel Salazar during the time of the experiments.
This study was carried out in strict accordance with the recommendations of the European Convention for the Protection of Vertebrate Animals used for Experimental and Other Scientific Purposes (ETS 123) and 86/609/EEC Directive and Portuguese rules (DL 129/92). The animal experimental protocol was approved by the competent national authority Direcção Geral de Veterinária (DGV) (Protocol Permit Number: 0420/000/000/2008). All animal experiments were planned in order to minimize mice suffering.
Recombinant GAPDH (rGAPDH) was purified as described in detail previously [29]. Enzymatically inactive rGAPDH (inact-rGAPDH) was obtained by pretreatment of the enzyme with 500 µM H2O2. The lack of enzymatic activity upon inactivation was confirmed using a previously described enzymatic assay for GAPDH [29].
Recombinant GAPDH was used for maternal immunization assays. Female mice were injected intraperitoneally (i.p.) twice, with a 3-week intervening period, with 200 µL of a preparation containing 25 µg of rGAPDH in a 1:20 PBS/alum suspension (Aluminium hydroxide Gel; a kind gift of Dr Erik Lindblad, Biosector, Frederickssund, Denmark). The sham-immunized control animals received 200 µL of a 1:20 PBS/alum suspension. Immediately after the second injection, female and male mice were paired. Females were monitored closely during gestation and the day of delivery was recorded. Serum anti-rGAPDH antibody titers were determined by ELISA as previously described [29].
Antibody treatments were performed in newborn BALB/c mice 12 h prior to GBS infection. For passive immunizations, pups were injected i.p. with 100 µg of anti-rGAPDH IgG antibodies or anti-rGAPDH F(ab')2 fragments. Control animals received the same amount of control IgG's or F(ab')2 fragments issued from control IgG's. For IL-10 signaling blocking, 100 µg of anti-IL10R antibodies (1B1.3a, Schering-Plough Corporation) were administered i.p. and control animals received the same amount of matched isotype control antibody.
Newborn mice were infected i.p. with 5×106 cells of GBS NEM316 or 106 cells of GBS BM110 (ST-17), 48 h after birth in a maximum volume of 40 µL. Subcutaneous (s.c.) infections were performed 48 h with 2.5×104 cells of GBS BM110 after birth in a total volume of 20 µL. Survival curves were determined in a 12-day experiment period and newborns were kept with their mothers during the entire time of the experiment. The liver, lungs, and brain of infected pups were aseptically removed at indicated time points and homogenized in PBS and serial dilutions of homogenized organs were plated on Todd-Hewitt agar to enumerate bacterial CFU.
Adult mice or rabbit were immunized twice with 25 µg of rGAPDH in a PBS/alum suspension as described above and sera were collected 10 days after the second immunization. Pooled serum samples were applied to a Protein G HP affinity column (HiTrap, GE Healthcare Bio-Sciences AB) and purified IgG antibodies were then passed through an affinity column with immobilized rGAPDH (Hi-trap NHS-activated HP, GE Healthcare Bio-Sciences AB). Control IgGs were obtained from sera of mice or rabbits sham-immunized with a PBS/alum suspension and purified on a Protein G HP affinity column. Purified IgG antibody fractions were further equilibrated in PBS and stored at -80°C in frozen aliquots.
GBS-specific antiserum was obtained from mice immunized i.p. twice (with a 3-week interval) with isopropanol-fixed 105 GBS cells plus alum (total volume). Serum from immunized animals (Anti-GBS serum) was obtained from retro-orbital bleeding 10 days after the second immunization.
F(ab')2 fragments from anti-rGAPDH or control IgGs were obtained using IgG1 F(ab) and F(ab')2 Preparation Kit (Pierce) used according to manufacturer's instructions.
Bone marrow-derived macrophages (BMM) purified as described previously [74] were plated in 96-well plates (105 BMM/ well) and stimulated for 30 min at 37°C 5%CO2 with 106 CFU of GBS BM110 (or NEM316) in the medium alone (RPMI), the medium containing 25 µg/mL of anti-rGAPDH IgG's, or medium with 10% of serum containing anti-GBS IgG antibodies. After this incubation period, the plates were washed three times with HBSS to remove extracellular bacteria. To enumerate intracellular GBS CFU, 10% saponin (1∶100 dilution) was added to wells and serial dilutions of supernatant were plated onto agar plates.
Blood from adult mice was collected in heparinated tubes and diluted 1:1 in HBSS. 106 GBS NEM316 (or BM110) CFU with 25 µg/mL of anti-rGAPDH IgG's or 10% of serum containing anti-GBS IgG antibodies were then added. Rabbit serum (5%) was added to the mixture as a source of complement. After 2 h of incubation at 37°C, serial dilutions of the mixture were plated onto agar plates to evaluate complement-mediated GBS killing.
The BMM, obtained as described previously [74], were infected with GBS strains NEM316 and BM110 at a macrophage:bacteria ratio of 1:10 in RPMI containing 10% FCS. Microplates were incubated for 2 h at 37°C in 5% CO2 for GBS phagocytosis. After this period, culture supernatants of infected macrophages were removed by aspiration and cells were washed three times (10 min for each wash) with HBSS containing penicillin (100 IU/mL) and streptomycin (50 µg/mL) to kill extracellular bacteria. Infected macrophages were further incubated in RPMI medium containing 10% FCS and the same concentrations of antibiotics. To quantify intracellular GBS, the supernatants containing antibiotics were removed, the cells were washed with antibiotic-free HBSS, lysed with saponin (0,1% final concentration), and the CFU were estimated by plating serial dilutions of the lysate onto agar plates.
Neutrophil recruitment in liver and lungs of infected pups was evaluated by flow cytometry analysis. Briefly, 18 h after GBS infection, the organs were collected, gently homogenized in HBSS (Sigma), and passed through glass wool to remove cellular aggregates. PerCP/Cy5.5 anti-mouse Ly-6G antibody (clone 1A8; Biolegend) was used for neutrophil detection. Cells were analyzed by an Epics XL cytometer (Beckman Coulter).
Newborn mice were depleted of neutrophils by treatment with anti-Ly6G antibodies (clone 1A8, Biolegend). Antibody treatment was performed twice, 12 h before and immediately after GBS challenge. Each pup was injected with a total of 80 µg of anti-Ly-6G antibodies.
IL-10 was quantitated in the serum of newborn mice with an ELISA kit (eBioscience) used according to the manufacturer's instruction.
The presence of GAPDH in the culture supernatants of GBS strains was visualized by Western-blot analysis. Extracellular proteins were isolated as described previously [29]. The reactivity of purified anti-rGAPDH IgG antibodies obtained from the serum of rGAPDH immunized mice or rabbits against self or human GAPDH, was determined by Western-blot analysis or ELISA. Human, rabbit, and mouse GAPDH were purified from human erythrocytes, rabbit erythrocytes or mouse muscle as previously described [75], [76].
Student's T test was used to analyze the differences between groups. Survival studies were analyzed with the log-rank test. A P value<0.05 was considered statistically significant.
|
10.1371/journal.pntd.0005524 | Cognitive deficits and educational loss in children with schistosome infection—A systematic review and meta-analysis | By means of meta-analysis of information from all relevant epidemiologic studies, we examined the hypothesis that Schistosoma infection in school-aged children (SAC) is associated with educational loss and cognitive deficits.
This review was prospectively registered in the PROSPERO database (CRD42016040052). Medline, Biosis, and Web of Science were searched for studies published before August 2016 that evaluated associations between Schistosoma infection and cognitive or educational outcomes. Cognitive function was defined in four domains—learning, memory, reaction time, and innate intelligence. Educational outcome measures were defined as attendance and scholastic achievement. Risk of bias (ROB) was evaluated using the Newcastle-Ottawa quality assessment scale. Standardized mean differences (SMD) and 95% confidence intervals (CI) were calculated to compare cognitive and educational measures for Schistosoma infected /not dewormed vs. uninfected/dewormed children. Sensitivity analyses by study design, ROB, and sequential exclusion of individual studies were implemented. Thirty studies from 14 countries, including 38,992 SAC between 5–19 years old, were identified. Compared to uninfected children and children dewormed with praziquantel, the presence of Schistosoma infection and/or non-dewormed status was associated with deficits in school attendance (SMD = -0.36, 95%CI: -0.60, -0.12), scholastic achievement (SMD = -0.58, 95%CI: -0.96, -0.20), learning (SMD = -0.39, 95%CI: -0.70, -0.09) and memory (SMD = -0.28, 95%CI: -0.52, -0.04) tests. By contrast, Schistosoma-infected/non-dewormed and uninfected/dewormed children were similar with respect to performance in tests of reaction time (SMD = -0.06, 95%CI: -0.42, 0.30) and intelligence (SMD = -0.25, 95%CI: -0.57, 0.06). Schistosoma infection-associated deficits in educational measures were robust among observational studies, but not among interventional studies. The significance of infection-associated deficits in scholastic achievement was sensitive to ROB. Schistosoma infection-related deficits in learning and memory tests were invariant by ROB and study design.
Schistosoma infection/non-treatment was significantly associated with educational, learning, and memory deficits in SAC. Early treatment of children in Schistosoma-endemic regions could potentially mitigate these deficits.
ClinicalTrials.gov CRD42016040052
| Empirical evidence for cognitive or educational benefits of anti-Schistosoma treatment is currently uncertain, despite the recommended practice of wide-scale deworming with praziquantel. We addressed this knowledge gap by synthesizing information from 30 relevant epidemiologic studies reporting on 38,992 children between 5–19 years old from 14 countries. In those studies, Schistosoma infection or non-dewormed status was associated with educational loss and cognitive deficits. Specifically, there were small to moderate deficits in both school attendance and scholastic achievement. Similarly, Schistosoma infection or non-dewormed status was associated with deficits in learning and memory domains of psychometrically tested cognitive function. However, there was no evidence of Schistosoma infection- or non-deworming-associated deficits on tests of innate intelligence or reaction-time. Overall, compared to Schistosoma-uninfected or to dewormed children, the presence of Schistosoma infection or non-dewormed status was associated with educational, learning, and memory deficits in school-aged children. The combined evidence suggests that early treatment of children in Schistosoma-endemic regions could mitigate these deficits.
| An estimated 800 million persons in tropical and sub-tropical countries are at risk of infection by one of three main human Schistosoma parasites–S. mansoni, S. haematobium, and S. japonicum [1]. As many as 240 million adults and children are actively infected [2–4] resulting in as much as 3.3 million disability-adjusted life years (DALYs) lost per annum due to overt and subclinical morbidities of Schistosoma infection [4, 5]. Sub-Saharan Africa is most affected; children from endemic areas are often infected by two years of age and many remain chronically infected throughout their school-age years [6–8]. Periodic mass drug administration (MDA) with praziquantel in school-aged children has been recommended for morbidity control by the World Health Organization [6]. However, Schistosoma-infected pre-school children are not routinely treated in such settings, and they constitute a potentially high risk group for accumulation of morbidity [9]. At present, there is no specific guidance for anti-Schistosoma drug treatment of preschool children, partly because of the lack of a child-friendly pediatric formulation [10].
Treatment with praziquantel has a demonstrated effectiveness in reducing infection intensity within individuals and in reducing the prevalence of infection within communities [11]. Such treatment results in clear-cut improvements with respect to advanced schistosomiasis-associated morbidities such as urinary tract fibrosis and hepatosplenic disease, including peri-portal fibrosis [12, 13]. Epidemiologic studies have associated Schistosoma infections with adverse impacts on anemia, growth [14], fitness [15], pediatric quality-of-life [16], and sub-optimal child development [17]. Definitively linking Schistosoma infections to these non-specific and sub-clinical morbidities is complicated in the context of poverty.[8, 12, 17, 18] However, plausible biologic mechanisms of these adverse impacts have been described [2, 19] and the likely underestimation of their morbidity-related health impact has been highlighted [3].
In helminth-endemic regions, the recommendation of periodic deworming of children is explained on the basis of its expected salutary impact on a range of child-health outcomes including anemia [12], nutritional status [14, 17, 18, 20], and overall well-being [16]. In addition, periodic deworming has been linked directly or indirectly to enhancement of school attendance and educational achievement among children enrolled in school [6, 14, 21, 22]. However, the empirical evidence-base for cognitive and educational benefits of deworming remains controversial [23–27]. Recent criticism of the supposed benefits of deworming for educational enhancement has emphasized the undue influence of a single study in evidence reviews [28], which may have led to over-optimistic appraisals regarding the potential health and poverty alleviation benefits of deworming programs [29, 30].
Recent reviews of MDA effects have, thus far, focused on the impact of soil-transmitted helminth infections (STH), but health policy discussions–including those at the World Health Organization [30, 31], have tended to generalize findings to all helminths. Because different parasites can have dramatically different effects in terms of organ-specific and systemic pathologies, it is important now to distinguish the impact and potential benefits of individual anti-helminthic therapies [32]. To date, the evidence base for cognitive or educational benefits of anti-Schistosoma treatment has not undergone systematic review. The present systematic review and meta-analysis addresses the following questions: a) among school-aged children examined in the context of cross-sectional or case-control studies, is Schistosoma infection associated with worse performance in neurocognitive tests or with educational loss? b) among school-aged children enrolled in prospective studies with specific treatment for Schistosoma infection, is lack of treatment with praziquantel associated with worse performance in neurocognitive tests or with educational loss? For our current meta-analysis, we hypothesize that non-treatment or infection with Schistosoma infection is associated with educational loss and cognitive deficits in school-aged children from schistosomiasis-endemic regions.
This review, with pertinent information regarding our review protocol, was prospectively registered with the PROSPERO database as follows: http://www.crd.york.ac.uk/PROSPERO/display_record.asp?ID=CRD42016040052 (see Supporting Information file S1 Text). We searched Medline, Web of Science, and Biosis electronic databases for original research articles, conference abstracts, or dissertations available as of August 22, 2016. Databases were searched with pre-specified keywords including “bilharzia”, “schistosomiasis”, “Schistosoma”, “school attendance”, “attention”, “impairment”, “memory”, “cognition”, among others. The complete search strategy is detailed in S2 Text.
This systematic review and meta-analysis is focused on school-aged children five years and older. We did not restrict studies according to language, design, or publication date. Both interventional and observational studies were included in this review if they evaluated cognitive function in school-aged children using any psychometric test, or measured school attendance or achievement in relation to infection by Schistosoma parasites of any species. We excluded studies exclusively focused on pre-school aged children because of the absence of educational measures and the use of neurocognitive tests (e.g., the Mullen test) that were difficult to classify in terms of neurocognitive domains. We excluded meta-analyses and primary studies of soil-transmitted helminth infections where Schistosoma coinfection was absent.
Schistosoma infection status was determined by microscopic examination of stool or urine as appropriate for species. Praziquantel was the primary deworming agent in interventional studies. The primary exposure for this meta-analysis included presence of infection or, operationally, infection was categorically defined based on study design as follows: 1) untreated/placebo versus praziquantel-treated in a randomized controlled trial, 2) any, versus no Schistosoma infection in cross-sectional studies, or 3) pre-, versus post-praziquantel treatment, or infection-free versus persistent infection, among Schistosoma-infected individuals in a longitudinal design study. “Untreated” refers to children determined to be infected but not dewormed.
In the reviewed studies, psychometrically-assessed cognitive function was measured by a range of instruments that, for the purposes of this meta-analysis, were categorized into four domains: memory, learning/executive function, attention/reaction time, and intelligence (see S1 Table for details). The memory domain instruments included tests of working (short-term) memory as well as those of long-term memory. Attention/reaction time tests were those that measured the ability of a child to sustain concentration on a particular object, action, or thought, including their capacity to manage competing demands in their environments. The learning/executive function domain included tests to evaluate children’s performance in goal-oriented behavior, particularly components that are important for scholastic advancement. These test a cluster of cognitive processes that enable children to connect past experience with present action, and by so doing, engage in planning, organizing, strategizing, paying of attention to details, and to emotionally self-regulate, make necessary efforts to remember important details required for attainment of future goals [29]. We included in the ‘intelligence’ domain psychometric tests of intelligence quotient (IQ) that likely measures largely biologically-determined cognitive abilities, in contrast to cognitive performance measures that are environmentally pliable [33].
It was common for studies to use a suite of psychometric instruments to assess a single or multiple cognitive domains in enrolled children. When multiple instruments were used to measure the same cognitive domain, a grand mean of scores and a grand mean of standard deviation (SD) across all instruments were calculated. Thus, for each publication, one overall mean and SD value was determined for each domain. A study could contribute data to different cognitive domains if it used tools spanning across several cognitive domains. However, each instrument only contributed to one single domain of function, as shown in S1 Table.
Two dimensions of educational loss were tabulated–school attendance and scholastic achievement for children enrolled in school.
School attendance: For children enrolled in cross-sectional and longitudinal studies, attendance rate was respectively defined as the number of days children attended school over the past month or over the study period. In case-control studies, the percentage of children enrolled vs. not enrolled in school was calculated for Schistosoma-infected and non-infected children.
School achievement was assessed across studies based on: i) children’s pass rate on standardized teacher-generated tests; ii) the percent of children who were in appropriate class for age; iii) their enrollment in elite vs. non-elite schools; iv) their scores in the school function domain of pediatric quality-of-life inventory; v) their change in class position after treatment for Schistosoma infection; vi) an above average vs. average/below average scholastic performance as rated by a teacher; or vii) their pass rate in any kind of educational test, whether teacher-administered or not.
Two researchers (LM and AK) independently screened individual articles by title and abstract, after which eighty-eight full text articles were assessed for eligibility for inclusion in this review. Studies were excluded on the following basis: no outcome measure reported (n = 39), non-primary literature or a review article (n = 7), absence of both infection and outcome measures (n = 6), no variation in exposure (n = 2), a limited meeting abstract duplicated by later full publication (n = 3) and a nonhuman study (n = 1) Disagreements between reviewers on inclusion of a given study were resolved by consensus between LM and AK. If no consensus was reached, the article was further evaluated by an additional reviewer (AEE). Thereafter, two researchers (AEE and NP) independently extracted relevant data for meta-analyses. Where differences in approach to standard error (SE) estimation were noted, discrepancy was resolved by consensus between AEE and NP. When a potentially relevant publication did not present needed information for meta-analysis, the authors were contacted to request additional data. If the dataset was publicly available, we obtained needed values directly [28].
The method for deriving SD from respective studies depended on how data were presented in the original reports. Some papers presented median (m) and range (a to b) instead of means and SD. These measures were converted into approximate mean and SD as follows: x¯≈a+2m+b4,S2≈112((a−2m+b)24+(b−a)2), see [34], where x¯ and S2 refer to the values of mean and variance, respectively. Some studies reported mean of respective measures and 95% confidence intervals. For such studies, SDs were derived as follows: SD = sqrt(N)*(upper limit–lower limit)/3.92. Other studies presented data on means and their standard errors (SE), and SD was estimated as SD = SE*the square root of N, the study size. For studies presenting data on differences in mean scores between two time points for treated/infected vs. untreated/uninfected groups, appropriate SD for mean difference was calculated using the approach recommended by the Cochrane Collaboration [35].
For the meta-analysis, studies were grouped into six categories of outcomes used to measure cognitive or school-based function: school attendance, school achievement, memory, learning, IQ, or attention. Standardized mean difference (SMD) estimates and 95% confidence intervals (CI) were calculated for each test. SMD estimates were classified as robustly statistically different if their confidence intervals excluded zero. SMD estimates were interpreted based on thresholds described by Cohen [36], as follows: ‘trivial’ (< |0.20|), ‘small’ (≥| 0.20| to < |0.50|), ‘moderate’ (≥ |50| to < |0.80|) or ‘large’ (≥ |0.80|) effects, according to standard practice in social science research.
All analyses and plots were implemented in STATA, versions 11 or 12. Heterogeneity between studies was measured with Higgins’s and Thompson’s I2 statistic and chi-square p-values [37]. Where between-study heterogeneity was high, random effects modelling was used to estimate a pooled summary effect across studies [38, 39]. In the absence of heterogeneity, fixed effects modeling was performed [39]. Publication bias was assessed using the Egger test [40]. In sensitivity analysis, we evaluated potential heterogeneity in pooled impact estimates based on: i) observational vs. intervention study design; ii) the quality of original studies based on the Newcastle-Ottawa quality assessment scale; and iii) by Schistosoma species. For sensitivity analyses by species, we distinguished between urogenital schistosomiasis (S. haematobium), which is often obvious to affected children, and intestinal/hepatosplenic schistosomiasis (S. mansoni/S. japonicum). The latter two infections are similarly diagnosed by stool exam and infection is seldom obvious to most children. We examined the potential for overly influential publications using the ‘metabias’ function in STATA to evaluate robustness of our pooled estimate, based on sequential removal of individual publications from the calculation of summary estimates. Lastly, we evaluated the impact of year of publication on the stability of the pooled estimate by iteratively including studies based on year of publication–i.e. starting from the earliest to the most recent publication using the ‘metacum’ function in STATA.
Our investigation was guided by recommendations of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) initiative and the Meta-analysis of Observational Studies in Epidemiology (MOOSE) guidelines for observational studies [41]. Quality ranking of each study was implemented using an adapted version of the Newcastle-Ottawa quality assessment scale (NOQAS) [42, 43] to derive a quality score for each investigation with respect to: i) representativeness of the infected population sample or the selection cases and controls (this yielded a score with range 0 to 3*); ii) comparability with respect to known correlates of cognitive function/educational attainment (score range 0 to 6*); iii) the absence of bias in relation to outcome assessment in prospective cohort studies (0 to 3*) or exposure assessment in cross-sectional and case-control studies (0 to 3*). We adapted the comparability segment of this scale to account for confounding effects of age, sex, nutrition, and socioeconomic status in the relation between Schistosoma infection and educational/cognitive outcomes. Comparability with respect to these factors was either achieved by design (i.e. age or sex restriction for observational studies, or randomization for RCTs), or analytically, via stratified analyses or multivariable adjustment in regression models. Scores were assigned for attainment of comparability with respect to these factors as follows: age (score = 1*), sex (score = 1*), nutritional status (score = 2*) and socioeconomic status (score = 2*). For each study, the initial raw quality score (max = 12*) was rescaled to match the scale of 9* and then classified as low, high, or very high risk of bias based on precedents in prior literature [42].
Our database search yielded a total 2914 unique records. The screening of titles and abstracts resulted in the exclusion of 2846 records leaving 78 unique papers for full text review. We identified an additional set of ten relevant studies from the bibliographies of relevant articles. Of these 88 articles, 58 were excluded based on full text review for reasons pecified in Fig 1.
A total of 30 epidemiologic studies that assessed differences in cognitive test scores (based on psychometric tests) and/or educational status (measured as scholastic achievement or school attendance rate) in relation to Schistosoma infection or treatment were selected for inclusion (Table 1). Of these, 21 studies were cross-sectional [16, 44–56] or case-control [57–63] design where children were classified based on presence vs. absence of Schistosoma infection. Seven were longitudinal studies or pre-post intervention studies that featured screening and treatment for Schistosoma infections at the time of first assessment [28, 64–68]. In these studies, outcome contrasts were made according to: a) the duration of persistent infection vs. duration of an infection-free interval, and b) the number of children for whom intensity of infection at last follow-up remained lower vs. the number of children for whom there was no change from baseline infection status after treatment with praziquantel.
Two studies were randomized controlled studies. Only one included study was a classic placbo-controlled randomized-trial intervention in which children with Schistosoma infection were randomized to praziquantel vs. placebo/no treatment [69]. The other study randomized children to screening vs. non-screening for Schistosoma infections [70]. Children in the non-screened arm remained untreated (although that sample was subsequently tested for infection to distinguish infected from uninfected children). In that study, among children randomized to screening, those found to be Schistosoma-infected were given treatment, and we were thus able to derive differences in cognitive test scores between Schistosoma-infected/treated vs. Schistosoma-infected/not-treated children (Table 2). The median follow-up duration across the nine logitudinal studies was 12 months. Minimum follow-up duration was one month and maximum follow-up duration was 36 months. In four of the longitudinal studies, follow-up duration was 6 months or less. In another four studies, follow-up duration was more than 12 months. One study had a 12 month follow-up.
In all, a total of 38,992 children between the ages of 5 to 19 years from 14 countries in three continents–Africa, Asia and North America, were included in this review. The vast majority of studies were of children from Africa–Nigeria (n = 4), Egypt (n = 4), South Africa (n = 4), Tanzania (n = 3), Kenya (n = 4), Mali (n = 1), Côte D’Ivoire (n = 1), Zambia (n = 1); Ghana (n = 1) and Ethiopia (n = 1). Three studies were conducted in Southeast Asia or Asia [The Philippines (n = 2) and China (n = 1)], one study was conducted in St. Lucia (Caribbean), and another study was implemented in Madagascar.
A total of 36,626 children were studied in the context of cross-sectional or case control studies. Of these, cognitive test scores and/or indicators of educational loss were measured in 23,126 (59.3%) children infected with one of three schistosomiasis species (S. haematobium (20 studies), S. japonicum (3 studies), or S. mansoni (10 studies)). These infected children were compared to 13,835 (40.6%) children without Schistosoma infection. A total of 2,366 children were studied in the context of randomized intervention studies including praziquantel vs. placebo, or prospective treatment studies that included baseline and follow-up assessments of cognitive function (Table 2). Of the 30 studies included in this review, three (10%) and 16 (53.3%) were judged to be at very high or high risk of bias, respectively, per the NOQAS (Table 3).
Sixteen studies evaluated Schistosoma infection-related differences in school attendance (Table 2). Schistosoma infection-associated attendance deficits varied in magnitude and direction by study design. Specifically, we did not find any evidence of association between Schistosoma infection and school attendance among the two interventional studies (SMD = 0.03, 95%CI: -0.73, 0.78; Table 4); however, the observed infection-associated deficit in school-attendance was robust for the pooled estimate of 14 observational studies (SMD = -0.42, 95%CI: -0.70, -0.14). Compared to uninfected or praziquantel-treated children, the magnitude and direction of infection-associated deficit in school attendance was similar for children infected with S. haematobium or S. mansoni. Within strata of study quality, the association between infection and scholastic achievement was directionally consistent and statistically robust (Table 4). Across all studies, regardless of design or ROB, a deficit in school-attendance was evident for Schistosoma-infected or non-praziquantel-treated children compared to uninfected or praziquantel treated children (n = 15 studies; SMD = -0.36, 95% CI: -0.60, -0.12).
The overall finding of schistosomiasis-associated achievement deficit (n = 16 studies; SMD = -0.58, 95%CI: -0.96, -0.20) was directionally consistent but varied in magnitude by study design and study quality (Tables 3 and 4). Specifically, achievement deficit with Schistosoma infection was noted in the pooled estimate derived from four intervention studies, but the magnitude of this association was lower and not statistically robust (SMD = -0.35, 95%CI: -0.71, 0.01; Table 4). Compared to uninfected or PZQ-treated children, the infection or non-treatment associated deficit in scholastic attainment was directionally consistent across different Schistosoma species, although the magnitude of effect was higher for infection with S. haematobium (SMD = -0.62; 95% CI:-1.09, -0.14) than for infection with S. mansoni (SMD = -0.22; 95% CI:-0.40, -0.05). Among observational study designs, scholastic achievement deficit was statistically robust and of larger magnitude (n = 12 studies; SMD = -0.65, 95% CI: -1.12, -0.17, Table 4). Similarly, Schistosoma infection was not associated with deficits in scholastic achievement among studies identified as low risk of bias (n = 6 studies; SMD = -0.08, 95%CI: -0.21, 0.02; Table 4). The estimates of infection-related deficit in scholastic achievement increased when study quality was lower: for studies with high risk of bias (n = 7) the SMD = -0.84 (95% CI:-1.52, -0.16); for studies with very high risk of bias (n = 7) SMD = -0.92 95% CI: -2.1, 0.28) although the estimated difference was statistically imprecise (i.e., with a wider CI) among the studies with greatest risk of bias.
We found Schistosoma infection-associated deficits in memory tests (n = 8 studies; SMD = -0.28, 95% CI: -0.52, -0.04; Table 2). Similarly, Schistosoma infection was associated with small-to-moderate deficits in learning tests (n = 6 studies; SMD = -0.39, 95%CI: -0.70, -0.09; Table 2). The Schistosoma infection-related deficits in memory tests were directionally consistent by study design; although separate pooled estimates for interventional studies (n = 4 studies; SMD = -0.36, 95% CI: -0.61, 0.09) and observational studies (n = 4 studies; SMD = -0.15, 95% CI: -0.34, 0.03) were not statistically robust. For Schistosoma infection-related association with learning, pooled estimates suggested the presence of deficits for infected/non-dewormed vs. uninfected/praziquantel-treated children (n = 6 studies; SMD = -0.39, 95% CI: -0.70, -0.09). However, the magnitude of this association differed according to study design. Here, larger pooled standardized differences were realized for interventional studies (n = 2 studies; SMD = -0.79, 95% CI:-1.19, -0.39) than for observational studies (n = 4 studies; SMD = -0.18, 95% CI: -0.35, -0.01). Schistosoma infection was not significantly associated with performance on tests of reaction time (n = 6 studies; SMD = -0.06, 95% CI: -0.42, 0.30) or performance in tests of innate intelligence (n = 4 studies; SMD = -0.25, 95% CI: -0.57, 0.06). For all psychometrically-assessed cognitive domains, the overall findings were not sensitive to study quality. Of note, the majority of publications with psychometrically-evaluated endpoints had low risk of bias and produced pooled estimates similar in magnitude and direction to the overall results (Table 4).
High levels of heterogeneity were observed across included studies (P ≤ 0.03, I2 ≥ 74.8%; Tables 2 and 4) for all outcome measures, but there was no evidence of undue influence by individual studies or of publication bias among the included studies (Egger’s test, all P-value ≥ 0.142, Table 2). Overall, inferences based on pooled estimates were generally insensitive to differences in study design (Table 4), to the exclusion of individual studies (S2 Table), or to the influence of publication year (S1 Fig).
This systematic review and meta-analysis of the cognitive and educational impact of Schistosoma infection in school-aged children supports the hypothesis that infection is associated with reduced school-attendance, with deficits in scholastic achievement and deficits in memory and learning domains of psychometrically evaluated cognitive function. It has previously been conjectured that Schistosoma infection may affect school attendance, scholastic achievement, and cognitive function, either directly through via deposition of Schistosoma eggs within the central nervous system, via physical discomfort and subsequent distraction due to the presence of the worms, or indirectly, via iron-deficiency and malnutrition [3, 71, 72]. However, Schistosoma infection or non-treatment was not associated with performance in tests of innate intelligence or reaction time. Inferences based on most pooled estimates for psychometrically assessed endpoints were generally insensitive to study design, Schistosoma species, and risk of study bias. However, associations between infection and educational outcomes were sensitive to study design and study quality–especially estimates for impact on scholastic achievement. The association between infection and scholastic achievement was directionally similar and statistically robust regardless of Schistosoma species; however, average effects were substantially larger for S. haematobium compared to S. mansoni infection. Cohen’s criteria for effect size suggest that the average Schistosoma infection-related deficits in education, learning, and memory performance range from ‘small’ to ‘moderate’. S. haematobium infection was associated with relatively larger deficits in scholastic achievement.
Prior meta-analysis of four randomized controlled trials that evaluated cognitive impacts of STH infections–which sometimes co-occur with schistosomiasis, have reached a different conclusion about the cognitive and scholastic effects of STH infection and the impact of interval treatments for STH [25, 26]. Those reviews concluded that there was substantial evidence that deworming for soil-transmitted helminth infections does not yield a cognitive or educational benefit. We note that our approach differed from these prior STH-based meta-analyses on several grounds: a) we included both interventional and observational studies to take advantage of all research data available on this question, b) we evaluated Schistosoma-associated impacts on two domains of educational loss (attendance and achievement), c) we further defined domains of cognitive function based on psychometrically-assessed testing to include the following: learning, memory, attention, and intelligence and d) the intervention, as defined in this meta-analysis, denoted treatment for Schistosoma infection, whether or not the study was randomized.
We intentionally included data from all available epidemiologic studies–whether interventional or observational in design, in this first systematic review and meta-analysis of Schistosoma infection-related differences in cognitive and educational outcomes. The inclusion of all available evidence reflects current standards for clinical evidence-gathering to inform health policy, in order to shape clinical practice based on the ‘best available’ relevant information [23, 24]. Future randomized-controlled trials to address this question are expected to be limited in scope and may be considered ethically objectionable given the current widespread adoption of deworming for schistosomiasis and STH. The current adoption of ‘preventive chemotherapy’ guidelines has been based on helminth-associated adverse effects on anemia and child growth. In consequence, meta-analysis of available evidence, as performed in the present study, remains the most practical strategy to inform current policy. By this approach, we have identified ‘small to moderate’ infection-related deficits in education, learning, and memory performance using the Cohen’s criteria of effect size. However, such numerically small deficits (per Cohen’s criteria) may significantly underestimate clinical significance of infection for childhood development, as Schistosoma infection is an exposure affecting millions of children in endemic regions. Hence, small to moderate deficits at the individual level may amount to large and important differences in disease burden at the population level [73, 74]. Our interpretation of SMD estimates is ultimately grounded in the importance of summary measures for clarifying existing knowledge gaps regarding the relationship of Schistosoma infections to respective outcomes, and the expected benefit of systematically lowering infection-related deficits in the millions of children at risk.
As a limitation of our approach, we acknowledge the possibility of residual confounding and bias in the primary literature, given that majority of included studies were observational or non-randomized intervention trials. Only two of the 30 included studies used an RCT design, thus a sensitivity analysis based on RCT vs. non-RCT study design was not possible. However, as part of sensitivity analysis we evaluated the potential for differences in pooled estimates based on our expanded definition of intervention as including longitudinal studies that included praziquantel treatment. Investigation of pooled estimate sensitivity by region of study, the year of publication, and by risk of bias did not result in materially different statistical inferences. Specific investigation of publication bias suggests that any possible greater likelihood of publishing positive studies did not unduly influence observed findings. Our approach of including both interventional and non-interventional studies is consistent with theoretical and empirical evidence that meta-analyses based on observational studies generally produce estimates of effect similar to those from meta-analyses based on randomized controlled trials, and that a priori exclusion of observational studies in systematic reviews is inappropriate and inconsistent with the evidence-based medical decision-making approach [75, 76]. In addition, the often restrictive inclusion criteria and short follow-up duration in RCTs could easily result in outcomes largely different from when the same interventions are applied to a general population.
Despite the intuitive appeal of our cognitive domain based evaluation, we acknowledge the critique that our classification of psychometric instruments by domain required some level of subjectivity, especially for tools that capture performance across multiple domains. We explicitly identified the instruments used in each study and, based on literature description of the major cognitive domain assessed by each tool, combined related tools into four separate domains. Ultimately, each instrument was assigned to one cognitive domain only. We have described the logic for our choices in the supplementary information (S1 Table) to provide a basis for further discussion and give sufficient context for critical evaluation of our approach in developing future studies.
Our decision to combine psychometric evaluations of cognitive functions in four domains (based on the primary capacity being tested) is a strength of our empirical approach. By so doing, we recognize that Schistosoma infection may not have equal impact on all cognitive domains. For example, to the extent that innate intelligence is strongly influenced by fixed or heritable factors, we did not anticipate infection related differences on tests of ‘intelligence quotient’. Unlike intelligence tests, we considered the other tests of memory, learning, and reaction time to be more sensitive to infection, and thus modifiable by presence/non-treatment vs. absence/treatment of Schistosoma infection. Unexpectedly, reaction time was not associated with Schistosoma infections. However, our findings of infection-related reductions in learning and memory tests are consistent with our hypotheses that these cognitive domains are sensitive to adverse environmental perturbations–including Schistosoma infection.
The finding of infection-related cognitive deficits and educational loss reported here is clinically and health policy-relevant for mitigating the cognitive and functional morbidities of Schistosoma spp. infection in children. The ‘small-to-moderate’ effects demonstrated may, in reality, be an underestimate of the lifetime impact of Schistosoma infection on personal performance (as affected via ultimately irreversible educational and cognitive losses). Typical epidemiologic studies necessarily include only a constrained portion of the relevant etiologic period. Among school-age children, Schistosoma infection is often effectively already chronic and/or recurrent, with reinfection rates extremely high in the absence of meaningful environmental interventions to reduce re-infection following treatment. The cumulative cognitive and educational impact of persistent infection may not be adequately captured by the relatively short-term investigations of treatment impact included in this meta-analysis.
In schistosomiasis-endemic regions, many children are infected by age two and remain chronically infected through school-age and late adolescence [9]. Under current national Schistosoma control treatment guidelines, preschool-age children are not treated as part of routine deworming programs for STH or schistosomiasis [1, 6, 30]. These children may therefore suffer cumulative damage to their health and function that is currently not reflected in most short-term study outcomes (or this meta-analysis). Of note, recent investigations have demonstrated the safety and efficacy of praziquantel for treatment of Schistosoma infection in preschool children [9]. The existence of an adverse developmental impact of Schistosoma infection on cognitive/educational domains would be a major justification for expanding the age-bracket of children who should be treated with praziquantel. Currently, evidence suggests that the timing of infection across a life path is especially consequential in terms of the severity of cognitive and physiologic impairments experienced [77, 78]. Future investigations evaluating the relative differences in cognitive outcomes for pre-school children with and without Schistosoma infection will be important for understanding the magnitude of potential impact of better prevention of Schistosoma infection.
It is currently unknown whether the cognitive and educational loss associated with Schistosoma infection can be reversed with treatment alone. The fact that infection often occurs in the context of malnutrition, coincident parasitic infections, and extreme poverty suggests that cognitive remediation efforts will need to be multi-faceted, using an integrated disease management framework. We expect that educational and cognitive interventions will be most effective if initiated earlier in life and that the package of interventions may need to include remedial instruction, the prevention of reinfection for treated/cured children, management of comorbid health conditions, and interventions for improvement of nutritional status.
Our investigation suggests that Schistosoma infection/non-treatment is associated with educational and cognitive loss. Our findings further suggest a definite cognitive and educational benefit of anti-schistosomal deworming among school-age children. Future complex intervention studies of early childhood interventions, focused on improving child well-being and cognitive potential, are needed to determine to what extent these observed deficits are preventable or reversible. Interventions that employ an integrated disease management framework will likely identify cost-efficiencies for leveraging existing disease and nutrition treatment programs in helminth-affected regions.
|
10.1371/journal.pgen.1008146 | Identification and validation of genetic variants predictive of gait in standardbred horses | Several horse breeds have been specifically selected for the ability to exhibit alternative patterns of locomotion, or gaits. A premature stop codon in the gene DMRT3 is permissive for “gaitedness” across breeds. However, this mutation is nearly fixed in both American Standardbred trotters and pacers, which perform a diagonal and lateral gait, respectively, during harness racing. This suggests that modifying alleles must influence the preferred gait at racing speeds in these populations. A genome-wide association analysis for the ability to pace was performed in 542 Standardbred horses (n = 176 pacers, n = 366 trotters) with genotype data imputed to ~74,000 single nucleotide polymorphisms (SNPs). Nineteen SNPs on nine chromosomes (ECA1, 2, 6, 9, 17, 19, 23, 25, 31) reached genome-wide significance (p < 1.44 x 10−6). Variant discovery in regions of interest was carried out via whole-genome sequencing. A set of 303 variants from 22 chromosomes with putative modifying effects on gait was genotyped in 659 Standardbreds (n = 231 pacers, n = 428 trotters) using a high-throughput assay. Random forest classification analysis resulted in an out-of-box error rate of 0.61%. A conditional inference tree algorithm containing seven SNPs predicted status as a pacer or trotter with 99.1% accuracy and subsequently performed with 99.4% accuracy in an independently sampled population of 166 Standardbreds (n = 83 pacers, n = 83 trotters). This highly accurate algorithm could be used by owners/trainers to identify Standardbred horses with the potential to race as pacers or as trotters, according to the genotype identified, prior to initiating training and would enable fine-tuning of breeding programs with designed matings. Additional work is needed to determine both the algorithm’s utility in other gaited breeds and whether any of the predictive SNPs play a physiologically functional role in the tendency to pace or tag true functional alleles.
| Certain horse breeds have been developed over generations specifically for the ability to perform alternative patterns of movement, or gaits. Current understanding of the genetic basis for these gaits is limited to one known mutation apparently necessary, but not sufficient, for explaining variability in “gaitedness.” The Standardbred breed includes two distinct groups, trotters, which exhibit a two-beat gait in which the opposite forelimb and hind limb move together, and pacers, which exhibit an alternative two-beat gait where the legs on the same side of the body move together. Our long-term objective is to identify variants underlying the ability of certain Standardbreds to pace. In this study, we were able to identify several regions of the genome highly associated with pacing and, within these regions, a number of specific highly associated variants. Although the biological function of these variants has yet to be determined, we developed a model based on seven variants that was > 99% accurate in predicting whether an individual was a pacer or a trotter in two independent populations. This predictive model can be used by horse owners to make breeding and training decisions related to this economically important trait, and by scientists interested in understanding the biology of coordinated gait development.
| Gait refers to a pattern of limb movement during locomotion, and can be defined by patterns of footfall and symmetry or asymmetry, among other factors. In quadrupeds, a limited number of gaits are conserved among species, including the walk (4-beat, symmetric), trot (two-beat, symmetrical, diagonal), and gallop (4-beat, asymmetric). Deviations from normal gait patterns are suggestive of underlying musculoskeletal or neurologic abnormalities. However, certain breeds of horses, including the Standardbred, Icelandic horse, Tennessee Walking Horse, and Paso Fino, have been specifically selected over generations of breeding for their ability to perform alternative patterns of locomotion. These alternative gaits are typically of intermediate speed and replace the trot. For example, the pace is a 2-beat lateral symmetrical gait, and the tolt is a 4-beat lateral symmetrical gait (S1 Fig).[1] There are strong signatures of selection evident when comparing gaited and non-gaited breeds[2], and the trait is highly heritable; for example, heritabilities of the pace and tölt in the Icelandic horse have been estimated to range between 0.53 and 0.73.[3] However, until recently, the specific genetic determinants underlying these alternative gaits were completely unknown.
In 2012, a genome-wide association study (GWAS) in four-gaited (walk, trot, tölt, and gallop) and five-gaited (walk, trot, tölt, gallop, and pace) Icelandic horses revealed a strongly associated SNP on equine (ECA) chromosome 23.[1] Deep (30x coverage) whole-genome sequencing of one four-gaited and one five-gaited individual revealed a premature stop codon in the last exon of DMRT3 (an isoform of the doublesex and mab-3 related transcription factor). Subsequent genotyping of additional Icelandic horses revealed that nearly all five-gaited individuals were homozygous for the mutation, compared to only a third of the four-gaited horses. Of even greater interest, when horses of other breeds were genotyped for the mutation, it was found to be nearly fixed in gaited breeds (e.g. Paso Fino, Peruvian Paso, Tennessee Walking Horse, Standardbred.), but absent in non-gaited breeds (e.g. Arabian, Thoroughbred).[1] The functional importance of DMRT3 was confirmed in a mouse model, where mice null for DMRT3 exhibited an abnormal gait characterized by an increased stride, prolonged, stance and swing phases of both the thoracic and pelvic limbs, and near absence of coordinated pelvic limb movements. Further, DMRT3 expression was localized to the spinal cord both pre- and postnatally, and null mice had fewer commissural interneurons, suggesting that this gene is important for the development of normal locomotor coordination.[1]
Although the DMRT3 mutation appears to be necessary for “gaitedness” in horses it is not sufficient to explain the variation of this trait, as demonstrated by the fact that it is nearly fixed in Standardbreds, although not all individuals exhibit that breed’s alternative gait, pacing.[1] It is noteworthy that approximately 20% of the offspring of Standardbred trotter stallions go on to race as pacers.[4] It is unknown whether this is due to genetic predisposition, training, or a combination of the two, but it is likely that modifying genetic factors segregate in the Standardbred population and determine an individual’s ability to pace. The purpose of this study was to identify putative modifying alleles associated with gait in a large cohort of Standardbred pacers and trotters using a combination of GWAS and variant discovery via whole-genome sequencing.
Horses included in the GWAS cohort (n = 542) were genotyped on either the first generation (Equine SNP50; n = 306) or second generation (Equine SNP70; n = 236) Illumina equine beadchip. After genotype imputation, 73,691 markers were available for analysis; after pruning, 62,901 SNPs were included in the mixed model association analysis. After correction for relatedness and population structure, mixed model association analysis in GEMMA[5] revealed 19 SNPs on nine chromosomes that reached genome-wide significance (p < 1.44 x 10−6, as determined by the likelihood ratio test) (Fig 1, Table 1). Seven SNPs were located on equine (ECA) chromosome 17, three on ECA1, two on ECA6, two on ECA31, and one each on ECA2, 9, 19, 23, and 25. The ECA17 SNPs were from two distinct regions, 28.4–41.9Mb (n = 4) and 60.50–60.55Mb (n = 3) (Table 1). Only eight of the genome-wide significant SNPs were found within protein coding genes; all were intronic (Table 1). An additional 37 SNPs on 14 chromosomes reached significance considered to be moderately associated with gait (p < 1 x 10−5) [6] (S1 Table).
We have previously reported the results of whole-genome sequencing in this cohort.[7] Briefly, 12 individuals (6 pacers, 6 trotters) were sequenced at an average coverage depth of 6.4x (range 4.7x-7.9x). Six individuals (3 pacers, 3 trotters) were sequenced at an average coverage depth of 12.2x (range 10x-13.1x). After filtering, 14,588,812 variants were called, of which 13,157,608 were SNPs, 671,144 were insertions, and 760,060 were deletions. Of these variants, 99.1% were predicted to have no functional effect, 0.5% (85,916) were predicted to have minor functional effect, 0.4% (57,122) were predicted to have moderate functional effect, and 0.07% (9,662) were predicted to have major functional effect.
Two pools of genomic DNA (20 pacers, 20 trotters) were sequenced at a target depth of 30x. A total of eighty-nine 50kb regions met one of the filtering criteria of either high differentiation between pacers and trotters (FST ≥ 0.35) or a combination of low pool heterozygosity (Hp < 0.1) in one of the groups and high differentiation (FST ≥ 0.30). Some of these regions were contiguous or overlapping, creating larger regions of interest ranging in size from 75-300kb (S2 Table). A total of 1,885 SNPs were called across all regions of interest. Of these, 1,273 were annotated by Ensembl (Equ Cab 2; GCA_000002305.1) as intergenic, 184 were located upstream of a gene, 138 were located downstream of a gene, 270 were located within an intron, and 20 were located within an exon (S3 Table).
Approximately 62,000 SNPs were evaluated from regions on 13 chromosomes identified as being of interest based on our GWA analysis (regions from 9 chromosomes that contained genome-wide significant SNPs, and regions from an additional 4 chromosomes that contained SNPs approaching genome-wide significance). An additional 1,885 SNPs were identified within regions of interest on 19 chromosomes from the pooled sequencing data. Three hundred three SNPs were included in the final Sequenom assay, including 190 SNPs from the whole-genome sequencing regions of interest (based on the GWA analysis) and 113 SNPs from the pooled sequencing regions of interest (S4 Table). Additionally, 98 ancestry informative markers (AIMs) were included on the assay to help control for population structure during downstream analysis (see Methods and [7]).
Genotyping was performed in 720 individuals (n = 458 trotters; n = 262 pacers). After pruning, 245 SNPs were available for mixed model association analysis in GEMMA. With Bonferroni correction for multiple testing, statistical significance was set at p < 2 x 10−4; after correcting for relatedness and population structure, 177 SNPs met this criteria for statistical significance (Table 2, S5 Table). Pacers were more likely than trotters to carry the derived alleles (compared to the reference, a Thoroughbred) (Table 2). Nearly all of the trotters carried only a single copy of the alternate allele in each case, while it was more common for pacers to be homozygous for the alternate allele.
Random forest classification analysis of genotyping data from the Sequenom assay in 659 Standardbreds with racing records (n = 428 trotters; n = 231 pacers) yielded an out-of-box (OOB) error rate of 0.61%, with a total of four misclassified individuals (three trotters misclassified as pacers, and one pacer misclassified as a trotter). Interestingly, one of the trotters who was predicted to be a pacer was in fact out of a line of pacing Standardbreds; it is not known whether this horse was trained as a pacer at any point before starting its racing career. There were 21 SNPs with a mean reduction of node impurity score (GINI index) > 5 (Table 3). When the random forest analysis was repeated, the relative importance of these SNPs varied only slightly over multiple iterations. The most important SNPs for classification as a pacer or trotter according to this analysis were located on ECA1, 17, 23, and 30. Ten-fold cross-validation of this data using linear discriminate analysis resulted in a misclassification error of 0.0106.
A conditional inference tree was constructed to determine the hierarchical organization of the most informative SNPs identified by random forest analysis. A tree composed of seven SNPs predicted status as a pacer or trotter among the 659 genotyped individuals with 99.1% accuracy, with only six horses misclassified (three pacers and three trotters). Again, one of these misclassified trotters came from a line of pacers. Considering pacing as the outcome of interest, this prediction model demonstrated a sensitivity of 98.7% (95% CI 96.25% - 99.73%) and a specificity of 99.3% (95% CI 97.97% - 99.86%). The seven SNPs were located on six chromosomes (ECA1, 6, 17, 23, 15, and 30) (Fig 2). For four SNPs, the alternate allele was more common in pacers, and in three SNPs the alternate allele was more common in trotters. In either case, the group with the lower allele frequency included very few homozygotes (Table 4).
One hundred sixty-six horses (n = 83 trotters, n = 83 pacers) were genotyped on the same custom Sequenom assay as the discovery cohort. The genotypes for the seven SNPs included in the conditional inference tree were extracted. Two individuals had missing genotypes at one or more of the SNPs and could not be classified. Of the remaining 164 horses, 163 were correctly classified as pacers or trotters, resulting in an overall accuracy of 99.4%. The single misclassified horse was a pacer.
The horse is unique among quadrupeds in that certain breeds have been strongly selected for the ability to exhibit alternative patterns of locomotion as a physiologic adaptation. Beyond giving insight into an economically important trait, improved understanding of the pathways that underlie alternative gaits in the horse may also provide insight into pathways that are dysregulated with disease in other species, as well as basic insight into the underlying neurobiology of locomotion. GWA analysis in our population identified 19 SNP markers that were associated with gait at a level of genome-wide significance. These SNPs defined regions of interest on nine chromosomes that contained more than two dozen named genes. However, a challenge arises in identifying biologically compelling candidate genes for gait because there is still much that is not known about the development of normal limb coordination. It is likely that many genes that play a role in the development of alternative gaits have not previously been associated with any aspect of neurobiology. This is aptly illustrated by DMRT3 which had initially been described as primarily playing a role in gonadal development and sexual differentiation.[8] The DMRT3 nonsense mutation originally reported by Andersson et al. in 2012 [1] has now been reported to occur in 68 out of 141 breeds tested from around the world, and at high frequency (>50%) in all “gaited” breeds.[9] This example demonstrates that a strongly associated mutation cannot be ruled out as having a functional role in the development of alternative gaits simply because it falls within a gene that does not have a described role in neural development or locomotion.
As SNPs are chosen for inclusion in genotyping panels based on their distribution and frequency, rather than on predicted effect, it is unlikely that any of the markers in the GWA analysis play a functional role in gait. Rather, it is more likely that they are “tagging” truly functional variants with which they are in linkage disequilibrium (LD).[10, 11] Standardbreds have been reported to have the greatest long-range LD (> 1,200kb) among horse breeds; thus, it is not unreasonable to expect that a significant SNP in a GWA might be “tagging” a functional sequence variant up to 1Mb distant (or further). Given this, we chose whole-genome sequencing as the most efficient way to catalogue variants within 1Mb of the regions defined by our genome-wide significant SNPs. This approach also allowed variant discovery in a larger cohort of individuals (9 trotters and 9 pacers) than would have been feasible using a traditional candidate gene approach, giving us a better picture of the alleles present in our population, as well as the segregation of these alleles with gait status. Pooled whole-genome sequencing offered a complementary approach to identifying regions and variants of interest based on population parameters of differentiation (FST) and heterozygosity (Hp). Pooled whole-genome sequencing has previously been used to identify genomic regions under selection, and identify candidate variants within those regions, in a number of plant and animal species.[12–14]
Of the tens of thousands of SNPs discovered within regions of interest via whole-genome sequencing, only a small fraction could be selected for follow-up. Thus, it is likely that we have not identified any truly functional alleles despite our prioritization process. Indeed, of the top 40 variants from GEMMA analysis of Sequenom genotyping in 720 individuals, only two resulted in amino acid changes. However, given the strength of association of our selected variants with gait in this large population, it is highly likely that one or more of these genotyped variants are “tagging” specific functional alleles, and additional investigation of nearby variants is warranted. It must be noted that given the strong population structure of our cohorts, inherent to the history of this breed which has undergone strong selection for our trait of interest over many generations, it is also possible that one or more of these variants are reflecting some other aspect of population structure unrelated to gait. Future work to address the issue of functionality will include tissue expression profiles of the cerebellum and specific regions of the proximal cervical spinal cord that contain neuronal tracts known to play a role in coordinated locomotion.
Random forest classification analysis was selected to help prioritize among the numerous statistically significant variants in our Sequenom assay. In a random forest approach to a binary trait, the predicted probability of an individual expressing or not expressing a trait (in this case, being able to pace) is based on the aggregation of a number of decision trees.[15, 16] Within these decision trees, each node is an attribute–in this case, the genotype at a given SNP. The importance of each SNP is determined by quantifying the increase of misclassified individuals when the genotype at that SNP is randomly permuted.[15] This approach requires no prior knowledge of gene function and can accommodate multiple variants within the same gene. Random forest analysis has previously been successfully used to identify SNPs associated with feed intake in dairy cattle[17] as well as pathway-phenotype associations in human bladder cancer.[16] In our population, random forest analysis revealed that just a few SNPs were of large importance in correctly classifying individuals, while a large number of SNPs were of minor to minimal importance. This suggested that an accurate prediction algorithm might be constructed, despite not knowing the functional importance of the individual SNPs. In fact, we were able to develop a prediction algorithm consisting of only seven SNPs that was > 99% accurate in two independently sampled cohorts of Standardbreds. The small number of misclassified individuals may have had an environmental component (i.e. training) that explains why they were competing at a different gait than expected, or it may simply reflect room for refinement of the algorithm by the identification of truly functional SNPs. This is the first time that a prediction algorithm for gait has been reported and it could be used by owners/breeders/trainers for both marker-assisted selection and making training decisions by identifying young horses that have the genetic background to race successfully at the pace. This model will need to be tested in other breeds to determine if its predictive value is specific to Standardbreds, to breeds that pace (e.g. Icelandic Horses), or if it is universally applicable across gaited breeds.
DNA was isolated from whole blood samples using the Gentra Puregene Blood Kit (Qiagen, Valencia, CA) per manufacturer recommendations. Briefly, RBC lysis solution was added to samples at a 3:1 ratio, incubated, and centrifuged. After discarding the supernatant, Cell lysis solution was added to the white blood cell pellet and the cells were re-suspended, after which protein was precipitated and discarded. DNA was precipitated in isopropanol and subsequently washed in ethanol prior to final hydration. Quantity and purity of extracted DNA were assessed using spectrophotometric readings at 260 and 280nm (NanoDrop 1000, Thermo Scientific, Wilmington, DE).
Genome-wide genotyping of single nucleotide polymorphism (SNP) markers was performed by Neogen GeneSeek (Lincoln, NE) using an Illumina Custom Infinum SNP genotyping platform. Horses were genotyped at either 54,602 SNPs using the first generation Illumina Equine SNP50 chip (n = 306), or at 65,157 SNPs using the second generation Illumina Equine SNP70 chip (n = 236).
The two genotyping platforms used in the GWAS cohort share 45,703 SNPs. As an alternative to losing information from tens of thousands of SNPs by pruning to this shared marker list prior to merging files, genotype imputation may be used. This technique statistically estimates genotypes from non-assayed SNPs based on a comparison of haplotype blocks between one population and a second, more densely genotyped reference population. An established pipeline for imputation of equine genotyping data [18] was used to impute the ~18,000 markers unique to the SNP70 chip in those horses genotyped on the SNP50 chip, and likewise to impute the ~9,000 markers unique to the SNP50 chip in those horses genotyped on the SNP70 chip. Imputed files were merged for subsequent analysis using the—merge command in PLINK.[19]
A GWA analysis with gait as the phenotype of interest was performed after genotype imputation using GEMMA (Genome-Wide Mixed Model Analysis) software.[5] A centered relatedness matrix (-gk 2) was constructed using a LD-pruned set of approximately 6600 markers (100 SNP windows, sliding by 25 SNPs along the genome, pruned at r2 > 0.2; PLINK command—indep-pairwise 100 25 0.2).[20] All three possible frequentists tests were performed: Wald, likelihood ratio, and score (-fa 4). A covariate file including sex and origin (North America or Europe) was incorporated into the mixed model (-c) and SNPs were pruned according to GEMMA default parameters (MAF <1%, missingness <95%). Association plots were generated using the base graphics package in the R statistical computing environment.[21] Genome-wide significance was set at p < 1.44 x 10−6 based on the effective number of independent tests in our data.[22]
We have previously reported whole-genome sequencing in this cohort.[7] For the purposes of the previous study, horses were selected for sequencing based on haplotypes in regions of interest associated with osteochondrosis; however, they were also selected in a balanced manner for their gait phenotype, with 9 pacers and 9 trotters included. Briefly, genomic DNA (2–6μg) from the 18 horses was submitted to the University of Minnesota Biomedical Genomics Center (UMGC) for quality control, library preparation, and sequencing. Samples were subjected to library preparation including fragmentation, polishing, and adaptor ligation, and were prepared with an indexed barcode for a 100bp paired-end run on the Illumina HiSeq sequencer, per standard protocols. Targeted depth of coverage was 12x for six horses and 6x for twelve horses, with each group balanced for gait.
Data analysis, including quality control, alignment, and variant detection, was carried out following published best practices[23, 24] within the Galaxy framework hosted by the Minnesota Supercomputing Institute. Briefly, reads that passed quality control were mapped to the reference sequence (EquCab 2.0, Sept. 2007 [25]) using BWA for Illumina[26]. Ambiguously mapped reads, low quality reads, and PCR duplicates were removed, after which reads were realigned around indels. Base quality recalibration was performed to remove systematic bias. This process was completed for the reads from each of the eight lanes for every individual before merging the mapped and recalibrated “lane-level” BAM files into a single “sample-level” file. Removal of duplicates and realignment around indels was repeated on the merged file. The eighteen sample-level files were merged into three groups of six, evenly divided between pacers and trotters, for the purposes of variant calling using the UnifiedGenotyper utility of the Broad Institute’s Genome Analysis ToolKit (GATK)[27] with a threshold phred-scale score of 20.0. Variants were filtered using the following thresholds: Quality Depth (QD) < 2.0 (assesses variant quality score taking into account depth of coverage at that variant), Read Position Rank Sum < -20.0 (Mann-Whitney Rank Sum test on the distance of the variant from the end of each read covering it), Fisher Strand (FS) > 200.0 (phred-scaled p-value to detect strand bias). Filtered variant lists from the three groups were combined into a single variant calling file (VCF) for subsequent analysis. Predicted functional effect for each called variant was determined based on the current equine reference genome annotation using SnpEff.[28] Frequency of variants within cases and controls, and the significance of frequency differences, was calculated using SnpSift CaseControl.[29] Variants from particular chromosomal regions of interest were selected using SnpSift Intervals and converted into Excel format for further evaluation.
Genomic DNA from 20 pacers and 20 trotters were combined into two pools, each comprising equimolar amounts of DNA from all individuals, for the purpose of whole-genome sequencing. These individuals were selected from the GWAS cohort and were chosen to be as unrelated as possible on the basis of coancestry coefficient generated from the whole-genome genotyping data (PLINK command—genome). Selected pacers had coancestry coefficients <0.06 (no more closely related than first cousins); selected trotters had coancestry coefficients <0.14 (one pair of half-siblings, the rest less closely related).
The two DNA pools were sequenced to 30X average depth of coverage using an Illumina HiSeq2500 sequencer at Uppsala University. The resulting paired reads were subjected to sequencing adaptor trimming and were subsequently aligned to the horse reference genome (EquCab2.1)[25] using the Burrows-Wheeler alignment algorithm as implemented in BWA for Illumina [26] (bwa sampe) using default alignment settings. Aligned reads were subjected to duplicate removal using the algorithm MarkDuplicates implemented in the Picard-tools software (https://broadinstitute.github.io/picard/). SNP and small insertion/deletion calling was performed using the UnifiedGenotyper algorithm of the Genome Analysis Toolkit (GATK) [27], and the resulting SNP calls were filtered using published best practice variant filtration settings.[23, 24] Numbers of sequence reads corresponding to the reference and variant alleles at filtered SNP sites were determined and were used to estimate allele frequencies for the pacer and trotter pools at each SNP locus. The allele counts and frequencies were then used to calculate the fixation index (FST) for the contrast between the two pools and to calculate estimated pool heterozygosity (Hp) within each pool for 50% overlapping sliding windows of 50 kilobases along the genome as previously described.[12] Distributions of the genome-wide FST and Hp values were consulted to determine the genomic intervals displaying the most strongly differentiated loci between the pools and the most strongly fixed loci within each pool, respectively. From these distributions we used a strategy where windows fulfilling one of two criteria, (1) FST ≥ 0.35 or (2) FST ≥ 0.30 with the additional criteria of at least one pool showing Hp <0.1, were selected as regions of interest.
A custom Sequenom genotyping assay was designed for high-throughput evaluation of prioritized variants. Variants were selected from top regions of interest identified in the GWA, as well as from regions from the pooled sequencing data with high differentiation between pacers and trotters (FST ≥ 0.35) or a combination of low pool heterozygosity (Hp < 0.1) in one of the groups and high differentiation (FST ≥ 0.30). SNPs discovered via whole-genome sequencing that passed quality control filters were prioritized according to the following parameters: 1) segregation with gait (preferentially with the alternate allele found in all or nearly all pacers and less than half of the trotters); 2) not intergenic; 3) non-synonymous, then synonymous changes; 4) if intronic, close to the exon-intron border (preferably < 100bp); 5) coding genes preferred over non-coding; and 6) if upstream/downstream, as close as possible to start/stop codon. Variants from pooled sequencing data were prioritized according to criteria 2–6.
When possible, at least one variant was selected from each coding gene within each region of interest. Among adjacent variants with equal magnitude of predicted functional effect, the one with the higher genomic p-value was selected for inclusion. Ancestry informative markers (AIMs) were also included in the assay to help control for population structure.[30] These 98 markers have previously been reported [7]; they describe more than 97% of the genetic variation captured by principal components from genome-wide genotyping data in the Standardbred breed.
|
10.1371/journal.pgen.1004799 | Stratification by Smoking Status Reveals an Association of CHRNA5-A3-B4 Genotype with Body Mass Index in Never Smokers | We previously used a single nucleotide polymorphism (SNP) in the CHRNA5-A3-B4 gene cluster associated with heaviness of smoking within smokers to confirm the causal effect of smoking in reducing body mass index (BMI) in a Mendelian randomisation analysis. While seeking to extend these findings in a larger sample we found that this SNP is associated with 0.74% lower body mass index (BMI) per minor allele in current smokers (95% CI -0.97 to -0.51, P = 2.00×10−10), but also unexpectedly found that it was associated with 0.35% higher BMI in never smokers (95% CI +0.18 to +0.52, P = 6.38×10−5). An interaction test confirmed that these estimates differed from each other (P = 4.95×10−13). This difference in effects suggests the variant influences BMI both via pathways unrelated to smoking, and via the weight-reducing effects of smoking. It would therefore be essentially undetectable in an unstratified genome-wide association study of BMI, given the opposite association with BMI in never and current smokers. This demonstrates that novel associations may be obscured by hidden population sub-structure. Stratification on well-characterized environmental factors known to impact on health outcomes may therefore reveal novel genetic associations.
| We found that a single nucleotide polymorphism in the CHRNA5-A3-B4 gene cluster, which is known to influence smoking heaviness, is associated with lower body mass index (BMI) in current smokers, but higher BMI in never smokers. This difference in effects suggests that the variant influences BMI both via pathways other than smoking, and via the weight-reducing effects of smoking, in opposite directions. The overall effect on BMI would therefore be undetectable in an unstratified genome-wide association study, indicating that novel associations may be obscured by hidden population sub-structure.
| As obesity represents a substantial and growing threat to public health, efforts to identify the determinants of obesity are of considerable scientific and societal importance. Genome-wide association studies (GWAS) have identified numerous variants associated with body mass index (BMI) [1], but a substantial proportion of the estimated heritability remains to be accounted for. At the same time, a number of modifiable environmental factors have been identified that influence BMI, with cigarette smoking a strong lifestyle influence on BMI [2]. In a previous Mendelian randomisation analysis, we used a single nucleotide polymorphism in the CHRNA5-A3–B4 gene cluster associated with heaviness of smoking within smokers [3] to confirm the causal effect of smoking in reducing BMI [4].
We sought to extend these findings in a larger sample drawn from the Causal Analysis Research in Tobacco and Alcohol (CARTA) consortium (http://www.bris.ac.uk/expsych/research/brain/targ/research/collaborations/carta/). We used the same genetic variant, characterised by two SNPs (rs16969968 and rs1051730) which are in perfect linkage disequilibrium (LD) in samples of European ancestry, and therefore reflect the same genetic signal (hereafter rs16969968-rs1051730). This variant is associated with approximately 1% phenotypic variance in cigarettes per day and approximately 4% variance in cotinine levels (the primary metabolite of nicotine, and a more precise measure of exposure) [5], [6]. Mendelian randomisation analyses of the causal effects of smoking heaviness require stratification according to smoking status – any causal effects of the exposure (i.e., smoking heaviness) should be reflected in an association of the instrument (i.e., genotype) among current smokers only, and not never smokers (former smokers might be expected to be intermediate between current and never smokers) [7]. The never smoking group therefore enables a test of the specificity of the instrument (i.e., that the variant only affects the outcome through the exposure of interest) [8]. Critically, the rs16969968-rs1051730 variant has not been shown to be associated with smoking initiation (i.e., it does not influence risk of being an ever versus a never smoker) in previous GWAS of smoking behaviour [9], which reduces the risk of introducing collider bias when stratifying on smoking status.
In the course of these analyses, we observed an unexpected finding, which we report here. Specifically, we observed an association of rs16969968-rs1051730 with higher BMI in never smokers. This association has not previously been reported in GWAS of BMI published to date. We therefore focus on the implications of this novel finding, and not the Mendelian randomisation analysis of the causal effects of smoking on BMI.
Our total sample size comprised 148,730 never smokers, former smokers and current smokers. In the 66,809 never smokers, we observed positive association of rs16969968-rs1051730 with BMI (Table 1), indicating an association operating via pathways other than smoking (percentage change per minor allele +0.35, 95% CI +0.18 to +0.52, P = 6.38×10−5). We also confirmed the expected inverse association of rs16969968-rs1051730 with BMI in the 38,913 current smokers (percentage change −0.74, 95% CI −0.97 to −0.51, P = 2.00×10−10), consistent with a causal, weight-reducing effect of cigarette smoking on BMI. There was no evidence of association in the 43,009 former smokers (percentage change −0.14, 95% CI −0.34 to +0.07, P = 0.19). An interaction test indicated that these estimates differed from each other (P = 4.95×10−13). Similar associations were observed for weight (Table 1) and waist circumference (data available on request), but not height (Ps ≥0.27 for all smoking categories). Between-study heterogeneity was low (I2 values ≤36%), and there was no evidence for effect modification by sex. Critically, when data were examined without stratification by smoking status no clear evidence of association with BMI was observed (P = 0.22), indicating that a conventional GWAS would have failed to detect this signal.
The 0.35% per minor allele BMI increase in never smokers represents a change of approximately 0.09 kg/m2. This is smaller than the effect of rs9939609 in FTO (∼0.4 kg/m2) [10] but is comparable in terms of variance explained to the other variants identified by Speliotes and colleagues [1]. As noted above, the rs16969968-rs1051730 variant has not been shown to be associated with smoking initiation in previous GWAS of smoking behaviour [9]. This is also true in our data (ever smoker versus never smoker: OR per minor allele 1.01, 95% CI 0.99 to 1.03, P = 0.50), although we observed an association with smoking cessation (current smoker versus former smoker: OR per minor allele 1.08, 95% CI 1.06 to 1.10, P = 1.44×10−12), consistent with previous studies [11]. Therefore, we do not believe that these findings are due to collider bias, whereby stratifying on the exposure measure can induce associations between instrument and outcome [12].
Our results indicate that rs16969968-rs1051730 may be associated with BMI in never smokers, via pathways other than smoking, as well as with heaviness of smoking among current smokers. At this stage we can only speculate as to the mechanism through which rs16969968-rs1051730 may exert a positive effect on BMI in never smokers. In GWAS, the CHRNA5-A3-B4 gene cluster was confirmed to be associated with heaviness of smoking, and downstream health outcomes including lung cancer and peripheral arterial disease [9], [13], [14]. It has been shown that the rs16969968 variant is functional and leads to an amino acid change (D398N) in the α5 nicotinic acetylcholine receptor (nAChR) subunit protein [15]. Animal models indicate that this subunit modulates tolerance to high doses of nicotine [16]. Candidate gene studies have suggested an association of rs16969968-rs1051730 with other substance use phenotypes, such as cocaine use [17], while other variants in this region have been reported to be associated with alcohol consumption [18], although the evidence for these associations is currently weak. Therefore, one possibility is that nAChRs play a role in central mechanisms mediating responding to rewarding stimuli in general, which could include natural rewards such as food.
It is also notable that rs3743075, located within the CHRNA3 gene and correlated with rs16969968-rs1051730 (r2 = 0.34, D′ = 1.00), shows association (N = 974, P = 9.06×10−5) with BMI (defined as <30 kg/m2 vs ≥30 kg/m2) (dbGaP Study Accession: pha003015.1). There is evidence from animal models that activation of hypothalamic α3β4 nAChRs leads to activation of pro-opiomelanocortin neurons, and subsequent activation of melanocortin 4 receptors, which have been shown to be critical for nicotine-induced decreases in food intake [19]. Therefore, another possibility is that nAChR sub-units play a role specifically in mediating food intake, through as yet undescribed mechanisms. In other words, the effects we have observed operate via other nAChRs, and other genes in this region (namely CHRNA3 and CHRNB4) may contribute to our finding. Clearly further work is required to explore this possibility. The use of more detailed body composition measures such as percent body fat and its distribution may also serve to refine the nature of the association.
Our results, if confirmed, have important implications for the design of future GWAS. The association we observed in never smokers would essentially be undetectable in an unstratified sample, since the effect size observed in the combined sample would require approximately 791,000 participants to detect even at an uncorrected P-value of 0.05, and even then would indicate an inaccurate effect size. This is essentially because the effect of rs16969968-rs1051730 on BMI that operates via pathways other than smoking is countered by the weight-reducing effect of smoking. Therefore, since there are roughly twice as many never smokers as current smokers on average across our sample, these two effects negate each other. On the other hand, a sample of approximately 160,000 never smokers would be required to detect the effect we observed with genome-wide significance. Assuming the proportions of never, former and current smokers in our sample, this would imply a total sample size of around 350,000. While this is larger than published GWAS of BMI [1], it is achievable. Therefore, although we cannot say how frequent a scenario such as the one we observed here will be, additional variants may be identified in GWAS stratified by environmental exposures known to have pronounced effects on the phenotype of interest, such as cigarette smoking or physical activity on BMI.
The pleiotropic effect of rs16969968-rs1051730 (or LD of this variant with another variant causally influencing BMI), if shown to be robust via replication, has important implications for Mendelian randomisation studies assessing the causal effects of smoking. In this case, we can be reasonably confident that the BMI-reducing effect of the variant operates through smoking because the association with BMI in current smokers is in the opposite direction to the association in never smokers. Furthermore, if the effects on BMI that operate via pathways other than smoking and the effects that operate via the weight-reducing effects of smoking are independent, then the true causal estimate of the magnitude of effect of smoking in reducing BMI is likely to be larger than estimated with this variant. However, some caution must be exercised in conducting and interpreting the results of other Mendelian randomisation analyses using this variant because rs16969968-rs1051730 may influence outcomes through its effects on BMI, instead of or in addition to smoking heaviness. One possible solution is to use genetic variants for BMI as a method of reciprocal randomization to determine the direction of causation within inter-correlated networks of mechanistic pathways (i.e., network Mendelian randomisation) [20].
A limitation to our analysis is that we were only able to control for potential population stratification indirectly in most samples, by restricting analyses to participants of self-reported European ancestry. We were not able to use other methods, such as adjustment for principal components, given that not all contributing studies hold the necessary genetic data. However, we note that the minor allele frequency of the rs16969968-rs1051730 differed only slightly across studies (between 0.30 and 0.36).
Testing for gene-environment interaction in GWAS is not novel [21], and examples exist which incorporate smoking status as an environmental factor [22]. However, this remains relatively uncommon, due to methodological challenges (e.g., introducing collider bias) and sample size constraints. A key challenge is the identification of suitable environmental variables on which to stratify GWAS analyses, from the multitude available. We suggest that focusing on environmental factors that are most strongly associated with the phenotype of interest, are likely to have profound biological effects, and which can be characterised in a relatively consistent way across studies, is likely to be the best strategy. Smoking status meets all of these criteria, and the data presented here demonstrate how stratification on well-characterized environmental factors known to impact on health outcomes (such as smoking status) may reveal novel genetic associations with health outcomes. As our data indicate, these associations may operate through genetic influences on the environmental factors themselves, or through new pathways which are masked by the environmental factors.
We used data on individuals (≥16 years) of European ancestry (ascertained via self report, or based on the genome-wide genotype data where available) from 29 studies in the Causal Analysis Research in Tobacco and Alcohol (CARTA) consortium (http://www.bris.ac.uk/expsych/research/brain/targ/research/collaborations/carta/): the 1958 Birth Cohort (1958 BC), the Avon Longitudinal Study of Parents and Children (ALSPAC, including both mothers and children), the British Regional Heart Study (BRHS), the British Women's Heart and Health Study (BWHHS), the Caerphilly Prospective Study (CaPS), the Christchurch Health and Development Study (CHDS), the Cohorte Lausannoise (CoLaus) study, the Exeter Family Study of Child Health (EFSOCH), the English Longitudinal Study of Ageing (ELSA), FINRISK, the Danish GEMINAKAR twin study, Generation Scotland, the Genomics of Overweight Young Adults (GOYA) females, GOYA males, the Helsinki Birth Cohort Study (HBCS), Health2006, Health2008, the Nord-Trøndelag health study (HUNT), Inter99, the Northern Finland Birth Cohorts (NFBC 1966 and NFBC 1986), MIDSPAN, the Danish MONICA study, the National Health and Nutrition Examination Survey (NHANES), the MRC National Survey of Health & Development (NSHD), the Netherlands Twin Registry (NTR), the Prospective Study of Pravastatin in the Elderly at Risk (PROSPER) and Whitehall II. References to these individual studies are available on request. All studies received ethics approval from local research ethics committees (see Text S1 for full details).
Within each study, individuals were genotyped for one of two single nucleotide polymorphisms (SNPs) in the CHRNA5-A3-B4 nicotinic receptor subunit gene cluster, rs16969968 or rs1051730. These single nucleotide polymorphisms are in perfect linkage disequilibrium with each other in Europeans (R2 = 1.00 in HapMap 3, http://hapmap.ncbi.nlm.nih.gov/) and therefore represent the same genetic signal. Where studies had data available for both SNPs, we used the SNP that was genotyped in the largest number of individuals.
Height (m), weight (kg) and waist circumference (cm) were assessed within each study, directly measured for 99% of participants, and self-reported for GOYA females (N = 1,015) and a sub-set of NTR (N = 602). Body mass index (BMI) was calculated as weight/height2.
Smoking status was self-reported (either by questionnaire or interview). Individuals were classified as current, former, or never cigarette smokers. Where information on smoking frequency was available, current smokers were restricted to individuals who smoked regularly (typically at least one cigarette per day). Where information on pipe and cigar smoking was available, individuals reporting being current or former smokers of pipes or cigars but not cigarettes were excluded from all analyses. For studies with adolescent populations (ALSPAC children and NFBC 1986), analyses were restricted to current daily smokers who reported smoking at least one cigarette per day (current smokers) and individuals who had never tried smoking (never smokers). Descriptive characteristics of smoking frequency data are provided in Text S2.
Analyses were conducted within each contributing study using Stata and R software, following the same analysis plan. Analyses were restricted to individuals with full data on smoking status and rs16969968-rs1051730 genotype. Within each study, genotype frequencies were tested for deviation from Hardy Weinberg Equilibrium (HWE) using a chi-squared test. Mendelian randomisation analyses of the association between rs16969968-rs1051730 and BMI were performed using linear regression, stratified by smoking status (never, former and current) and sex, and adjusted for age. BMI was log transformed prior to analysis. An additive genetic model was assumed on log values, so that each effect size could be exponentiated to represent the percentage increase in BMI per minor (risk) allele.
For NHANES, which has a survey design, Taylor series linearization was implemented to estimate variances. For studies including related family members appropriate methods were used to adjust standard errors: in GEMINAKAR, twin pair identity was included as a cluster variable in the model, in MIDSPAN linear mixed effects regression models fitted using restricted maximum likelihood were used to account for related individuals. ALSPAC mothers and children were analysed as separate samples; as there are related individuals across these samples, sensitivity analyses were performed excluding each of these studies in turn.
Results from individual studies were meta-analysed in Stata (version 13) using the “metan” command. As I2 values were all equal to or below 36% (indicating low to moderate heterogeneity), fixed effects meta-analyses were performed. The “metareg” command was used to examine whether SNP effects varied by sex and estimates were combined as there was no evidence for effect modification by sex. Evidence for interaction between genotype and smoking status was assessed using the Cochran Q statistic. Data are available from the Institutional Data Access/Ethics Committees of the individual studies that contributed to this analysis, for researchers who meet the criteria for access to confidential data. Full details are provided in Text S3.
Sample size calculations were performed using Quanto software (http://biostats.usc.edu/Quanto.html). The following parameters were used: 80% power to detect associations, minor allele frequency of 0.33, mean and standard deviation for BMI of 25 kg/m2 and 3.8 kg/m2 respectively, alpha values of 0.05 and 5×10−8.
|
10.1371/journal.pntd.0004263 | Coevolution of the Ile1,016 and Cys1,534 Mutations in the Voltage Gated Sodium Channel Gene of Aedes aegypti in Mexico | Worldwide the mosquito Aedes aegypti (L.) is the principal urban vector of dengue viruses. Currently 2.5 billion people are at risk for infection and reduction of Ae. aegypti populations is the most effective means to reduce the risk of transmission. Pyrethroids are used extensively for adult mosquito control, especially during dengue outbreaks. Pyrethroids promote activation and prolong the activation of the voltage gated sodium channel protein (VGSC) by interacting with two distinct pyrethroid receptor sites [1], formed by the interfaces of the transmembrane helix subunit 6 (S6) of domains II and III. Mutations of S6 in domains II and III synergize so that double mutants have higher pyrethroid resistance than mutants in either domain alone. Computer models predict an allosteric interaction between mutations in the two domains. In Ae. aegypti, a Ile1,016 mutation in the S6 of domain II was discovered in 2006 and found to be associated with pyrethroid resistance in field populations in Mexico. In 2010 a second mutation, Cys1,534 in the S6 of domain III was discovered and also found to be associated with pyrethroid resistance and correlated with the frequency of Ile1,016.
A linkage disequilibrium analysis was performed on Ile1,016 and Cys1,534 in Ae. aegypti collected in Mexico from 2000–2012 to test for statistical associations between S6 in domains II and III in natural populations. We estimated the frequency of the four dilocus haplotypes in 1,016 and 1,534: Val1,016/Phe1,534 (susceptible), Val1,016/Cys1,534, Ile1,016/Phe1,534, and Ile1,016/Cys1,534 (resistant). The susceptible Val1,016/Phe1,534 haplotype went from near fixation to extinction and the resistant Ile1,016/Cys1,534 haplotype increased in all collections from a frequency close to zero to frequencies ranging from 0.5–0.9. The Val1,016/Cys1,534 haplotype increased in all collections until 2008 after which it began to decline as Ile1,016/Cys1,534 increased. However, the Ile1,016/Phe1,534 haplotype was rarely detected; it reached a frequency of only 0.09 in one collection and subsequently declined.
Pyrethroid resistance in the vgsc gene requires the sequential evolution of two mutations. The Ile1,016/Phe1,534 haplotype appears to have low fitness suggesting that Ile1,016 was unlikely to have evolved independently. Instead the Cys1,534 mutation evolved first but conferred only a low level of resistance. Ile1,016 in S6 of domain II then arose from the Val1,016/Cys1,534 haplotype and was rapidly selected because double mutants confer higher pyrethroid resistance. This pattern suggests that knowledge of the frequencies of mutations in both S6 in domains II and III are important to predict the potential of a population to evolve kdr. Susceptible populations with high Val1,016/Cys1,534 frequencies are at high risk for kdr evolution, whereas susceptible populations without either mutation are less likely to evolve high levels of kdr, at least over a 10 year period.
| Constant use of pyrethroid insecticides has driven mosquito populations to develop resistance. In Aedes aegypti, the primary mosquito vector of dengue, yellow Fever, and chikungunya viruses, pyrethroid resistance is primarily associated with mutations in the voltage-gated sodium channel protein. One mutation occurs in codon 1,016 and involves a replacement of valine with isoleucine (Ile1, 016), and a second located in subunit 6 of domain III in codon 1,534, replaces phenylalanine with cysteine (Cys1,534). In Mexico, we found that Cys1,534 was present in the same mosquito collections that were previously analyzed for Ile1,016. In this study, we performed a linkage disequilibrium analysis on both Ile1,016 and Cys1,534 in Mexican collections from 2000–2012. Our analysis suggests that pyrethroid resistance requires the sequential evolution of the two mutations and that Cys1,534 must occur first and appears to enable the Ile1,016 mutation to survive.
| Worldwide Aedes aegypti (L.) mosquitoes are the principal urban vectors of dengue, chikungunya, and yellow fever viruses. Approximately 2.5 billion people (40% of the human population) currently live with the risk of dengue transmission. In Mexico, Ae. aegypti is the primary vector of the four dengue virus serotypes (DENV1-4), the causative agents of dengue fever (DF), dengue hemorrhagic fever (DHF) and dengue shock syndrome (DSS). Mexico is severely affected by DF, DSS, and DHF because all four dengue serotypes co-occur in most states of Mexico. A recent review of dengue disease in Mexico [2] reported an increase in incidences from 1.72 per 100,000 in 2000 to 14.12 per 100,000 in 2011.
Currently the most effective means to reduce dengue transmission by Ae. aegypti is through reduction of larval and adult populations. In Mexico larval reduction is accomplished chiefly through the application of the organophosphate temephos to peridomestic larval breeding sites and through physical source reduction or alteration of potential water-holding containers. Following recommendations of the official Mexican policy for vector control, (NOM-032-SSA2-2002), pyrethroids were almost exclusively used to control adults in and around homes from 1999 to 2010.
Pyrethroid insecticides prolong the opening of the voltage gated sodium channel protein (VGSC) in insect nerves to produce instant paralysis and ‘‘knock-down.” The α-subunit of VGSC has four repeat domains, labeled I-IV, each of which contains six transmembrane helix segments, S1-S6. Pyrethroids preferentially bind to the open state of vgsc by interacting with two distinct receptor sites formed by the interfaces of the transmembrane helix S6 of domains II and III, respectively [1]. The original computer modeling studies [3] suggest that simultaneous binding of pyrethroids to S6 in both domains II and III is necessary to efficiently lock sodium channels in the open state. These models also predict that mutations in the S6 of domain III allosterically alter S6 in domain II via a small shift of IIS6 thus establishing a molecular basis for the coevolution of S6 mutations in domains II and III in conditioning pyrethroid resistance.
In 2006 we described a mutation, Ile1,016, in the S6 of domain II in Ae. aegypti that is associated with very high knock-down resistance (kdr) to the pyrethroid insecticide permethrin in mosquitoes homozygous for this mutation. We examined collections of Ae. aegypti from Mexico during 1996–2009 [4] and found that the overall Ile1,016 frequency increased from 0.1% in 1996–2000, to 2%–5% in 2003–2006, to 38.3%–88.3% in 2007–2009 depending upon collection location. In 2010 another vgsc mutation was described in the S6 of domain III in Ae. aegypti that was also strongly correlated with kdr and involved a cysteine replacement (Cys1,534Phe) [5–7]. A general trend in these studies was that Cys1,534 frequencies were generally higher and increased more rapidly than Ile1,016 frequencies in natural populations.
Based upon these observations and on the dual binding model [3], we analyzed fresly collected DNA from Ae. aegypti for Ile1,016 and Cys1,534 while DNA previously analyzed for Ile1,016 [4] were tested for the presence of Cys1,534. The purpose of this study was to test the hypothesis that mutations in the S6 of domains II and III coevolve in a dependent manner through various allosteric interactions as suggested by computer models [3, 8]. An analysis of linkage disequilibrium was performed on the two alleles in 1,016 (Val 1,016 (susceptible), Ile 1,016(resistant)) and on the two alleles in 1,534 (Phe 1,534 (susceptible), Cys1,534 (resistant)) to assess whether alleles at 1,534 and 1,016 evolve independently or in a correlated fashion through epistasis.
Larval mosquitoes were collected from the locations mapped in Fig 1 and listed in Table 1. At each collection site, we collected immatures from at least 30 different containers in each of three different areas located at least 100 m apart. This included water storage containers and discarded trash containers such as plastic pails, tires, and cans. Larvae were returned to the laboratory where they were reared to adults and then identified to species. The Viva Caucel collection was west of the city of Merida in Yucatán State (20.9979639°, 089.7174611°). The Vergel collection was from eastern Merida (Fig 1) (20.9575694°, -89.5886889°). Both were collected in 2011 by Universidad Autónoma de Yucatán. DNA was isolated from individual adult mosquitoes by the salt extraction method [9] and suspended in 150 mL of TE buffer (10 mM Tris-HCl, 1 mM EDTA, pH 8.0). The SNP identification, allele-specific polymerase chain reaction (PCR), melting curve conditions, and genotype readings followed published procedures [6, 10–12].
The F3 generation of the Viva Caucel and Vergel strains were exposed to 25 μg permethrin (Chem Service, West Chester, PA) coated 250 mL Wheaton bottles. In each bottle approximately fifty 3–4 days old mosquitoes were exposed for one hour. Active mosquitoes were transferred to cardboard cups and frozen at -80°C and formed the ‘alive’ group. Knocked down mosquitoes were transferred to a second cardboard cup and placed into an incubator at 28°C and 70% humidity. After four hours, newly recovered mosquitoes were aspirated, frozen and labeled as ‘recovered’. The mosquitoes that remained inactive were scored as ‘dead’.
There are four potential 1,016/1,534 dilocus haplotypes: Val1,016/Phe1,534 (VF), Val1,016/Cys1,534 (VC), Ile1,016/Phe1,534 (IF), Ile1,016/Cys1,534 (IC). The number of times (Tij) that an allele at locus i = 1,016 appears with an allele at locus j = 1,534 was determined by the program LINKDIS [13]. The program then calculated composite disequilibrium frequencies [14] because the phase of alleles at 1,016 and 1,534 are unknown in double heterozygotes. An unbiased estimate of the composite disequilibrium coefficient Δij [14, 15] was calculated as:
Δij=(N/(N-1))((Tij/N)−2pipj)
Where N is the sample size and pi and pj are the frequencies of alleles at locus i = 1,016 and locus j = 1,534 respectively. Bayesian 95% Highest Density Intervals (HDI) around pi and pj were calculated in WinBUGS[16].
A χ2 test was performed to determine if significant disequilibrium exists among all alleles at 1,016 and 1,534. The statistic was calculated and summed over all two-allele-interactions [15]:
χ[1d.f.]2=N∑i∑j(Δij2pipj)
The linkage disequilibrium correlation coefficient Rij [15] is distributed from -1 (both mutations trans) to 0 (1,534 and 1,016 mutations occur independently), to 1 (both mutations cis) and therefore provides a standardized measure of disequilibrium:
Rij=Δij/(pi(1−pi)+Ci)(pj(1−pj)+Cj)
Where the Ci term corrects for departures from Hardy-Weinberg expectations:
Ci=Hobs(i)−pi2
where Hobs (i) is the observed frequency of i homozygotes. Departures from Hardy-Weinberg expectations were also expressed as Wright’s inbreeding coefficient (FIS) and calculated as 1- (Hexp/2p (1- p)) where Hexp is the observed frequency of heterozygotes. A χ2 test of the hypothesis FIS = 0 with one degree of freedom is:
χ[1d.f.]2=N(Hexp−Hobs)2∑ipi2+(∑ipi2)2−2∑ipi3
The locations of all sampling sites are shown in Fig 1 and the latitude and longitude coordinates are provided in Table 1. The sample sizes and numbers of the nine dilocus genotypes (Three 1,534 genotypes x Three 1,016 genotypes) are listed in Table 1. From a total of 615 treated mosquitoes in Viva Caucel, 17.6% (n = 108) were scored as alive, 15.6% (n = 96) as recovered and 66.8% (n = 411) as dead (Table 2). Genotypes at 1,016 and 1,534 were identified in 95 randomly chosen individuals from each group. From a total of 337 treated Vergel mosquitoes, 48.1% (n = 162) were scored as alive, 20.5% (n = 68) as recovered and 31.5% (n = 106) as dead. We randomly chose 95, 68 and 95 Vergel individuals from each group, respectively to obtain the genotypes at 1,016 and 1,534 (Table 2).
In Viva Caucel, the frequency of the Ile1,106 allele was 0.746 and the frequency of the Cys1,534 allele was 0.926 (Table 3), while in Vergel Ile1,016 was at a slightly higher frequency of 0.80 while the Cys1,534 allele was close to fixation at 0.988. The Ile1,106 and Cys1,534 alleles were in positive disequilibrium in Viva Caucel, but were only marginally significant in Vergel.
Genotypes at the 1,016 and 1,534 loci were not independent, in agreement with the linkage disequilibrium analysis in Table 4. Table 5 is a three-way contingency analysis of genotypes at loci 1,016, 1,534 and numbers alive or dead individuals in Viva Caucel. Numbers of alive were not independent of genotypes at the 1,016 locus; specifically, numbers of alive were significantly greater in Ile1,016 homozygous mosquitoes than in heterozygotes or in Val1,016 homozygotes. Numbers of alive were also not independent of genotypes at the 1,534 locus; specifically, numbers of alive were significantly greater in Cys1,534 homozygous mosquitoes than in heterozygotes or in Phe1,534 homozygotes. In general, numbers of alive in the Viva Caucel strain were not independent of genotypes at either locus.
However, a problem with this analysis is that genotypes at the two loci are not independent. In this and previous studies [10, 11], Ile1,016 homozygous mosquitoes have the greatest survival, while few, if any heterozygotes or Val1,016 homozygotes survive. To evaluate Cys1,534 genotypes independently of Ile1,016 homozygous mosquitoes, we only compared the three Cys1,534 genotypes among Ile1,016 heterozygotes and Val1,016 homozygotes. A significantly larger proportion of Cys1,534 homozygotes survived.
Table 5 also shows the contingency analyses of Vergel mosquitoes. Genotypes at the 1,016 and 1,534 loci were not independent, while they were marginally significant in the linkage disequilibrium analysis in Table 4. Numbers of alive were not independent of genotypes at the 1,016 locus again because numbers of alive were significantly greater in Ile1,016 homozygous mosquitoes than in heterozygotes or in Val1,016 homozygotes. Numbers of alive were however independent of genotypes at the 1,534 locus; specifically because Cys1,534 was almost fixed in the Vergel strain.
Table 5 also shows the three-way contingency analysis between genotypes at loci 1,016 and 1,534 and the numbers recovered or dead in Viva Caucel. As in Table 4, genotypes at the 1,016 and 1,534 loci were not independent. The numbers of recovered mosquitoes were not independent of genotypes at the 1,016 locus; specifically-numbers recovered were significantly greater in Ile1,016 homozygous mosquitoes than in heterozygotes or in Val1,016 homozygotes. Numbers of recovered were also not independent of genotypes at the 1,534 locus; specifically, numbers of alive were significantly greater in Cys1,534 homozygous mosquitoes than in heterozygotes or in Phe1,534 homozygotes.
In general, numbers of recovered in the Viva Caucel strain were heavily dependent on genotypes at both loci. An interesting difference between the two loci is that 32% (28/88) of Ile1,016 heterozygotes recovered while only 3.6% (1/28) of Cys1,534 heterozygotes recovered. This difference was significant (χ2 = 7.59, df = 1, p-value = 0.006).
Table 5 also shows the same analysis of recovery but in Vergel mosquitoes. Genotypes at the 1,016 and 1,534 loci were not independent, while they were marginally significant in the linkage disequilibrium analysis in Table 4. Numbers of recovered were not independent of genotypes at the 1,016 locus, again because numbers of recovered were significantly greater in Ile1,016 homozygous mosquitoes than in heterozygotes or in Val1,016 homozygotes. However, numbers of recovered were independent of genotypes at the 1,534 locus; specifically because Cys1,534 was approaching fixation in the Vergel strain.
Table 6 contains the frequencies of Ile1,016 and Cys1,534 and their Bayesian 95% HDI. FIS was significantly greater than zero (heterozygote deficiency) in two of the 36 collections where Ile1,016 and Val1,016 alleles were segregating. In contrast, a significant heterozygote deficiency occurred in eight of the 53 collections where Cys1,534 and Phe1,534 were segregating and an heterozygote excess occurred in two collections.
The frequencies of the Ile1,016 and Cys1,534 alleles from 1999 to 2012 are plotted in Fig 2. The Cys1,534 allele appears sooner and increases more rapidly than Ile1,016. Only the states of Veracruz and Chiapas had sufficient samples over the years to compare the spatial distributions of Ile1,016 and Cys1,534 (Fig 3). It is very clear that Ile1,016 and Cys1,534 were increasing in frequency much earlier in Veracruz state in eastern Mexico than in Chiapas state in southwestern Mexico. It is also clear that in both states Cys1,534 was increasing in frequency much earlier than in Ile1,016. Starting in 2002, the frequency of Cys1,534 was greater than or equal to that of Ile1,016. In a yearly comparison of Ae. aegypti collection sites, 80 out of 87 sites (Table 6) had a frequency of Cys 1,534 being greater than the frequency of Ile1,016. In 6 of the 7 cases where the frequency of Ile1,016 exceeded that of Cys1,534, the difference was only from 1–2% and values were not different (overlapping 95% HDI). Only in Martınez de la Torre in 2002 was there a credible difference of 9%.
Linkage disequilibrium analysis can only be performed in datasets where alleles are segregating at both loci. There were 34 datasets that met this criteria of the 87 collections listed in Table 1. Table 7 lists the state, city and year of the 34 datasets along with linkage disequilibrium correlation coefficient Rij and its associated χ2 values and the probability of a greater χ2. Ile1,016 and Cys1,534 were in disequilibrium in the majority (21/34 = 62%) of datasets. For the most part, alleles in 1,534 and 1,016 were evolving in a correlated, dependent fashion. However, this analysis does not provide specific information about the four haplotypes.
The frequencies of the four potential dilocus haplotypes are plotted by year in Fig 4. The frequency of the susceptible Val1,016/Phe1,534 (VF) haplotype remained high from 1999–2003 (Fig 4A). No collections were made again until 2008, by which time frequencies had dropped to 0–0.6. Four years later, VF was approaching extinction in all collections. Fig 4B plots the frequency of the Val1,016/Cys1,534 (VC) haplotype. From 1999–2003, VC frequencies remained low (0–0.10). By 2008, frequencies had increased to 0.1–0.75. Four years later, VC was declining in frequency in two collections and was increasing in four collections. A very different trajectory occurred for Ile1,016/Phe1,534 (IF) (Fig 4C). From 1999–2002, the IF frequency remained low and only reached as high as 0.1 in two collections. By 2008 frequencies were approaching extinction and four years later similar trends were seen, even though VC and IC frequencies had increased dramatically. Fig 4D is a plot of the frequency of the resistant Ile1,016/Cys1,534 (IC) haplotype. From 1999–2002, the IC frequency was low and only reached 0.1 in one collection. By 2008 frequencies had increased dramatically in all collections and continued to increase in all collections up to 2012 when frequencies ranged from 0.5–0.9.
The frequency of Cys1,534 has increased dramatically in the last decade in several states in Mexico including Nuevo Leon in the north, Veracruz on the central Atlantic Coast, and Chiapas, Quintana Roo and Yucatan in the south. The linkage disequilibrium analysis on the Ile1,016 and Cys1,534 alleles in Ae. aegypti collected in Mexico from 2000–2012 (Table 7) strongly supports statistical associations between 1,534 and 1,016 mutations in natural populations. Furthermore, the dynamics of haplotype frequencies during that time suggest pyrethroid resistance in the vgsc gene requires the sequential evolution of 1,534 and 1,016 mutations. Fig 4C suggests that the Ile1,016/Phe1,534 haplotype has a low fitness, even when pyrethroids are being released. For this reason Ile1,016 is unlikely to have evolved independently. Instead it is much more likely that the Cys1,534 mutation evolved first but conferred only a low level of resistance. This conjecture is strongly supported by the fact that in 80 of 87 collections (92%), the frequency of Cys1,534 was greater than the frequency of Ile1,016.
The findings of this study are different in many respects from those in a study of a Tyr1,575 substitution in Anopheles gambiae that occurs just beyond the S6 of domain III, within the linker between domains III and IV [17]. This linker contains a sequence of three amino acids (IFM) that close the sodium channel pore following activation, block the influx of sodium into the cell and restore the membrane resting potential. In contrast, Cys1,534 in Ae. aegypti occurs in the S6 of domain III. This is close to a Met1,524Ile substitution that has been associated with knockdown resistance in Drosophila melanogaster [18] and a Phe1,538Ile mutation that reduces sensitivity to deltamethrin in arthropods and mammals [19, 20].
Mutations in S6 of domain II, such as Phe1,014, Ser1,014 in An. gambie and Ile1,016 and Gly1,016 in Ae. aegypti are not directly in the binding pocket, but affect the resistance phenotype by preventing binding of insecticides and changing the conformation of the VGSC [3, 21]. In contrast, a binding site located in a hydrophobic cavity delimited by the IIS4-S5 linker and the IIS5/IIIS6 helices has recently been proposed [22] that it is accessible to the lipid bilayer and lipid-soluble insecticides. The methyl-cyclopropane (or equivalent structure) of pyrethroids and the trichloromethyl group of DDT appear to be critical features for the action of both pyrethroids and DDT. Both insecticides fit into a slot in a small pocket in the main hydrophobic cavity, flanked by Val1,529 and Phe1,530 on IIIS6. The binding site is formed upon opening of the sodium channel and is consistent with observations that pyrethroids bind preferentially to open channels. This binding pocket includes several known mutations in the S6 of domain III that reduce sensitivity to pyrethroids. Two nearby residues (Gly1,535 and Phe1,538) have been previously implicated in resistance in other insect species (23).
Study in which An. gambiae mosquitos were collected from a range of approximately 2000 km throughout West/Central Africa and had Tyr1,575 occurring at frequencies up to 30% in both M and S forms. Even though this mutation is seen over a large range of the continent, only a single Tyr1,575 haplotype occurred with a Phe1,014 haplotype background (possibly analogous in function to Ile,1016), which infers strong positive selection acting on a recent mutant [17]. In contrast to the present study, Phe1,014 is almost fixed in West Africa and the Tyr1,575 allele is increasing in frequency in M form but not in S form. Thus in contrast to the apparent evolution of Ile1,016 on a Cys1,534 background as reported here in An. gambiae, Tyr1,575 appears to have evolved on a Phe1,014 background. There are many potential reasons for this difference including the possibility that mutations within the S6 of domain III may produce a different resistance mechanism and have a different impact on fitness than mutations in the linker between domains III and IV. It is also possible that the specific changes of amino acids at these sites are unique and may confer different resistance phenotypes. In either case it seems likely that one of the mutations compensates for deleterious fitness effects of the other mutation and/or confers additional resistance to insecticides.
An interesting difference between the two mutations in the present study is that 32% of Ile1,016 heterozygotes recover from pyrethroid exposure but only 3.6% of Cys1,534 heterozygotes recover. Thus while Cys1,534 in synergy with Ile1,016 may confer greater survival following pyrethroid exposure, Ile1,016 may confer a greater ability to recover following knockdown in heterozygotes.
There was evidence of heterozygote deficiency in eight of the 53 collections and the average FIS among these eight collections was large and positive (0.580) while the average among all collections was 0.052. This suggests that the fitness of Phe1,534 and Cys1,534 homozygotes may be greater than the fitness of G/T heterozygotes (i.e. underdominance). While these parameters have been estimated at the 1,016 locus [23], no similar studies have involved the 1,534 locus and so the stability point beyond which the frequency of either allele would increase has not been determined. Since the Cys1,534 confers some degree of pyrethroid resistance (Tables 2–5), directional selection could increase the frequency of Cys1,534 beyond the underdominance stability point, at which stage the frequency of Cys1,534 would rapidly increase towards fixation.
Little is known of other mutations in the Ae. aegypti vgsc that may affect pyrethroid resistance. Codon 989 in the “super-kdr” region of domain II was assessed and no mutations were found [11]. Ile, Met and Val alleles occur at codon 1,011 [11] but these alleles were not associated with resistance in our initial survey of 1,318 mosquitoes from the 32 strains throughout Latin America [11]. The recombination dynamics of the Ae. aegypti vgsc are also poorly understood. Analysis of segregation between alleles at the 1,011 and 1,016 codons in F3 showed a high rate of recombination even though the two codons are only separated by a approximately 250 bp intron [11]. A maximum parsimony phylogeny of the intron spanning exons 20 and 21 in 88 mosquitoes with different genotypes in exons 1,011 and 1,016 indicated the presence of three clades with bootstrap support > 80%. These were arbitrarily labelled clades 1–3. The frequencies of Ile1,011, Met1,011, Val1,011, Val1,016, Ile1,016 and Gly1,016 (from Thailand only) were then compared among the three clades. The frequency of Ile1,011 was distributed independently among the three clades, as was Val1,011 and Met1,011. However, there was a very evident excess of Val1,016 alleles in clade 1 and an excess of Ile1,016 alleles in clade 2. Ile1,016 alleles occurred in disequilibrium with a large number of segregating sites in the intron and a large excess of Ile1,016 alleles were found to be associated with clade 2 in the phylogenetic analysis. This pattern is consistent with a hypothesis that a genetic sweep of the Ile1,016 allele and proximate intron sequences has occurred through DDT exposure and subsequently pyrethroid selection. Furthermore, the genetic sweep was recent enough that there has been insufficient time for recombination to disrupt the disequilibrium between the Ile1,016 allele and proximate intron sequences.
Recent work on the dual binding model may shed some light on the next steps in the evolution of pyrethroid resistance in the vgsc [8]. The Tyr1,575 mutation in An. gambiae was introduced alone into an Ae.aegypti sodium channel (AaNav1-1) [8] and then in combination with Phe1,014. Both substitutions were then functionally examined in Xenopus oocytes [8]. Tyr1,575 alone did not alter AaNav1-1 sensitivity to pyrethroids. However, the Tyr1575- Phe1014 double mutant was more resistant to pyrethroids than the Phe1014 mutant channel alone. Further mutational analysis showed that Tyr1,575 could also synergize the effect of Ser1,014 and Trp1,014, but not Gly1,014, or other pyrethroid-resistant mutations in subunit 6 of domains I or II. Computer modeling predicted that Tyr1,575 allosterically alters pyrethroid binding via a small shift of the subunit 6 of domain II. This establishes a molecular basis for the coexistence of Tyr1,575 with Phe1,014 in pyrethroid resistance, and suggests an allosteric interaction between IIS6 and Loop III/IV in the sodium channel.
The rapid increase in Cys1,534 (Fig 4B and 4D) cannot be the result of neutral forces such as genetic drift or founder’s effects. Parallel increases in Cys1,534 frequency occurred throughout Mexico. Even though the forces that caused an increase in the frequency of Cys1,534 are unclear, our results suggest that Ile1,016 in domain IIS6 arose from a Val1,016/Cys1,534 haplotype and was rapidly selected possibly because double mutants confer higher pyrethroid resistance. When combined with Phe1014, the Tyr1,575 mutation in An. gambiae increased resistance to permethrin and deltamethrin by 9.8- and 3.4-fold, respectively [8].
Fig 5 illustrates two models for the evolution of 1,534 and 1,016 mutations. Model 1 proposes that the 1,534 and 1,016 mutations occurred independently and became cis by crossing over. Model 2 instead proposes that 1,534 mutations occurred first because 1,016 mutations confer low fitness. Ile1,016 mutations then arose on a Val1,016/Cys1,534 background. These results suggest that knowledge of the frequencies of both 1,534 and 1,016 mutations are important to predict the potential of a population to evolve kdr. Obviously, the frequency of Ile1,016 by itself is a poor predictor (Fig 4C). Populations that are pyrethroid susceptible, but have high Val1,016/Cys1,534 frequencies, are at high risk for rapid kdr evolution. If our experience in tracking the frequencies of Ile 1,016 and Cys1,534 mutations over the past 15 years can be extended to other Ae. aegypti populations, then populations with intermediate to high frequencies of Cys1,534 might only be susceptible for 5–10 years. Conversely, pyrethroid susceptible populations without either mutation are unlikely to develop kdr quickly and might be susceptible for up to 10–15 years.
|
10.1371/journal.ppat.1000200 | A Sterol-Regulatory Element Binding Protein Is Required for Cell Polarity, Hypoxia Adaptation, Azole Drug Resistance, and Virulence in Aspergillus fumigatus | At the site of microbial infections, the significant influx of immune effector cells and the necrosis of tissue by the invading pathogen generate hypoxic microenvironments in which both the pathogen and host cells must survive. Currently, whether hypoxia adaptation is an important virulence attribute of opportunistic pathogenic molds is unknown. Here we report the characterization of a sterol-regulatory element binding protein, SrbA, in the opportunistic pathogenic mold, Aspergillus fumigatus. Loss of SrbA results in a mutant strain of the fungus that is incapable of growth in a hypoxic environment and consequently incapable of causing disease in two distinct murine models of invasive pulmonary aspergillosis (IPA). Transcriptional profiling revealed 87 genes that are affected by loss of SrbA function. Annotation of these genes implicated SrbA in maintaining sterol biosynthesis and hyphal morphology. Further examination of the SrbA null mutant consequently revealed that SrbA plays a critical role in ergosterol biosynthesis, resistance to the azole class of antifungal drugs, and in maintenance of cell polarity in A. fumigatus. Significantly, the SrbA null mutant was highly susceptible to fluconazole and voriconazole. Thus, these findings present a new function of SREBP proteins in filamentous fungi, and demonstrate for the first time that hypoxia adaptation is likely an important virulence attribute of pathogenic molds.
| The incidence of potentially lethal infections caused by normally benign molds has increased tremendously over the last two decades. One disease in particular, invasive pulmonary aspergillosis (IPA), caused by the common mold Aspergillus fumigatus, has become the leading cause of death due to invasive mycoses. Currently, we have a limited understanding of how this opportunistic pathogen causes disease in immunocompromised patients. In this study, we discover a previously unexplored mechanism required by this mold to cause disease, hypoxia (low oxygen) adaptation. We report that hypoxia adaptation in A. fumigatus is mediated in part by a highly conserved transcription factor, SrbA, a protein in the sterol regulatory element binding protein family. A null mutant of SrbA was unable to grow in hypoxia, displayed increased susceptibility to the azole class of antifungal drugs, and was avirulent in two distinct murine models of IPA. Importantly, we report the discovery of a novel function of SrbA in molds related to maintenance of cell polarity. The finding that SrbA regulates resistance to the azole class of antifungal drugs presents an opportunity to uncover new mechanisms of antifungal drug resistance in A. fumigatus.
| Aspergillus fumigatus is a normally benign saprophytic fungus that may cause an often lethal invasive disease in immunocompromised patients, invasive pulmonary aspergillosis (IPA) [1],[2]. Interestingly, while IPA can be caused by several Aspergillus species, the majority of IPA cases are caused by A. fumigatus. This may suggest that A. fumigatus contains unique attributes that allow it to cause disease [3]. Yet, the mechanisms utilized by A. fumigatus to survive and cause disease in immunocompromised hosts are not fully understood [4]. During infection, A. fumigatus causes significant damage to host tissue through invasive growth by hyphae and subsequent recruitment of immune effector cells. Thus, infection generates significant inflammation and necrosis in lung tissue that can be visualized by histopathology. These pathologic lesions also likely represent areas of poor oxygen availability to the pathogen and host.
At sites of Aspergillus infection, direct measurements of oxygen tension have not been recorded, however, it is well established that sites of inflammation contain significantly low levels of oxygen (hypoxia) [5]–[7]. Moreover, low oxygen tension has been observed in many compartments of inflamed as well as normal tissues [5]–[7]. In inflamed tissues, the blood supply is often interrupted because the vessels are congested with phagocytes [8],[9]. Indeed, immune effector cells such as neutrophils often function effectively in severely hypoxic microenvironments and have evolved distinct mechanisms to deal with the absence of oxygen that are dependent upon the transcription factor hypoxia inducible factor (HIF) 1.
HIF1 is a heterodimeric transcription factor that consists of a constitutively expressed HIF1β subunit and an oxygen-tension-regulated HIF1α subunit. [10]. Increased HIF1α protein stability and activity of the HIF1 complex, in turn, regulate the transcription of many hypoxia-responsive genes, including those encoding many glycolytic enzymes, erythropoietin, adrenomedullin, and growth factors [11],[12]. Genetic evidence of the importance of hypoxic environments in the regulation of immune responses was recently provided by a study of neutrophil-mediated lung inflammation [13]. Thus, since immune cells of the host have evolved mechanisms to function in hypoxia, it follows that invasive fungal pathogens like A. fumigatus are likely subjected to hypoxia during fungal pathogenesis.
While hypoxic adaptation has not been studied in the context of A. fumigatus pathogenesis, circumstantial evidence suggests that hypoxia plays a key role in the pathophysiology of IPA. For example, it has been postulated that the low rate of Aspergillus recovery from clinical specimens is due to adaptation by the fungus to hypoxic microenvironments found at sites of infection [14],[15]. Furthermore, there are often significant differences in the in vivo and in vitro test results of antifungal drug efficacies. These differences have been postulated to be linked to the occurrence of hypoxia in vivo as demonstrated by recent in vitro antifungal drug efficacy tests conducted in hypoxia [16],[17]. Consequently, it seems probable that pathogenic molds such as A. fumigatus must possess mechanisms to adapt to hypoxic microenvironments found in vivo during infection.
In fact, switching from aerobic respiration to various forms of anaerobic respiration to deal with low oxygen levels has been implicated as an important virulence attribute in several prokaryotic pathogens [18],[19]. However, in eukaryotic pathogens, mechanisms of how these organisms respond and adapt to hypoxia are largely unknown. Most of our knowledge on how fungi respond to hypoxia comes from studies in the model yeast Saccharomyces cerevisiae. Under aerobic conditions, heme biosynthesis activates the transcriptional regulator Hap1p [20]. Hap1p induces genes involved in respiration and oxidative stress-responses, but also activates the transcriptional repressors Rox1p and Mot3p, that down-regulate genes required for hypoxia adaptation [21]. However, in hypoxic conditions, Rox1p and Mot3p are expressed and this leads to transcriptional induction of genes involved in hypoxia adaptation [22]. Thus, hypoxic gene expression in yeast requires transcription factors that utilize Rox1p-binding sequences, low oxygen-response elements (LORE), and other regulatory elements within promoters [23],[24]. Since S. cerevisiae is a facultative anaerobe, it is not surprising that homologs of these key hypoxia gene regulators have not been found in obligate aerobic filamentous fungi such as A. fumigatus.
Recently, a novel mechanism of hypoxia adaptation mediated by a highly conserved family of transcription factors, sterol regulatory element-binding proteins (SREBPs), was characterized in fission yeast, Schizosaccharomyces pombe [25]. SREBPs are a unique family of membrane bound transcription factors first identified in mammals as regulators of cholesterol and lipid metabolism [26]–[30]. Hughes et al. [25] proposed a model in S. pombe where SREBP (Sre1) and a sterol cleavage activating protein (SCAP, Scp1) monitor-oxygen dependent sterol synthesis as an indirect measure of oxygen supply. Importantly, Sre1 was found to be required for adaptation to hypoxia and regulated approximately 68% of the genes transcriptionally induced greater than 2-fold in response to anaerobic conditions [31].
Orthologs of Sre1 and Scp1 were recently identified and characterized in the human fungal pathogenic yeast, Cryptococcus neoformans [32],[33]. As in fission yeast, the SREBP pathway mediated by Sre1 and Scp1 in C. neoformans was crucial for adaptation to hypoxia and sterol biosynthesis. Importantly, these mutants also failed to proliferate in host tissue, failed to cause fatal meningoencephalitis, and displayed hypersensitivity to the azole class of antifungal drugs [32],[33]. In the yeast S. cerevisiae and Candida albicans, orthologs of SREBPs do not appear to exist. However, two similar genes, Upc2 and Ecm22, appear to serve similar functions as SREBPs. Conserved functions of these genes include their involvement in the ability of yeast to grow in hypoxia as well as regulation of sterol biosynthesis and resistance to antifungal drugs [34]–[40]. Taken together, these observations demonstrate an important link between sterol biosynthesis, hypoxia adaptation, azole drug resistance, and the virulence of pathogenic yeasts.
In this study, we report the identification and first characterization of a Sre1 homolog, SrbA, in an opportunistic pathogenic mold, A. fumigatus. Our results suggest that while certain aspects of SREBP function are conserved in yeast and filamentous fungi, significant differences exist that are unique to molds. Thus, our results further expand the spectrum of important functions mediated by SREBPs in eukaryotes, emphasize the importance of this pathway in human fungal pathogenesis, and suggest possible clinical significance of SREBPs related to antifungal drug efficacy.
In order to determine if hypoxia adaptation is an important virulence attribute of filamentous fungi, we first conducted transcriptional profiling experiments using a long-oligo A. fumigatus microarray (version 3.0) of wild type A. fumigatus grown under hypoxic (1% O2) conditions compared to fungus grown under normal conditions (∼21% O2). Analysis of this data revealed ten putative transcription factors that were transcriptionally induced more than 2-fold in response to hypoxia and thus could possibly act as a regulators of genes required for hypoxic adaptation in A. fumigatus (Willger and Cramer, unpublished data). Further bioinformatic analyses of these genes revealed that only one, AFUA_2g01260 (induced 5.02 fold in response to hypoxia), had similarity with a functionally characterized protein, Sre1 from S. pombe [25]. Sre1 shares similarity with mammalian SREBP proteins that regulate lipid and cholesterol homeostasis (reviewed in [29],[30]). In addition, a Sre1 homolog has also recently been described in the human pathogenic yeast, C. neoformans as a regulator of hypoxia adaptation and fungal virulence [32],[33]. AFUA_2g01260 contains 988 amino acid residues, which displayed low sequence percent identity with Sre1 from S. pombe (∼13%) and C. neoformans (∼10%). However, like Sre1 in both yeasts, the amino terminus (amino acids 1–425) of AFUA_2g01260 contains a basic helix-loop-helix (bHLH) leucine zipper DNA binding domain. In addition, AFUA_2g01260 is predicted to contain at least one, and likely two, transmembrane domains. The carboxyl terminus of AFUA_2g01260 is also predicted to contain a conserved domain of unknown function (DUF2014) that is found in other SREBP homologs. Consequently, these results suggest that AFUA_2g01260 is likely the SREBP homolog in A. fumigatus, and we consequently named this gene srbA (sreA is already in use in A. nidulans for an unrelated gene). Additional BLAST analyses revealed that SrbA is highly conserved amongst the filamentous fungi with putative orthologs found in plant pathogens such as Magnaporthe grisea and Alternaria brassicicola and saprophytic molds such as Neurospora crassa and Aspergillus nidulans.
To determine whether SrbA is involved in hypoxia adaptation and fungal virulence in filamentous fungi, we generated a null mutant of the gene encoding SrbA by replacement of the srbA coding sequence in A. fumigatus strain CEA17 with the auxotrophic marker pyrG from A. parasiticus as previously described [41],[42] (Figure 1). The resulting ΔsrbA strain was named SDW1. Ectopic re-introduction of the wild type srbA allele into SDW1 (resulting in strain SDW2) allowed us to attribute all resulting phenotypes specifically to the absence of srbA in SDW1. All strains were rigorously confirmed with Southern blot (Figure 1) and PCR analyses (data not shown). The re-introduced srbA allele in SDW2 displayed similar mRNA abundance in response to hypoxia as the srbA allele in the wild type strain (data not shown). SDW1 and SDW2 both displayed normal hyphal growth rates compared to the wild type strain CEA10 in normoxic conditions on glucose minimal medium (GMM) (Figure 2A) (P>0.01). However, no hyphal growth of SDW1 was observed in hypoxic (1% O2, 5% CO2, 94% N2) conditions whereas wild type strain CEA10 and reconstituted strain SDW2 grew at a normal rate with visual phenotypic differences in colony color and conidiation compared to growth in normoxia (Figure 2A and 2B). In hypoxia, the wild type strains displayed increased aerial hyphae, decreased conidia production, and consequently exhibited a fluffy colony morphology (Figure 2B). After 96 hours of incubation in hypoxia, SDW1 continued to display undetectable growth. However, upon transfer back to normoxic conditions, wild type growth rate was restored (data not shown). Addition of exogenous ergosterol or lanosterol did not rescue the SDW1 growth defect or alter wild type growth morphology in hypoxia (data not shown). These results indicate that A. fumigatus can rapidly adapt to hypoxic microenvironments, and that SrbA in A. fumigatus is involved in mediating this response by an undefined mechanism.
Given the dramatic phenotype observed in strain SDW1 in hypoxia and the sequence annotation that SrbA likely functions as a transcription factor, we next sought to determine which genes are regulated by SrbA under hypoxic conditions. Microarray experiments comparing the transcriptional profiles of wild type strain CEA10 and SDW1 exposed to hypoxia for 24 hours revealed 87 significant genes possibly regulated by SrbA (Table 1). Several genes previously shown to be involved in ergosterol biosynthesis in fungi were found to be transcriptionally repressed in the absence of SrbA including ERG25, ERG24, and ERG3 (Table 1). Interestingly, besides srbA itself, the gene with the highest fold difference in expression in SDW1 is a non-ribosomal peptide synthetase, AFUA_1g10380 (NRPS1 or pes1) [43],[44]. In addition, a significant number of genes involved in cell wall biosynthesis or homeostasis were observed to be repressed in SDW1 compared to wild type. These included genes known to be involved in cell wall biosynthesis such as alpha-galactosidase, alpha-glucosidase B, and genes involved in cell wall homeostasis such as chitinase. However, no obvious defects in cell wall biosynthesis were observed in the mutant strain, and thus the transcriptional profiling results are likely indirect effects of the altered cell polarity of the mutant (discussed below). Genes encoding several transporters were also found to be regulated by SrbA. Overall, these results suggest some similarities, such as with regard to ergosterol biosynthesis, with genes regulated by SREBPs in S. pombe and C. neoformans. However, the overall set of genes putatively regulated by SrbA in A. fumigatus is significantly different from data obtained from Sre1 mutants in the yeast S. pombe and C. neoformans. Consequently, these results strongly suggest that SrbA plays a distinct role in filamentous fungal biology. These results subsequently directed experiments to further characterize the role of SrbA in A. fumigatus biology.
Given the number of ergosterol biosynthesis genes apparently regulated by SrbA, we next asked the question whether SrbA mediated resistance to the azole class of antifungal drugs that target ergosterol biosynthesis. In a screen for susceptibility to antifungal drugs using E-Test strips (AB Biodisk, kindly provided by Dr. Theodore White, Seattle Biomedical Research Institute) we found that SrbA is required for resistance to Fluconazole and Voriconazole, but not Amphotericin B or Caspofungin. All 3 strains showed equivalent minimal inhibitory concentrations (MIC) to Amphotericin B (0.25 µg/ml) and Caspofungin (0.125 µg/ml). The lack of effect of Caspofungin provides support for our hypothesis that the mutant is likely not directly affected in cell wall biosynthesis as possibly suggested by the transcriptional profiling data. However, while CEA10 and SDW2 showed resistance to Fluconazole as expected, SDW1 growth was inhibited at the surprisingly low MIC of 1 µg/ml (Figure 3). On the plates with Voriconazole we could observe that CEA10 and SDW2 were susceptible as expected (MIC of 0.125 µg/ml respectively). Similar to the results with Fluconazole, SDW1 was significantly more susceptible to Voriconazole and showed a MIC of only 0.012 µg/ml (Figure 3). These clinically significant results suggest that SrbA mediates resistance to the azole class of antifungal drugs by an undefined mechanism.
Visual inspection of SDW1 colony morphology in standard laboratory conditions did not reveal any apparent morphological phenotypes (Figure 2B). However, our transcriptional profiling experiments suggested possible alterations in cell wall biosynthesis, a critical component of hyphal morphology and growth, in the absence of SrbA. Consequently, we performed a more in depth analysis of SDW1 morphology. First, we utilized light microscopy to examine the growing edges of SDW1 colonies in normoxia. We observed a significant defect in hyphal tip branching in SDW1 that is not apparent in strains CEA10 and SDW2 (Figure 4). SDW1 hyphal tips display hyper-branching and a “blunted” abnormal morphological phenotype (Figure 4). This phenotype suggests that SrbA is involved in maintaining cell polarity that directs hyphal growth. Interestingly, this phenotype does not appear to alter the growth rate of the colony, which was comparable to the wild type under normoxic conditions (Figure 2A). Next, we utilized transmission electron microscopy (TEM) to further examine the cell wall and morphology of conidia and hyphae of SDW1. Confirming our suspicions that the mutant was not directly affected in cell wall biosynthesis we observed no clear cell wall defects. However, a general thickening of the intracellular space between the cell wall and plasma membrane is observed in SDW1 conidia and hyphae compared with the wild type (Figure 5A and 5B and 5D and 5E). A striking phenotype was consequently observed in conidia from SDW1 that suggested a significant defect in the cell wall-plasma membrane interface occurs in the absence of SrbA (Figure 5A and 5B). This defect is apparently exacerbated by the electron beam, which causes a separation between the cell wall and plasma membrane in SDW1 conidia (Figure 5C). This phenotype was observed in over 80% of the SDW1 conidia examined. However, the size and density of the mutant conidia were comparable to the wild type strain as measured by flow cytometry (data not shown). Since a defect in the cell wall plasma membrane interface was suggested, we examined viability of the SDW1 conidia by monitoring germination. These experiments revealed that viability, as measured by conidia germination, was not significantly different between the wild type, SDW1 and SDW2 strains (Figure 6) (P>0.01). Similar cell wall-plasma membrane defects were observed in the SDW1 hyphae compared with the wild type hyphae (Figure 5D and 5E). Importantly, an accumulation of electron dense objects was observed in the SDW1 hyphae. We hypothesize that these objects may be vesicles of the Spitzenkörper, and their abnormal location in the SDW1 hyphae may cause the observed altered cell polarity (Figure 5E and 5H). This phenotype was observed in over 50% of the SDW1 hyphae examined and never observed in the wild type strain. These results suggest that SrbA is critical for maintaining the cell wall – plasma membrane interface, and that SrbA is critical for normal hyphal branching and cell polarity in filamentous fungi by an undefined mechanism.
Transcriptional profiling of SDW1 under hypoxia suggested that SrbA was involved in both early and late steps of the sterol biosynthesis pathway. In addition, the abnormal conidial and hyphal morphology observed via light microscopy and TEM micrographs in SDW1 also suggested possible alterations in sterol content in the absence of SrbA. Thus, we examined the sterol profile of the SrbA null mutant SDW1 by GC-MS and compared it with the wild type strain CEA10. The GC-MS profiles demonstrated a significant accumulation of 4-methyl sterols in the SrbA null mutant, SDW1, that was not observed in the wild type strain CEA10 (Figure 7). Interestingly, both strains possessed significant amounts of ergosterol (Figure 7). The ratio of C-4 methylated sterols to ergosterol in the absence of SrbA is 1.94 whereas no C-4 methylated sterols accumulated in the wild type. Specifically, the accumulation of 4-methylfecosterol and 4,4-dimethylergosta-8,24(28)-dien-3β-ol in the absence of SrbA suggests a blockage at ERG25 in the sterol biosynthesis pathway in SDW1. These alterations are supported by the transcriptional profiling data, which suggests transcriptional regulation of ERG25 by SrbA in A. fumigatus (Table 1). Consequently, these results suggest a blockage of C4 demethylation in the absence of SrbA in A. fumigatus. In addition, these results suggest that ergosterol can still be synthesized in the absence of SrbA in A. fumigatus.
Next, we sought to determine whether SrbA was required for A. fumigatus virulence. To answer this important question, we utilized two distinct murine models of IPA. In the first model, outbred CD1 neutropenic mice infected with SDW1 displayed no symptoms associated with IPA (Figure 8A). This was in contrast to mice infected with the wild type CEA10 and reconstituted strain SDW2 that displayed well described symptoms of A. fumigatus infection including hunched posture, ruffled fur, weight loss, and increased respiration. Consequently, a significant difference in mortality was observed between the mice infected with SDW1 and mice infected with either SDW2 or CEA1O (P = 0.0002). Indeed, in this murine model, the SDW1 strain was completely avirulent (Figure 8A). We next asked the question whether mice infected with SDW1 were able to clear the infection. After 28 days, SDW1 infected mice displayed no visible or microscopic signs of infection. In particular, at days 14, 21, and 28 lung homogenates were taken from SDW1 infected mice and with the exception of one mouse, no fungal colonies were recoverable indicating that the mice had cleared the infection. Histopathological analyses of mice on days 14, 21 and 28 in this neutropenic model also confirmed the lack of fungal persistence and inflammation in mice infected with SDW1 (Figure 9).
Next, we examined the virulence of SDW1 in a murine model of X-linked chronic granulomatous disease (X-CGD) utilizing gp91phox−/− mice. These mice are deficient in NADPH oxidase activity and display hyper-susceptibility to Aspergillus species without the need for immunosuppression with chemotherapeautic agents [45],[46]. Similar to the neutropenic mouse model, X-CGD mice infected with strain SDW1 had significant differences in survival compared with mice infected with wild type and reconstituted strains (Figure 8B) (P = 0.005). Unlike the neutropenic mouse model, these mice all displayed symptoms of IPA during the preliminary stages of infection. These symptoms, likely due to the large inflammatory response characteristic of these mice when exposed to fungal antigens, included ruffled fur, hunched posture, and lethargic movement as early as 24 hours post-infection. However, only one mouse infected with SDW1 succumbed to the infection. In a repeat experiment, 3 additional X-CGD mice infected with SDW1 also succumbed on day 4 to the infection. Most likely, this was due to the hyper-inflammatory response that occurs in X-CGD mice and not death due to invasive fungal growth. Regardless, the majority of X-CGD mice infected with SDW1 survived the infection and displayed no symptoms of IPA by day 14. Histopathological analyses of these mice displayed standard pathological findings associated with Aspergillus infections in X-CGD mice including the development of granulomatous like lesions, massive influx of inflammatory cells (primarily neutrophils) to sites of infection, subsequent peribronchiolar and alveolar inflammation, and substantial fungal growth in silver stained tissue (Figures 10 and 11). On day 1 of the infection, fungal germination and growth is observed in mice infected respectively with all 3 strains of the fungus. This observation confirms the viability of SDW1 conidia in vivo (Figure 10). Semi-quantitative assessment of the percent of the lung affected by the infection, measured by inflammation and necrosis, of mice infected with the 3 strains respectively revealed no difference at this early time point (CEA10 = 1.3±0.5, SDW1 = 1.3±0.5, SDW2 = 1±0.0). Histopathology on day 4 of the infection, however, revealed extensive growth and proliferation of the wild type and reconstituted SDW2 strain, but minimal fungal growth and proliferation in mice infected with the SrbA null mutant SDW1 (Figure 11). Semi-quantitative assessment of the inflammation and necrosis observed in the lungs of mice infected with the 3 strains respectively at this time point revealed significant differences in the percent of the lung affected by the infection (CEA10 = 3.3±0.5, SDW1 = 2.3±0.5, SDW2 = 3.8±0.5). Lung homogenates from these mice also revealed that viable SDW1 fungus was recoverable from these mice at this time point. This data is consistent with the observed in vitro phenotype of the SDW1 strain in hypoxia. Histopathological analysis of SDW1 infected survivors in this model revealed persistence of granuloma like structures and fungal tissue (Figure 12). Lung homogenates from these animals revealed that the observed fungal tissue was still viable. These results indicate that despite normal growth rates in vitro in normoxic conditions, the SDW1 strain is severely attenuated in its ability to cause lethal disease in two distinct murine models of IPA.
One possible mechanism that could explain the virulence defect of strain SDW1 is an increased susceptibility to oxidative stress as suggested by transcriptional profiling and altered conidia morphology. We examined the growth of CEA10, SDW1, and SDW2 in the presence of 1 mM and 2.5 mM hydrogen peroxide on glucose minimal media. After 48 hours, we observed no detectable difference in growth morphology or colony diameter. In addition, we next examined the ability of RAW264.7 macrophage-like cells to kill SDW1 conidia (Figure 13). As presented in figure 13, no significant difference in conidia killing was observed between CEA10, SDW1, and SDW2 (P>0.01). We conclude that increased susceptibility to oxidative stress and macrophage killing is not responsible for the virulence defect observed in the absence of SrbA.
In this manuscript we present the first characterization of a SREBP in a filamentous fungus. In the yeasts S. pombe and C. neoformans, SREBP homologs are crucial for sterol biosynthesis, survival under hypoxic conditions, resistance to azole antifungal agents, and fungal virulence [25],[32],[33]. Our results confirm that some roles of SREBPs in filamentous fungi are conserved with yeast including, the response to hypoxia, sterol biosynthesis, and susceptibility to the azole class of antifungal drugs. However, our results suggest additional functions of SREBPs in filamentous fungi, most importantly a role in maintenance of cell polarity.
Similarities and differences between SrbA in A. fumigatus and Sre1 in the yeast S. pombe and C. neoformans were apparent from transcriptional profiles comparing the SREBP null mutants to their respective wild type strains in response to hypoxia. Unlike C. neoformans, we did not observe SrbA dependent genes involved in iron or copper uptake in A. fumigatus [32]. This may, however, be a reflection of the experimental conditions that did not place iron stress on the fungus in these experiments. Similar to C. neoformans and S. pombe, we observed SrbA dependent genes involved in ergosterol biosynthesis including ERG25, ERG24, and ERG3 [31]–[33]. This result suggests that regulation of ergosterol biosynthesis is a conserved function of SREBPs in fungi. In A. fumigatus, the SrbA dependent regulation of ERG25 seems to be of particular significance as sterol profiles of the SrbA mutant indicated an accumulation of C-4 methyl sterols suggesting a block in ERG25 function. The effects of decreased ERG3 and ERG24 transcription in the SrbA null mutant is less clear. The accumulation of pathway intermediates may subsequently affect the expression of these genes, and thus, their regulation by SrbA may be indirect. Moreover, A. fumigatus is predicted to have 3 possible orthologs of ERG3 and two of ERG24, which likely indicates a complex regulatory mechanism for ergosterol biosynthesis in A. fumigatus that is mediated in part by SrbA under specific conditions such as hypoxia [47],[48]. Indeed, single mutants of erg3 genes result in no difference in their sterol profiles compared with wild type strains [49].
Other differences with yeast in the transcriptional profile of the SREBP mutant in A. fumigatus suggest important roles for SrbA in filamentous fungal biology. For example, a non-ribosomal peptide synthetase, NRPS1 (or pes1), had the highest change in expression between wild type and the SrbA null mutant [43],[44]. This NRPS has been observed to mediate resistance to oxidative stress in A. fumigatus and displayed an attenuated virulence phenotype in a Galleria mellonella (wax moth) model of aspergillosis depending on inoculum dose [44]. NRPSs are not generally found in most yeast and are particularly abundant in filamentous fungi. Thus, this result suggests that the uncharacterized peptide produced by this NRPS may possibly be involved in hypoxia adaptation as regulated by SrbA in filamentous fungi. Interestingly, we did not observe any increased susceptibility to oxidative stress in the SrbA null mutant. Overall, however, unlike C. neoformans and S. pombe, we did not observe any genes with an annotation that would clearly point to a role in allowing Aspergillus to adapt to hypoxia. This result further illustrates that mechanisms of hypoxia adaptation are almost certainly different in molds than yeast.
Our examination of the SrbA null mutant colony morphology subsequently revealed abnormal branching at the hyphal tips in normoxia and an inability of hyphal growth in hypoxia. Further examination of the mutant with TEM suggested altered vesicle translocation or formation in the hyphae. It is unclear whether these electron dense objects, which we hypothesize are vesicles, comprise the actual Spitzenkörper. At the apex of hyphae in filamentous fungi, the Spitzenkörper is an accumulation of vesicles that is critical for growth directionality [50],[51]. Interestingly, Takeshita et al. (2008) recently observed that localization of key deposition proteins involved in polarized growth at the hyphal tip requires apical sterol-rich membranes [52]. Thus, we hypothesize that the altered hyphal morphology and excessive branching at the tips observed in the SrbA null mutant is due to the alteration in sterol composition of the sterol-rich microdomains in the membrane that are critical for localization of important vesicles and landmark proteins [53]. The alteration in sterol content may cause improper sorting of the vesicles to the apex of the hyphal tip. It is also likely then that the inability of the mutant to grow in hypoxia is related to the perturbation in sterol biosynthesis, a highly oxygen dependent pathway reported to require at least 22 molecules of oxygen.
We could not rescue the SrbA phenotype in hypoxia with addition of ergosterol or lanosterol (data not shown). Nor did exogenous addition of these sterols alter growth of the wild type strain in hypoxia as is the case for S. cerevisiae, which requires exogenous sterols for anaerobic growth. These results may suggest that A. fumigatus does not import exogenous sterols in hypoxic conditions, that SrbA may be in part responsible for exogenous sterol uptake, or that the defect is not due to loss of ergosterol or lanosterol. We feel that the latter explanation is most likely as A. fumigatus has been observed to take up and utilize exogenous cholesterol [54]. We observed that the A. fumigatus SrbA null mutant produced substantial levels of ergosterol even in the absence of SrbA. Thus, even though the ergosterol biosynthesis pathway appears blocked at ERG25 in the SrbA mutant, alternative mechanisms exist for A. fumigatus to produce ergosterol in the absence of SrbA and presumably ERG25 activity. This finding is consistent with a recent report which suggested that A. fumigatus likely possess at least three alternative pathways for ergosterol biosynthesis [48]. Also, an analysis of the A. fumigatus genome sequence revealed that A. fumigatus contains duplicate and even triplicate copies of many of the ergosterol biosynthesis genes [47],[55]. Thus, it appears that A. fumigatus contains complex regulatory mechanisms, of which SrbA is a part, for the production of ergosterol that remain to be elucidated.
Based on our current knowledge of the pathophysiology of IPA, the in vitro phenotypes observed in the SrbA mutant would not predict a role for this protein in A. fumigatus virulence. However, the SrbA null mutant was virtually avirulent in two distinct murine models of IPA despite a normal growth rate of the fungus in standard laboratory conditions. Consequently, we believe two possible explanations exist for the observed avirulent phenotype of the SrbA null mutant. First, and we believe most likely, the inability of the SrbA null mutant to grow in hypoxia prevents invasive disease from being established. Once hypoxia is generated during Aspergillus infection, the mutant simply can no longer grow and proliferate, allowing what immune effector cells that remain functional the ability to ultimately clear the infection. An alternative hypothesis is that the altered hyphal morphology and excessive branching observed in the SrbA mutant in normoxic conditions results in a strain incapable of invasive growth or a strain more susceptible to clearance by the immune system. To examine these alternatives, we employed the use of two distinct murine models of IPA.
We first examined the SrbA mutant virulence phenotype in a persistently neutropenic mouse model characterized by the use of high doses of cyclophosphamide and Kenalog [42]. Currently, it is unclear what specific components of the immune system are affected in this model, but it is clear that differences in the immunosuppression regimen can significantly affect the outcome of infection [56],[57]. In this model, significant inflammation and tissue necrosis is observed in histopathological examinations. We hypothesize that these sites of infection and inflammation in this model are hypoxic. Thus, we believe that A. fumigatus must overcome significant hypoxia during pulmonary infections, and the inability of the SrbA null mutant to adapt to hypoxic conditions results in rapid cessation of invasive growth and a lack of lethal disease. Our histopathological findings with the SrbA mutant strain revealed fungal growth in this model early in the infection. However, by day 14, we were unable to recover viable colonies from mice infected with the SrbA null mutant strain. Indeed, by day 14 of the infection, little evidence of inflammation or fungal burden was evident in mice infected with the SrbA null mutant. These two results suggest that growth of the fungus was halted and what immune effector cells present in the immunosuppressed mice were able to clear the infection. Furthermore, our in vitro experiments revealed that the growth defect of the SrbA mutant in hypoxia was not fungicidal but fungistatic. Thus, if growth simply were halted in the animals without immune system clearance, we would have expected to recover viable fungal colonies from the infected mice.
To further examine the apparent virulence defect of the SrbA null mutant, we utilized a mouse strain highly susceptible to Aspergillus infections, the X-CGD gp91phox−/− mice [45],[46]. These mice exhibit a hyper-inflammatory response when exposed to A. fumigatus and other Aspergillus species. We chose this particular animal model for our experiments given the very specific defect in NADPH oxidase function in these mice, and with the hypothesis that the hyper-inflammatory response would generate significant hypoxia in the lung. Given the extreme susceptibility of these mice to A. fumigatus, we hypothesized that if the SrbA null mutant could grow and persist in vivo, even at a reduced rate, we should observe significant mortality in these mice. However, in contrast, we observed limited mortality in these mice when inoculated with the SrbA null mutant, strongly suggesting that the mutant simply cannot grow effectively in vivo to cause invasive disease. Unlike the neutropenic mouse model, extensive signs of chronic inflammation remained evident in the X-CGD mice post-day 14, consistent with previously reported results in these animals [45]. Furthermore, unlike the neutropenic mice, we could detect the persistence of viable SDW1 in the lungs of these surviving mice out to day 14.
Consequently, we conclude that these observations strongly suggest that the inability of the SrbA null mutant to grow in hypoxic microenvironments is primarily responsible for the avirulent phenotype of the mutant. Though the altered cell polarity of the SrbA mutant may contribute to the virulence defect, the fact that SrbA null mutant displayed normal growth rates in vitro in standard laboratory growth conditions suggests to us that the altered cell polarity did not significantly affect fungal growth. Furthermore, we also have examined the susceptibility of the SrbA null mutant conidia to macrophage (RAW264.7 cells) killing and found no difference with the wild type strain. In addition, the SrbA mutant did not display increased sensitivity to hydrogen peroxide. Taken together, we feel these observations strongly suggest that the virulence defect in the SrbA null mutant is due to its inability to grow in hypoxia.
An additional observation of clinical significance was the finding that SrbA mediates resistance to the azole class of antifungal drugs. Interestingly, loss of SrbA resulted in a strain of A. fumigatus highly susceptible to fluconazole, an azole that normally has minimal activity against A. fumigatus [58],[59]. The mechanism(s) behind this result are currently not known. Transcriptional profiling of the SrbA mutant revealed numerous transporters possibly regulated by SrbA. Thus, the mechanism behind the increased azole susceptibility may be due to loss of transcription in specific transporters in the SrbA mutant. This hypothesis is currently being tested in our laboratory. Second, a relationship between mitochondria function, sterol homeostasis, and azole drug resistance has been observed in the yeast S. cerevisiae and Candida glabrata [60],[61]. Thus, the altered accumulation of sterol intermediates in the SrbA mutant may alter the resulting interaction with fluconazole and mitochondria. With a similar increase in susceptibility to azoles in the SREBP mutant in C. neoformans, it seems clear that further study of the SREBP pathway and azole drug resistance in pathogenic fungi is highly warranted. Identification of ways to inhibit this pathway in vivo may increase the efficacy of current azole antifungal agents [32],[33]. Thus, further studies are needed to dissect this important pathway in yeast and molds to identify conserved targets that may be harnessed to treat patients with invasive mycoses.
Finally, in this study, we did not focus on elucidating the molecular mechanism behind SrbA regulation and activation in molds. However, several observations from our studies hint at possible mechanisms. First, we identified SrbA in a transcriptional profiling screen of A. fumigatus in response to hypoxia (induced >5 fold). This suggests that SrbA may be transcriptionally regulated in molds. However, HIF1 in humans also responds transcriptionally to hypoxia, but its activity is primarily post-translationally regulated [62],[63]. In the yeast S. pombe and C. neoformans, it seems clear that Sre1 is regulated post-translationally in response to sterol biosynthesis perturbation that occurs in low oxygen environments. Indeed, Hughes et al. (2007) have identified 4-methyl sterols as the primary activating agent of Sre1 in S. pombe [64]. Thus, our finding that the SrbA null mutant in A. fumigatus accumulates 4-methyl sterols may also suggest that these sterols are the trigger for SrbA activation in A. fumigatus.
While many of the phenotypes we observed in the SrbA mutant in A. fumigatus may suggest that SrbA is regulated in a similar manner as Sre1 in yeast, our results may also suggest an alternative model in molds. First, despite extensive bioinformatic analyses, we were unable to identify a clear homolog of the sterol cleavage activating protein (SCAP). SCAP is highly conserved in yeast, mammals, and insects and thus it is surprising that bioinformatic searches were unable to identify a clear homolog in any filamentous fungi with genome sequences available. However, some candidates with minimal sequence similarity are being pursued in our laboratory. Second, the observation that sterol biosynthesis was altered in normoxia, likely resulting in altered cell polarity, suggests that in molds, SrbA plays a significant role in the biology of filamentous fungi in normoxic conditions. Third, though sequence identity was extremely low, generation of null mutants in putative site-1 (S1P) and site-2 (S2P) protease homologs in A. fumigatus did not demonstrate expected defects in hypoxic growth (Willger and Cramer, unpublished data). Additional proteases remain to be explored. We could, however, identify a clear Insig1 homolog, which we have named InsA. In mammals, Insig is a key regulator of SREBP function where it binds to SCAP and prevents SREBP cleavage in the presence of sterols by maintaining the SREBP-SCAP complex in the endoplasmic reticulum membrane [65],[66]. We are currently characterizing a possible role for InsA in SREBP signalling in filamentous fungi. Interestingly, C. neoformans lacks an apparent Insig homolog and the Insig homolog in S. pombe does not appear to be required for regulation of SREBP signalling [25],[32]. Taken together, these results suggest that while aspects of SrbA signalling in filamentous fungi may be conserved in yeast and mammals, it is likely that significant differences exist in molds that remain to be elucidated. What is clear, however, is that SREBPs play critical roles in the biology of fungi that have important implications for fungal virulence and how we manage and treat invasive fungal infections. Future studies on this pathway in A. fumigatus are likely to yield important insights into sterol metabolism, hypoxia adaptation, fungal growth, and mechanisms of azole drug resistance.
A. fumigatus strain CEA17 (a gift from Dr. J.P. Latgé, Institut Pasteur) was used to generate the srbA null mutant strain, SDW1 (ΔsrbA::A. parasiticus pyrG pyrG1). A. fumigatus strain CEA17 is a uracil-auxotrophic (pyrG1) mutant of A. fumigatus strain CEA10 [67],[68]. In this study we used CEA10 (gift from Dr. Thomas Patterson, University of Texas- San Antonio Health Sciences Center) as the wild type, SDW1, and an ectopic complemented control strain SDW2 (Δsrb::A. parasiticus pyrG+srbA). All strains were stored as frozen stocks with 50% glycerol at −80°C. The strains were routinely grown in glucose minimal medium (GMM) with appropriate supplements as previously described [69] at 37°C. To prepare solid media 1.5% agar was added before autoclaving.
Generation of a srbA null mutant in A. fumigatus strain CEA17 was accomplished by replacing an ∼2.2-kb internal fragment of the srbA coding region (∼3.0 kb; GenBank accession no. XM_744169) with A. parasiticus pyrG. The replacement construct was generated by cloning a sequence homologous to the srbA locus into plasmid pJW24 (donated by Dr. Nancy Keller, University of Wisconsin—Madison). Homologous sequences, each ∼1 kb in length and 5′ and 3′ of the srbA coding sequence, were cloned to flank A. parasiticus pyrG in pJW24. The resulting plasmid, pSRBAKO, was used as a template to amplify the ∼5.1-kb disruption construct for use in fungal transformation. To complement the ΔsrbA strain SDW1 the srbA gene was amplified using genomic DNA of CEA10 as template and the primers 5′SrbAKOLF and 3′SrbAKORF. The ∼5.9-kb PCR product was used in a fungal transformation and selection was for colonies able to grow under hypoxic conditions. The primers utilized in vector construction are presented in Table S1.
Generation of fungal protoplasts and polyethylene glycol-mediated transformation of A. fumigatus were performed as previously described [70]. Briefly, 10 µg of the srbAKO PCR-generated replacement construct was incubated on ice for 50 min with 1×107 fungal protoplasts in a total volume of 100 µl. Transformants were initially screened by PCR to identify potential homologous recombination events at the srbA locus. PCR was performed with primers designed to amplify only the disrupted srbA locus (5′SrbAKOLF and 3′PyrGKOLF; 5′PyrGKORF and 3′SrbAKORF) (Table S1). Homologous recombination was confirmed by Southern analysis with the digoxigenin labeling system (Roche Molecular Biochemicals, Mannheim, Germany) as previously described [71]. To eliminate the chance of heterokaryons, each transformant was streaked with sterile toothpicks a minimum of two times to obtain colonies from single conidia.
Strains were grown on GMM plates at 37°C. Normoxic conditions were considered general atmospheric levels within the lab (∼21% O2). For hypoxic conditions a Hypoxia Incubation Chamber (MIC-101; Billups-Rothenberg, http://www.hypoxiaincubator.com) was used. The chamber was maintained at 37°C and kept at ∼1% oxygen level utilizing a gas mixture containing 1% O2, 5% CO2 and 94% N2. In addition, hypoxia experiments requiring shake-flask cultures were conducted in a Biospherix C-Chamber with O2 levels controlled by a PRO-Ox controller and CO2 levels controlled with PRO-CO2 controller (Biospherix, Lacona, NY). For these experiments, O2 set point was 1% and CO2 set point was 5%.
Colony growth was quantified as previously described [72]. Briefly, 5-µl aliquots containing 1×106 conidia from freshly harvested GMM plates were placed in the center of GMM agar plates. Plates were then cultured under normoxic or hypoxic conditions. Diameters of three colonies per A. fumigatus strain and condition were measured once daily over a period of 4 days. The average change in colony diameter per 24 h of growth was calculated from three independent cultures. Conidia were harvested with 20 ml of sterile 0.01% Tween 80, filtered through two layers of sterile miracloth (EMD Biosciences, La Jolla, CA), and quantified.
Conidia from freshly harvested GMM plates were inoculated in 5 ml GMM in a 6-well plate to a concentration of 1×107/ml. Cultures were grown aerobically for 24 h. For normoxic growth, cultures were maintained in atmospheric conditions. For hypoxic growth, cultures were placed in the hypoxic chamber for 24 h. Fungal mats were flash frozen in liquid nitrogen and lyophilized prior to total RNA extraction using TRIsure Reagent (Bioline) according to the manufacturer's instructions. RNA was further purified using the RNeasy Mini Kit (Qiagen) and re-suspended in DEPC-treated water. RNA integrity was confirmed with an Agilent Technologies Bioanalyzer.
Total RNA was reverse transcribed by priming with oligo dT and utilizing aminoallyl-dUTP. The resultant cDNA was then coupled to Cy3- and Cy5-labeled probes (GE Healthcare), and hybridized to Aspergillus fumigatus version 3 microarrays from the pathogen functional resources center (PFGRC) as described in the TIGR standard operating procedures found at http://atarray.tigr.org. Labeled cDNA from wild type grown in hypoxic conditions was hybridized against cDNA from SDW1 grown in hypoxic conditions. Data for each strain represents six independent experiments and includes three dye swaps. Arrays were scanned on an Axon 4000B scanner with GenePix software at the Montana State University Functional Genomics Core facility (Axon Instruments). Array signals were bulk-normalized and filtered for flagged spots using MIDAS (available at http://www.tm4.org/midas.html). Data were log-transformed (base 2) and filtered for genes that contained data for at least three out of four arrays from each strain, and missing values were calculated through K-nearest neighbor algorithm using Significance Analysis of Microarrays (SAM) software [73] prior to statistical analysis by SAM. Statistically significant genes identified by SAM with 2-fold or greater changes in expression are listed in Table 1. A Delta cutoff in SAM that captured the maximum number of significant genes with a false discovery rate of zero was utilized. Microarray data has been deposited in the Gene Expression Omnibus (GEO) at the National Center for Biotechnology Information (NCBI) series accession number GSE12376.
E-test strips (AB Biodisk, N.J.) plastic strips impregnated with a gradient of Fluconazole, Voriconazole, Caspofungin, or Amphotericin B were used per manufacturers' instructions. Each strip was placed onto a RPMI-1640 (Sigma Aldrich) agar plate containing a lawn of conidia and growth inhibition was measured after 24 and 48 h by direct observation of the plates at 37°C. No difference in results was observed between 24 and 48 h.
Sterols were extracted following published protocols [74]. Gas-chromatography-Mass spectrometry analyses were performed with a HP6890 GC coupled to a HP5973 mass selective detector. Electron impact MS (70 eV, scanning from 50 to 550amu, at 2.94 intervals/sec) was performed using the following conditions: HP-5 column (30 m×0.25 mm i.d., 0.25 mm film thickness), Helium as carrier gas (1 ml/min), detector temperature 180°C, column temperature 100°C to 300°C (100°C for 1 min, 7°C/min to 300°C then held for 15 min). All injections were run in splitless mode.
Conidia and mycelia of wild type and SDW1 were examined by transmission electron microscopy (TEM). Conidia released in sterile water from 5-day-old GMM plates and mycelia grown in liquid GMM for two days were collected by centrifugation at 5000× g for 10 min. The conidial and mycelial pellets were coated with 0.8% agarose and fixed in modified Karnovsky's fixative containing 2% paraformaldehyde and 2% (v/v) glutaraldehyde in 0.05 M sodium cacodylate buffer (pH 7.2) overnight at 4°C. After washing three times with 0.05 M sodium cacodylate buffer (pH 7.2) for 10 min each, samples were post-fixed with 1% (w/v) osmium tetraoxide in the same buffer for 2 hours at 4°C. The post-fixative was removed by washing briefly twice with distilled water at room temperature and the samples were en bloc stained with 0.5% uranyl acetate overnight at 4°C. The samples were then dehydrated in a graded ethanol series, rinsed with propylene oxide, and embedded in Eppon resin (Fluka AG, Zürich, CH). Ultrathin sections cut from the Eppon-embedded material with ultramicrotome (MT-X, RMC, USA) were collected on carbon-coated grids, stained with 2% uranyl acetate for 3 min, and with Reynold's lead solution [75] for 3 min. Examination was conducted with a JEM-1010 (JEOL, Tokyo, Japan) electron microscope operating at 60 kV.
For the conidia germination assay, A. fumigatus strains were grown in 25 ml GMM with 2% yeast extract. Cultures were inoculated with approximately 106 conidia/ml. After 7 hours the germination rate was determined by counting a total of 100 spores and noting the number of germinated spores. Counting was repeated three times for each strain and the mean and standard deviation are reported.
In this study two different mouse models were used to assess the role of the transcription factor SrbA in fungal virulence. For the persistently neutropenic mouse model we used outbred CD1 (Charles River Laboratory, Raleigh, NC) male mice (26 to 28 g in size, 6–8 weeks old), which were housed six per cage and had access to food and water ad libitum. Mice were immunosuppressed with intraperitoneal (i.p.) injections of cyclophosphamide at 150 mg/kg 2 days prior to infection and with Kenalog injected subcutaneously (s.c.) at 40 mg/kg 1 days prior to infection. On day 3 post-infection (p.i.), repeat injections were given with cyclophosphamide (150 mg/kg i.p.) and on day 6 p.i. with Kenalog (40 mg/kg s.c.). Twelve mice per A. fumigatus strain (CEA10, srbA-deficient mutant SDW1, or the reconstituted strain SDW2) were infected intranasally. For an alternative mouse model, we used breeder mice with a null allele corresponding to the X-linked gp91phox component of NADPH oxidase (B6.129S6-Cybbtm1Din). Breeding pairs of these mice were obtained from the Jackson Laboratory (Bar Harbor, Maine) and reared in the Animal Resource Center at Montana State University. All animals were kept in specific-pathogen-free housing, and all manipulations were approved by the institutional internal review board (IACUC). To avoid exposing gp91phox−/− mice to bacterial infections, they were housed in microisolator cages in an environment of filtered air and given autoclaved food ad libitum and prophylactic treatment with sulfamethoxazole-trimethoprim in their sterile drinking water. The animals were used at 8 to 13 weeks of age. The mice were inoculated intratracheally following brief isoflurane inhalation, returned to their cages, and monitored at least twice daily.
Infection inoculum was prepared by growing the A. fumigatus isolates on GMM agar plates at 37°C for 3 days. Conidia were harvested by washing the plate surface with sterile phosphate-buffered saline-0.01% Tween 80. The resultant conidial suspension was adjusted to the desired concentration of 1×106 conidia/40 µl by hemacytometer count. Mice were observed for survival for 14 days after A. fumigatus challenge. Any animals showing distress were immediately sacrificed and recorded as deaths within 24 h. Mock mice were included in all experiments and inoculated with sterile 0.01% Tween 80. Lungs from all mice sacrificed during the experiment were removed for fungal burden assessment and histopathology. Experiments were repeated in duplicates with similar results. Survival was plotted on a Kaplan-Meier curve and a log-rank test used to determine significance of pair-wise survival (two-tailed P<0.01). No mock infected animals perished in either murine model in all experiments.
For histopathology studies, additional gp91phox−/− mice were infected as described above, and sacrificed at set time points of day 1 and day 4 after A. fumigatus challenge. When mice were sacrificed, lungs were removed on that day. Lung tissue was fixed in 10% phosphate-buffered formalin, embedded in paraffin, sectioned at 5 µm, and stained with hematoxylin and eosin (H&E) or Grocott methenamine silver (GMS) by using standard histological techniques. Microscopic examinations were performed on a Zeiss Axioscope 2-plus microscope and imaging system using Zeiss Axiovision version 4.4 software. Semi-quantitative analysis of inflammation and necrosis were scored on a scale of 1 to 5. The scale consisted of: 1 = 0 to 24% lung involvement, 2 = 25–49%, 3 = 50–74%, 4 = 75–99% 5 = 100%. H&E stained whole lungs from 4 mice infected with each respective strain were assessed to determine the percentage involvement and scored accordingly on days 1 and 4 of the infection in consultation with a pulmonary immunologist.
Macrophage killing of conidia was measured by serial dilution as previously described with slight modifications [76]–[78]. Briefly, 2.5×105 RAW264.7 cells in a volume of 500 ml were inoculated into 24 well tissue culture treated cell culture plates (Corning Incorporated, Corning, NY) in DMEM complete media and incubated overnight at 37°C, 5% CO2. A total of 1.25×106 freshly harvested A. fumigatus conidia of the respective strains in DMEM complete media were inoculated into each well to give a conidia∶macrophage ratio of 5∶1. Co-incubation was performed at 37°C, 5% CO2 for 1 hour, after which media was removed and cells were gently washed with 1× phosphate buffered saline (PBS) to remove non-phagocytosed conidia. At this time point, conidia from each strain were harvested from macrophages in one well to establish the baseline number of conidia engulfed. DMEM complete media was added back to the non-harvested wells and incubation proceeded for an additional 5 hours. Lysis of macrophages was performed by treating the cells with 200 ml of a 0.5% SDS solution for 10 minutes followed by addition of 200 ml of 1× PBS. The percentage of colony forming units (CFU) from conidia∶macrophage co-incubations was determined relative to control conidia harvested at the one hour time point. Controls were performed by lysing macrophages as described above after phagocytosis of conidia for 1 hour and CFU counts were set to 100%. Experiments were performed with triplicate wells and repeated two times for each A. fumigatus strain.
For the oxidative stress assay, the A. fumigatus strains were grown on GMM plates with and without H2O2. GMM plates with 1 and 2.5 mM H2O2 were prepared. Plates were inoculated with approximately 100,000 spores in 5 µl and incubated at 37°C. Sensitivity to oxidative stress was determined by comparing the colony radius of 2-day-old cultures on plates with H2O2. The assay was repeated three times for each concentration. Growth of each strain on each plate was visually examined.
The software program Prism 5 (GraphPad, San Diego, Calif.) was used for all statistical tests of significance (to P values of ≤0.01). Normally, a two-sided t test was used to compare two groups of data, with Welch's correction being used if the groups had unequal variances. In cases in which a deviation from a normal distribution was suspected, a nonparametric test (Mann-Whitney test) was also applied. In those cases, we found that both the t test and Mann-Whitney test indicated the same results (i.e., both indicated significance or insignificance); however, typically one test gave a more conservative (larger, but still <0.01) P value. The P values we report are always the conservative values. Log-rank tests were utilized to determine significance of survival in animal studies.
|
10.1371/journal.pcbi.1003671 | Comparison of REST Cistromes across Human Cell Types Reveals Common and Context-Specific Functions | Recent studies have shown that the transcriptional functions of REST are much broader than repressing neuronal genes in non-neuronal systems. Whether REST occupies similar chromatin regions in different cell types and how it interacts with other transcriptional regulators to execute its functions in a context-dependent manner has not been adequately investigated. We have applied ChIP-seq analysis to identify the REST cistrome in human CD4+ T cells and compared it with published data from 15 other cell types. We found that REST cistromes were distinct among cell types, with REST binding to several tumor suppressors specifically in cancer cells, whereas 7% of the REST peaks in non-neuronal cells were ubiquitously called and <25% were identified for ≥5 cell types. Nevertheless, using a quantitative metric directly comparing raw ChIP-seq signals, we found the majority (∼80%) was shared by ≥2 cell types. Integration with RNA-seq data showed that REST binding was generally correlated with low gene expression. Close examination revealed that multiple contexts were correlated with reduced expression of REST targets, e.g., the presence of a cognate RE1 motif and cellular specificity of REST binding. These contexts were shown to play a role in differential corepressor recruitment. Furthermore, transcriptional outcome was highly influenced by REST cofactors, e.g., SIN3 and EZH2 co-occupancy marked higher and lower expression of REST targets, respectively. Unexpectedly, the REST cistrome in differentiated neurons exhibited unique features not observed in non-neuronal cells, e.g., the lack of RE1 motifs and an association with active gene expression. Finally, our analysis demonstrated how REST could differentially regulate a transcription network constituted of miRNAs, REST complex and neuronal factors. Overall, our findings of contexts playing critical roles in REST occupancy and regulatory outcome provide insights into the molecular interactions underlying REST's diverse functions, and point to novel roles of REST in differentiated neurons.
| The RE-1 silencing transcription factor (REST) binds to DNA and has been shown to repress neuronal genes in non-neuronal systems, but more recent studies have expanded its functions much beyond this. At the molecular level, REST acts cooperatively with other proteins to execute its transcriptional regulatory roles. The dynamics of REST binding and cofactor recruitment and its association with the underlying DNA sequence remain unclear. Here, we have applied chromatin immunoprecipitation and deep sequencing to identify REST binding across 16 different cell types, including neurons. Our results demonstrate that REST binding events are dynamic and quite distinct among cells and that REST binding is generally associated with low gene expression. Closer examination finds that the context of the DNA sequence at REST bound sites is associated with the lower expression of REST-associated targets and that different contexts correlate with different cofactor recruitment. These in turn have an effect on the expression of REST targets. REST targets in human neurons, however, are drastically different from those in other cell types. These findings provide insights into the effect of genomic and cellular contexts on REST's diverse functions and point to distinct and novel roles for REST in neurons.
| The REST (RE1-silencing transcription factor) [1], also known as NRSF (Neural Restrictive Silencing Factor) [2] and XBR (X2 Box Repressor) [3], encodes a zinc-finger transcription factor that was initially shown to repress neuronal genes in non-neuronal tissues and neural progenitors. It has since been shown to play a broad range of roles in neuronal differentiation and development [4]–[6], such as fine-tuning neural gene expression [7] and modulating synaptic plasticity [8]. REST is necessary for the maintenance of self-renewal capacity of neural stem cells (NSCs), as its knockdown led to a lower mitotic index and a higher rate of early neuronal differentiation [5]. REST has also been implicated as a tumor suppressor in breast cancer, colorectal cancer and small cell lung cancer, and as an oncogene in neuroblastomas, medulloblastomas and pheochromocytomas, which are associated with von Hippel-Lindau syndrome [9], [10]. These findings show that REST plays diverse roles in multiple cellular processes.
In addition to the 21-bp DNA sequence bound by REST (termed the RE1 motif), an array of cofactors have been found to interact and cooperate with REST, including SIN3, CoREST, Polycomb Repressive Complexes (PRCs), and various histone deacetylases (HDACs) [9], [11], [12]. Many of these cofactors are chromatin modifiers or are associated with enzymes that have effects on post-translational histone modifications, suggesting that at the molecular level REST functions as a platform for the recruitment of multiple chromatin modifiers and that together they orchestrate gene regulation [9], [11], [12]. In fact, REST occupancy has been found to correlate with an increase of repressive and a decrease of active histone modifications [13]. Not all of the REST cofactors, however, are recruited to each of the REST-bound loci concomitantly. For example, a study of REST occupancy in eight RE1 loci in mouse NSCs found the existence of four distinct configurations of REST and its cofactors: REST-Sin3b-CoREST-HDAC1/2, REST- Sin3b-CoREST, REST-CoREST, and REST-Sin3b-HDAC1/2 [14]. An independent genome-wide study showed that approximately half of the REST-bound sites in mouse ESCs were associated with REST cofactors (in various combinations of the Sin3 and CoREST family members) and that genes targeted by REST together with its cofactors showed more repression than genes bound by REST alone [15]. These studies indicate that differential recruitment of REST cofactors can potentially orchestrate distinct transcriptional outcomes. Moreover, REST acts as a repressor at only a subset of RE1 containing genes [16], [17]; in particular, REST and its splicing variants have been reported to activate a variety of genes in certain cell types and conditions, such as CHRNB2, CRH, PAX4, and OPRM1 [18]–[23].
Previous genome-wide studies have characterized REST targets in human Jurkat cells [24], K562 cells [25], APL blasts [26], mouse embryonic [15], [27]–[29] and neural stem cells [27], [28]. Widespread switches in REST targeting during mouse neuronal and glial differentiation have also been reported [7], [30], [31]. Moreover, analysis of 1% of the human genome (the ENCODE pilot regions [32]) using ChIP-chip technology has found interesting context-dependent REST functions among cell types [33]. While these previous studies have provided important information and principles about REST chromatin interaction and gene regulation, they have not systematically addressed how REST cistromes differ across diverse human cell types. In our current study we set out to address this by integrated analysis of ChIP-seq and RNA-seq data, including primary cells and differentiated neurons. By a comprehensive analysis of ChIP-seq data for 16 cell types, including one collected for this study for CD4+ T cells and 15 cell types from the ENCODE project [34], we have identified a total of 21,134 non-redundant REST binding sites (i.e., peaks) in the human genome. Among the REST sites in 15 “non-neuronal” cell types, only 7% were common to all. We then studied how RE1 motif status, genic location and the cellular context were related to REST binding and gene expression, as well as how these contexts were related to co-factor colocalization. Finally we compared REST occupancy between neurons and non-neuronal cells and found that REST bound to a distinct set of targets in neurons. Moreover, in contrast to other cell types, REST binding was largely localized to genes that were activated in differentiated neurons, as indicated by high levels of gene expression and active histone modifications. Our study provides valuable insights into the dynamic landscape of REST-chromatin interactions in the human genome and the importance of genomic and cellular contexts in modulating the outcome of REST regulation.
To investigate the roles of REST occupancy in cell types relevant to normal physiology, we have performed a ChIP-seq analysis on human CD4+ primary T cells and compared the results with REST ChIP-seq data for 28-day-old differentiated human neurons (derived from the H1-ESC line), which were obtained from the ENCODE project [34] (Table S1). In order to make the data uniform and comparable across cell types, we called REST binding sites (i.e., ChIP-seq peaks) by the same pipeline and with consistent parameters across datasets: using the program SPP [35] and the IDR methodology recommended by the ENCODE project [34] (see Methods). As a result, we called 4,404 REST binding sites for the T cells and 5,387 for the neurons (Table 1). From the T cell REST peaks, we selected 10 sites and confirmed the binding of nine by ChIP-qPCR; the qChIP enrichments were consistent with the peak enrichment scores provided by SPP (Fig. S1). 38% of the T cell REST peaks had the canonical 21 bp RE1 motif (cRE1) (Fig. 1A), while an additional 3% contained a non-canonical RE1 motif (ncRE1), in which two halves of the cRE1 motif were separated by a 1–10 nucleotide insertion [24], [27] (Fig. 1A, bottom). Consistent with previous reports, we also found that a large fraction (22%) of the REST peaks had only one of the two half-sites. In total, 63% of REST peaks in T cells had the cRE1 motif or one of its variants. For the neuron peaks, motif analysis revealed an unexpected and different picture. Less than 10% of the REST peaks in neurons contained either the cRE1 or ncRE1 motifs (Fig. 1B). Even the enrichment of RE1 motifs in the top 600 neuronal REST peaks was marginal, occurring in 22 sites (p = 4.9E-64 from MEME [36]). A second enriched motif, GGAAA/TA, was detected among these peaks (n = 180, p = 1.6E-50) (Fig. S2). It is similar to the DNA motifs recognized by transcription factors NFATC2, dl_2 and EDS1, and it was found in 12% of the total H1-derived neuron REST peaks, compared to 3% for the T cell REST peaks and 9% of randomly selected genomic sequences.
A comparison of the REST-bound genes further demonstrated the distinction between REST targeting in T cells and neurons. The 4,404 peaks identified in T cells were associated with 3,307 Refseq [37] genes and miRNAs [38], while the 5,387 peaks in H1-derived neurons were associate with 3,389 genes/miRNAs. Despite a previous report that RE1 motifs were prominently distributed in introns [24], 51% and 41% of the REST peaks were localized to promoter regions (−5 kb to +1 kb from transcription start sites, TSSs) in T cells and neurons, respectively. Other than promoter regions, 28% and 25% of the REST peaks were found within intragenic regions and another 14% and 12% were within 50 kb of genes (Fig. 1C). In total, 93% and 78% of the T-cell and neuronal peaks were assigned to one or more genes, respectively. Furthermore, functional analysis of the REST-bound genes using GREAT [39] revealed a dramatic disparity in the enriched pathways between these two cell types. One of the pathways defined by Pathway Commons [40] that was enriched in T cells showed clear involvement in neuronal function: neuronal system (p = 1.9E-5) (Fig. 1D), the other top enriched pathways, however, were related to functions important for lymphocytic cells. We found different categories of REST targets in neurons; the top four enriched pathways were involved in general gene expression (p = 1.1E-38) (Fig. 1D). The primary known REST function is to repress neuronal genes in non-neuronal cells [41]–[43], our data agreed with this, as this function was only found to be enriched in the T-cell REST targets. Only 487 (11%) of the T cell peaks overlapped with those from neurons, and the majority (n = 259; 53%) of these contained either a cRE1 or ncRE1 motif. Notably, 655 (73%) of the 894 genes targeted in both cell types had at least one peak that was detected in only one of the two cell types. Those common REST targets were enriched for neuronal functions, so were the T cell only REST targets, but not the neuron-only REST targets. These results suggest that REST cistromes of non-neuronal and neuronal systems may share limited overlap.
To extend our observation of distinct REST occupancy between neurons and non-neuronal cells, and also to gain insight into the dynamics of REST cistromes across human cell types, we decided to explore more publicly available data from the ENCODE project [34] and expand our comparison to include fourteen additional human cell lines: alveolar adenocarcinoma cells (A549), endometrial carcinoma cells (ECC1) lymphoblastoid cells (GM12878), embryonic stem cells (H1), colon carcinoma cells (HCT-116), cervical adenocarcinoma cells (HeLa S3), liver hepatocellular carcinoma cells (Hep G2), promyelocytic leukemia cells (HL-60), erythroleukemia cells (K562), breast adenocarcinoma cells (MCF-7), pancreatic carcinoma cells (PANC-1), primitive neuroectodermal tumor cells (PFSK-1), neuroblastoma cells (SK-N-SH), and glioblastoma cells (U87) (Table 1 and Table S1). These cell types represent a number of lineages. The SK-N-SH cell line is particularly interesting as it allows us to perform a comparison between normal neurons and tumorgenetic neuroblastoma cells, which have been reported to be associated with increased REST expression [44]. The resulting peak numbers from our ChIP-seq analysis pipeline are shown in Table 1, and range from 2,048 (in U87) to 8,199 (in H1 ESCs). (See Table S2 for list of peaks).
We next evaluated REST binding sites for cell specificity. Based on the above described difference in REST binding sites between T cells and neurons and the expectation of REST repression of neuronal genes in non-neuronal cells, we decided to compare REST cistromes across all 15 non-neuronal cells first, and then brought in neurons for a final comparison. Noted that we considered those tumor cell lines derived from neural tissues (e.g., SK-N-SH) as “non-neuronal” in this report. After overlapping peaks were merged, we obtained a set of 16,913 non-redundant REST binding regions from the total of 61,801 peaks in the 15 non-neuronal cell types. Analysis of the ChIP-seq signals showed different levels of REST enrichment across these peaks among the 15 cell types (Fig. 2A), and to our surprise, only 1,116 (7%) of the merged REST peaks were consistently called by SPP in all these cell types (referred to as “common” peaks). Nevertheless, 7% is much more than expected by chance, since we obtained 0 in common when we randomly picked genomic regions, with total number and size distribution matching to those of the REST peaks in individual cell lines, and performed the same merging procedure. A similar small fraction of REST peaks were found to be common in a previous analysis of REST bindings in 1% of the human genome [33]. Interestingly, these common peaks indeed exhibited the greatest enrichment of REST ChIP-seq signals in all cell types (Fig. 2A/B).
To study cellular specificity of REST occupancy more robustly, we compared several quantitative metrics for evaluating REST ChIP-seq signals at individual peaks for differential binding across cell types (see Supplementary Methods for more details). In brief, we analyzed the number of ChIP-seq reads at the summits of the non-redundant peaks using the program seqMiner [45] and then computed Z-scores to detect cell types with significantly more ChIP-seq reads than the rest. The results showed that ChIP-seq signals for 2,690 (16%) of those non-neuronal REST peaks were significantly stronger in one cell type than in any other cell type; these were termed “cell-specific” peaks (Fig. 2A/D and Table 1). Nevertheless, a large fraction (77%) of non-neuronal REST binding sites were shared by at least two or more cell types (referred to as “shared” peaks), as illustrated in Fig. 2A. In a comparison of the common, shared (i.e., 2–14 cell types) and cell-specific peaks, we found that common peaks were more enriched with the cRE1 motif (86%) than non-common ones (53%). The GGAAA/TA motif identified in neurons was not particularly enriched in any of the other cell types, as it occurred in 5.4% of the combined non-neuronal REST-binding sites (binomial test, p = 1.7E-133). Although 78% (n = 4,240) of the neuron peaks were called only for this cell type by SPP, 22% of them exhibited significant ChIP-seq signals in other cell types as well, thus resulting in 62% of neuronal REST peaks specific to neurons by our definition. This change indicates that comparison of transcription factor occupancy across samples by simple intersection of the genomic coordinates of the ChIP-seq peaks could exaggerate the true difference significantly. Interestingly, even for the neuronal REST peaks overlapping with peaks in other cell types, the peak summits were often shifted slightly to a new position in neuronal chromatin. While the average distance between the summits of overlapping peaks found in pairs of non-neuronal cells ranged from 6 bp (GM12878 vs. Hep G2, MCF-7 vs. HL-60, and MCF-7 vs. K562) to 24 bp (A549 vs. T cell), the mean distance of REST peak summits between neurons and other cell types ranged from 26 bp (vs HCT-116) to 79 bp (vs T cell) (Fig. S3). This observation again reveals the distinction of REST occupancy in neuronal cells.
To address if chromatin factors may contribute to cell-specific REST binding, we analyzed available DNase-seq data from the ENCODE project and related them to REST binding in A549, GM12878, Hep G2, H1 ES, K562, MCF-7 and T cells. We had expected that chromatin regions bound by REST would have greater DNAse-seq signals than “potential” but unbound REST candidate sites (i.e, REST-bound in other cell lines). While this was true in four of the cell types (A549, K562, MCF7 and T cells), with REST-bound regions showing 1.5–3.0× more overlapping with DNAse-seq peaks, no difference was observed for GM12878, Hep G2 and H1 cell lines. Nevertheless, we found that DNase-seq signals (measured by read densities) at the sites with stronger REST occupancy were generally higher than sites with weaker REST occupancy in all of these seven cells (data not shown). On the other hand, DNA-seq signals at many unbound REST candidate sites still showed much greater DNase-seq read enrichment in comparison to adjacent genomic regions. Taken together, these results indicate that chromatin accessibility is not the critical factor determining REST occupancy and thus the dynamics of REST cistromes. This observation is consistent with previous finding that neither DNase hypersensitivity nor chromatin features were a good predictor of REST binding [46].
Next, we compared the genes and miRNAs [37], [38] that were bound and thus potentially regulated by REST. Similar to REST peaks, the numbers of REST targets varied from one cell type to another (Table 1), with a total of 10,286 genes and miRNAs bound in at least one of the 16 cell types. There was approximately a 3-fold difference between the cell type with the highest (H1 ESCs, 4,509) and the one with the lowest (U87 cells, 1,682) number of REST targets. Five pathways: neuronal system, GPCR ligand-binding, potassium channels, transmission across chemical synapses, voltage-gated potassium channels were identified as significantly enriched in the REST targets for >10 cell types (Table S3). All of these pathways are important for neuronal function. Interestingly, pathways involved in translation (e.g., peptide chain elongation) were significantly enriched in REST targets in A549, HL-60, PFSK-1, and SK-N-SH cells, along with neuronal pathways, and in REST targets in neurons, but to the exclusion of top neuronal pathways (Table S4). Notably, REST-bound genes in these pathways were predominantly the same set of targets shared by these cell types. In addition, nearly all of the genes (n = 856) targeted by REST in all 15 non-neuronal cells contained a common REST peak (i.e., called in all cells). Among these common genes, only 1.2% (n = 10) were bound by REST at different genomic sites in any of the 15 non-neuronal cell types and 41% (n = 353) were also bound by REST in neurons. Interestingly, one of those genes targeted by REST in all cell types except neurons was REST itself, for which negative auto-regulatory feedback has been proposed [24].
REST was found to bind proximally to many of the well-characterized neuronal genes in various brain cancer cell lines: PFSK-1, SK-N-SH, U87 and in H1-derived neurons, a phenomenon previously reported [41]–[43]. We compiled a list of 15 known REST target genes from the literature, including BDNF [43], [47], CALB1 [48], L1CAM [49], CHAT [50], GRIA2 [51], CHRM4 [52], NRCAM [53], GRIN1 [54], STMN2 [52], SCG2 [55], SYN1 [52], SYP [56], SYT4 [48], GLRA1 [52], CHRNB2 [52]. All of these genes were REST-bound in 14 or more cell types. Among these, 6 of them (GLRA1, GRIA2, SCG2, CALB1, STMN2, CHRNB2) were bound in all 16 cell types, including neurons. The remaining nine genes displayed variable binding, with any lack of REST occupancy occurring in only four cell types: neurons, HCT-116 cells, PANC-1 cells, or Hep G2 cells. This result indicates that our observation of neurons as an “outlier” is unlikely due to some experimental technical biases (e.g., off target immunoprecipitation) in ChIP-seq analysis. Instead, it suggests that the prevalent view of REST having similar functions in non-neuronal systems through the repression of neuronal genes may have arisen from a systematic experimental bias that the same small set of genes has been examined repeatedly in previous studies. In addition, of the 52 genes that were upregulated upon REST knockdown in HEK293 cells [57], 54% (n = 28) of them showed variable REST binding among the 16 cells. 11 of these genes exhibited differential REST occupancy in neurons, with 10% (n = 5) of them bound by REST exclusively in neurons and 12% (n = 6) bound in other cell types, but not in neurons. This result suggests that a large fraction of genes repressed by REST are direct REST targets in non-neuronal cells.
As REST has been previously implicated in a variety of cancers, we decided to look into whether there were any cancer specific REST targets. A comparison of the REST occupancy in differentiated cell types (T cells and neurons) with the 13 cancer-derived cell lines revealed that several tumor suppressor genes (OSMR, MYO1A, THRB, FRMD3, LOXL4, CEACAM3, TRH) were bound by REST in all of the cancer cell lines, but in neither neurons nor T cells. Conversely, IL-7 receptor (IL7R) and LAMA2, two genes that are upregulated in a number of cancers [58], [59], were targeted by REST only in the two non-tumorigenic differentiated cell types. Notably, in H1 ESCs the REST binding patterns at all of these genes (except LAMA2) matched with those in the tumor cell lines, suggesting that REST regulation of these genes may have a role in cell proliferation, since active growth is a common feature of ESCs and tumor cells.
We next sought to better understand the transcriptional effect of REST occupancy, by integrating REST cistromics data and transcriptomics data. We utilized RNA-seq data (Table S5) and the TopHat/Cufflinks software suite [60] to determine gene expression levels in all of the cell types (Fig. S4). Hierarchical clustering analysis of gene expression showed that cell types of similar lineages and functions were grouped together, affirming the quality of our RNA-seq data (Fig. S4). We took the combined REST targets (n = 10,286) in all 16 cells and then for a particular cell type we compared the transcription of the bound (b) to the unbound REST candidates (ubrc = 10,286-b; approximate for potential REST targets) as well as to the transcription of all genes. For the majority (13) of the 16 cell types, REST bound genes were transcribed at a level significantly lower (2.4–36 fold lower; p<2E-16) than the unbound REST candidates (Fig. 3A), congruent with REST's primary role as a repressor [9]. In addition, in the majority (11) of the cell types lower expression of REST-bound genes compared to all genes was observed (significant in 8; p<2E-5). In A549 cells, T cells and neurons, REST-bound targets exhibited a significantly higher expression (Fig. 3A and Table S6). The predominantly repressive function of REST was further supported by very low expression of the 15 known REST targets described above (medians of FPKMs were 0.02–1.1 in all non-neuronal cell types but 16.2 in neurons). We were not surprised to find that REST targets were highly expressed in neurons as it has been reported that REST could bind and activate neuronal genes [9]. We were, however, surprised to see that REST-associated genes also exhibited greater expression than unbound REST candidate targets in A549 cells and T cells. Out of the top 200 most expressed genes in individual cell types, 37%–59% were REST-bound in A549 cells, neurons or HL-60 cells, while <23% were bound by REST in all other cell types. Interestingly, neurons, T cells and A549 cells had high proportions of REST peaks located to promoters and a higher percentage of REST peaks without an RE1 motif than the other cell types, the significance of which needs further investigation.
Current data also allows us to make an interesting comparison between the SK-N-SH neuroblastoma cell line and neurons. It has been reported that elevated expression of REST is associated with neuroblastomas [44]. RNA-seq data showed that REST was indeed more highly expressed in the SK-N-SH line (6.3 FPKM) than neurons (2.7 FPKM). Relatively few REST peaks (n = 386) overlapped in these two cell types. We found that a variety of neuronal pathways were enriched among the genes that were expressed at a lower level in SK-N-SH and bound by REST in SK-N-SH cells but not in neurons, while many genes involved with translation and RNA processing were uniquely bound and more highly expressed in neurons. As expected, genes in neuronal system pathways had higher expression in neurons than in SK-N-SH cells (median FPKM of 5.6 vs. 3.0), but even those bound by REST only in neurons (n = 17) were highly expressed (median FPKM 34.7 vs. 0.3 FKPM for those bound in SK-N-SH cells only). Interestingly, the tumor suppressors TRH and MTSS1 were not expressed in SK-N-SH cells (0 and 0.27 FPKM, respectively) but were in neurons (3.7 and 20 FPKM, respectively). In contrast, the oncogene β-catenin and proto-oncogene IL-7 were much more highly expressed in SK-N-SH cells (99 and 4.7 FPKM, respectively) than neurons (61 and 0.4 FPKM, respectively). These data support the view that REST may have a direct functional role in cancer progression by regulating oncogenes and tumor suppressors and the point that REST does not always repress its targets.
Previous studies show that REST binding does not always lead to gene repression and that in some cases it is conversely correlated with gene activation. In addition, it has been shown that the sequence bound by transcription factors can determine cofactor specificity [61]–[64] for a number of proteins such as for the transcription factor NF-κB [61], hormone activated estrogen receptors [62], and the glucocorticoid receptor [63], [64]; thus, we wondered if the context of REST binding plays a role in gene regulation.
By analyzing the REST peak numbers, we found that genes with more REST peaks generally exhibited lower levels of expression than those with a single peak. This observation persisted in all 16 of the cell types and was statistically significant in 15 (Fig. 3B shows data for GM12878, Hep G2, T cells and neurons; data for all cell types in Table S6; about 1.5–68 fold lower). This finding suggests that there may be an additive dosage effect of REST occupancy on the repression of its targets.
We also found that genes with REST peaks containing RE1 motifs (either cRE1 or ncRE1) generally exhibited lower expression levels than those without RE1 motifs or with half-sites, consistently across all cell types (Fig. 3C, Fig. S5, and Table S6; about 1.4–164 fold lower). Intriguingly, despite the low occurrence of RE1 motifs in neuron peaks, this trends held. However, even the REST bound genes with RE1 motifs had a higher expression level in neurons than in any other cell types (median of 5.6 FPKM in neurons vs. <1.3 FPKM in others, Table S6). Interestingly, genes with ncRE1-peaks tended to exhibit even lower levels of expression than those with cRE1, in agreement with a previous report [13]. This also persisted in 14 of the cell types and was statistically significant in 10 (Fig. 3C, Fig. S5, and Table S6). Notably, REST genes with RE1 motifs had lower expression than those without RE1 motifs and overall stronger peaks were associated with lower gene expression, except in neurons (data not shown).
Next, we examined whether genes consistently bound by REST in all cell types were as lowly expressed as genes that were bound variably (shared or cell-specific). Indeed, we found that genes with a common REST binding site exhibited lower levels of expression than those with shared or cell-specific sites (Fig. 3D and Table S6; about 1.7–195 fold lower). Interestingly, the highest levels of expression for these common genes occurred in neurons exclusively (median expression of 4.4 FPKM).
Finally, we compared the expression of different groups of REST targets that were separated based on the relative locations of REST binding. In 11 of the cell lines, genes bound by REST in their bodies (introns or exons) exhibited significantly lower expression levels than those with REST in their promoters (data for all cell types in Table S6; about 1.4–69 fold lower). This is quite interesting, since the effect of REST on gene expression has been mostly studied through its binding at promoter regions. On the other hand, it has been reported that REST binding within 50 bp of the TATA box in neuronal cells (but not in non-neuronal cells) was correlated with gene activation [18]. Indeed, those genes bound by REST at their promoters with peak summits located <50 bp from the TSS exhibited higher expression levels (Table S6; 1.8–5.4 fold higher) than even unbound REST candidate genes. This difference was found in 9 of the cell types, and was statistically significant in 7. This finding is probably related to the emerging view that transcription factor binding at enhancers has a greater effect on gene expression than binding at promoters, where many factors likely act competitively or coordinately.
As previously mentioned, data from A549 cells, T cells and neurons did not indicate repression of REST targets. No dosage-associated expression difference was detected for REST binding in neurons. Nevertheless the presence of RE1 motifs, cell specificity, and location of REST peaks still made a difference in terms of the extent of REST repression (Table S6). These three cell types had the largest percentages of REST peaks in the promoter regions (34.6%, 50.6%, and 41.2% for A549 cells, T cells, and neurons, respectively, in comparison with 14.2–28.1% of peaks in other cell types) and the smallest percentages of their peaks contained a cRE1 or ncRE1 motif (48.2%, 40.6%, and 8.5% A549 cells, T cells, and neurons, respectively, in comparison with 51.3%–83.8% in other cell types). Both of these two features showed a strong bias towards higher expression of the REST targets.
Since these aforementioned factors were not independent, we performed a logistic regression to determine the individual contributions of these factors to predict the transcription outcome of a REST target as expressed (FPKM≥1) or not expressed (FPKM<1). The results indicated that genes next to RE1 motif peaks, common REST peaks, and intragenic/distal peaks were 2.6, 1.9 and 1.3 times more likely to not be expressed than genes next to RE1-free peaks, non-common REST peaks, and promoter peaks, respectively, suggesting that motif status and cellular context were the two primary factors.
The above analyses show that context plays a role in REST regulation of its targets, which brings up a question as to whether this is due to different sets of REST cofactors being recruited. To address this, we investigated the co-occupancy of REST with CoREST, SIN3, and EZH2 (a core component of PRC2) using the ENCODE project ChIP-seq data from the two cell types: GM12878 and Hep G2 [34]. There were 17,590 SIN3, 44,065 CoREST, and 64,277 EZH2 peaks in GM12878. The corresponding numbers in Hep G2 cells were 32,019 SIN3, 51,883 CoREST and 79,275 EZH2 peaks. Note that we merged the two sets of SIN3 peaks in Hep G2. First, we found that 14%, 29% and 43% of the GM12878 REST peaks overlapped with SIN3, CoREST and EZH2, respectively. Interestingly, a previous analysis also showed that RE1 motifs were highly enriched in the SIN3A-occupied genomic sites in H1ESCs [65]. Further analysis showed that only a small fraction (7%) of REST-bound regions had all three cofactors, while a large percent (42%) were not occupied by any of these cofactors at all (Fig. 4A). In addition, most (85%) of the SIN3-REST co-binding regions were localized to sites bound by CoREST, and a large fraction (36%) of the REST-EZH2 sites were co-occupied by CoREST. Next we determined the enrichment of the contexts in each of the groups with different combinations of REST and its cofactors. We found that REST-SIN3 peaks were 2.7 times more likely than expected (in relation to all REST peaks) to be in promoters, 1.7 times more likely to contain either a half RE1 motif or no RE1 motifs. It should be noted that the TSS proximity of SIN3 has been reported previously [66]. REST-CoREST peaks were 1.5 times more likely than expected to be REST-bound in all 15 cell types (i.e. common REST peaks). REST-EZH2 peaks did not show an increased association with any of the examined features. Cofactor-free REST sites showed a depletion of both promoter and common REST binding (Fig. 4B). These observations were generally reproduced with data from the Hep G2 cell line (Fig. S6).
We then asked how variable cofactor association played into the transcriptional regulation on REST targets. A previous study [15] has shown that in mouse ESCs, REST-binding sites colocalized with SIN3 or CoREST occurred more frequently in genes whose transcription was directly repressed by REST, as measured by their upregulation upon shRNA knockdown of REST expression. Since we did not have REST knockdown data in any of the cell types studied here, we contrasted REST targets that were bound by cofactors to all REST targets. In contrast to the previous finding [15], we found that REST-bound sites with SIN3 and CoREST were located to genes exhibiting 28 and 4 times higher transcription, respectively, when compared to all REST targets in GM12878 (Fig. 4C; data for Hep G2 in Table S7). This was due to the strong positive association between SIN3 occupancy and active transcription, since REST-SIN3 only targets were expressed 47 times higher than REST targets (Table S8) and SIN3 was the transcription factor most associated with active expression among the 3 cofactors assessed. This result is surprising because SIN3 interacts with HDACs and reduction of histone acetylation is often associated with gene repression, although a previous study has reported that HDACs were localized to many active genes [67]. Moreover, REST sites colocalized with SIN3, CoREST and EZH2 were actually found at genes (e.g., CHRNB2) that were expressed 26 times more highly than the median of all REST targets. REST sites colocalized with EZH2 and CoREST, either alone or together, were associated with genes (e.g., DRD3, HTR3A, and BDNF, respectively) expressed at a level of 2.0, 3.6, and 1.9 times lower, respectively, in relation to all REST targets. REST targets without any of the three cofactors were expressed, unexpectedly, at 2.5 times lower levels than all targets in GM12878. The expression difference associated with differential cofactor binding generally held true by examining either all REST peaks or only promoter peaks. In addition, similar results were obtained with data from Hep G2 cells (Table S7), except that in Hep G2 the REST targets not exhibiting colocalization with any of the three cofactors were actually expressed at levels 3.5 times higher than all REST-bound genes. In Hep G2 cells, the REST sites that colocalized only with CoREST were also more highly expressed. The opposite trends observed for REST-alone targets and the REST-CoREST “only” targets between GM12878 and Hep G2 cells suggest that a different set of other REST cofactors may have been recruited to those targets, also in a cell-specific manner. It would be interesting to study in the future if G9A, CTBP, MECP2, LSD1 or other yet-to-be-identified REST interactors are involved. Notably, we did not observe pathways specifically enriched in any group of REST targets with different cofactor occupancies. In summary, our results indicate that SIN3 co-localization was correlated with higher expression while EZH2/CoREST co-occupancy was associated with lower levels of expression of REST targets, with SIN3 seemingly dominant over EZH2/CoREST. In the future, it would be interesting to test experimentally if SIN3 and EZH2/CoREST indeed confer opposite regulatory roles in some REST targets by knocking down these chromatin factors.
The fact that the outcome of REST regulation is largely dependent on its cofactors is not entirely surprising, but it reinforces the view that REST is a molecular platform for recruiting chromatin modifiers, which ultimately determine the transcription activity. To address this computationally, we combined the cofactor colocalization information with histone modification data in GM12878 cells from a previous study [68]. We observed that 82% of the REST-EZH2-only, 74% of the REST-CoREST-only, and 19% of REST-SIN3-only sites were located to regions enriched with H3K27me3 (a repressive histone mark). On the other hand, 74% and 48% of the REST-SIN3-only, 6% and 2% of REST-CoREST-only and 12% and 3% of REST-EZH2-only sites were located to either H3K4me3 (an active histone mark) or H3K27ac (an active histone acetylation mark) enriched regions, respectively. These enrichments are congruent with the known addition of H3K27me3 by EZH2 and removal of H3K4me by LSD1/CoREST. SIN3 acts on the chromatin through the recruitment of HDACs. Since there was no HDAC ChIP-seq data in GM12878 cells, we only studied the colocalization of SIN3 and HDAC2 in Hep G2 cells. A total of 32,895 HDAC2 peaks were called for Hep G2. We found that 69% of the REST sites that colocalized with HDAC2 (n = 374) were also enriched with SIN3 (binomial test, p = 7.5E-129), supporting SIN3-HDAC interaction at REST-bound sites. All together, our study demonstrates that distinct combinations of chromatin modifying cofactors are recruited to different REST-binding regions, and that they likely contribute to the transcriptional outcome of REST regulation. Whether and how these cofactors work together with REST to activate or repress gene expression cannot be directly addressed here, due to the limitation of computational work, and thus requires more study in the future.
Throughout our analysis the H1-derived neurons stood out from the other cell types. For instance, neuronal REST peaks were enriched with a different motif and REST targets in neurons were overall highly transcribed. This raises the question of whether different cofactors are recruited to the REST-bound sites in neurons. Unfortunately, TAF1 and RNA polymerase II were the only DNA-binding proteins that had been analyzed by ChIP-seq in the same neuron samples, both of which are unlikely to play a deterministic role in REST-specific gene regulation. We have, however, analyzed several available genomic and epigenomic data sets in order to gain a glimpse at the neuron-specific REST function.
As it has been suggested that a REST isoform (REST4) could inhibit REST function and activate REST targets [9] and that REST4 expression is neuron-specific [44], we examined whether this isoform was more abundant in neurons. Our analysis of RNA-seq reads mapped to REST4 specific exon, however, did not find evidence that REST4 was the dominant REST isoform in neurons. In fact, in comparison to the full length REST (2.7 FPKM), REST4 transcription was 7-fold lower at 0.39 FPKM (Fig. S4). Next, we asked if the presence of small RNAs within REST-binding sites might have altered REST-mediated gene regulation, since it has been shown that the transcription of enhancer RNAs (eRNAs) is often correlated with gene activation [69], [70] and that a double stranded small RNA was shown to activate REST targets [71]. We utilized small RNA sequencing data from the ENCODE project for H1-derived neurons, GM12878, and H1 ESCs [72], and determined small RNA read abundance at individual REST peaks using the algorithm HTSeq [73]. Interestingly, we found in neurons that (i) 804 of the REST sites (15%) had small RNAs mapped to them (≥1 reads per kb per million small RNA-seq reads (RPKM); binomial test, p<2.2E-16), and (ii) genes associated with these REST peaks (n = 761) were expressed at significantly higher levels (median FPKM = 15.0) than all REST bound genes (p = 2.8E-39). In contrast, very few genes were associated with REST peaks that could potentially produce eRNAs by the same criterion in both GM12878 (n = 33) and H1 ESC (n = 69). This difference remained when the threshold for small RNA presence was set to 0.1 RPKM (data not shown).
Next, we examined the chromatin modifications at REST binding sites. We obtained chromatin regions that were determined to be enriched with H3K27me3, H3K36me3, H3K4me1, H3K4me3, and H3K9me3 from a previous report, in which these modifications in H9-derived neurons, GM12878 cells, H1 ESCs, and many other tissues were studied [68]. Intersecting those histone modification regions with the REST peaks in neurons, GM12878, and H1 ESCs demonstrated that a much greater percentage (49%) of neuron REST peaks overlapped with H3K4me3, an active histone modification, than did GM12878 (20%) or H1 ESC (18%) REST peaks. This finding was further supported by the high enrichment of H3K4me3 ChIP-seq signal at the center of REST peaks in neurons but not in GM12878 cells, H1 ESCs, A549, HeLa S3, Hep G2, or K562 cells (Fig. 5A). The overlaps of REST peaks with chromatin regions enriched with two additional active histone marks: H3K4me1 and H3K36me3, were also 2× higher in neurons than in either GM12878 or H1 ESCs (Table 2). Conversely, a significantly greater percentage (69%) of GM12878 REST peaks overlapped with H3K27me3 regions, a repressive histone modification, than did neuron REST peaks (27%). This finding was further supported by the enrichment of H3K27me3 ChIP-seq signal at the center of REST peaks in GM12878, Hep G2 and other cell lines, but not in H1-derived neurons (Fig. 5B). Although only 18% of the REST peaks in ESCs overlapped with H3K27me3 regions (Table 2), H3K27me3 ChIP-seq signals showed significant enrichment at REST peaks in ESCs (Fig. 5B), which was upon further examination due to strong H3K27me3 signals in these 18% overlapping peaks (no enriched profile detected after their removal, data not shown). Similarly, a higher overlap of REST peaks with H3K9me3 regions was observed in GM12878 cells compared to either neurons or ESCs. It is possible that lower cofactor colocalization at these neuron REST sites, could explain the higher levels of H3K4me1/3 and lower levels of H3K27me3 and H3K9me3. Likewise, for the 15 known REST target genes, we found much higher association of their REST sites with repressive histone marks outside neuronal cells, 14 of the 15 genes with H3K27me3, and 4 of the 15 with H3K9me3 in GM12878 compared with 6 of the 10 genes with H3K27me3 and none with H3K9me3 in neurons. These differential histone modification colocalizations go hand in hand with the higher expression of these genes in neurons based on RNA-seq data. This finding underscores the hypothesis that distinct sets of REST cofactors are recruited to REST-bound regions in a context-dependent manner, as these 15 genes are occupied by REST across cell types but exhibit differential enrichment of histone marks.
We next examined how REST differentially regulated a mini regulatory circuitry that has previously been suggested to be important for neurogenesis. The circuitry includes REST cofactors, several miRNAs, and neurogenic factors critical for neuronal development (Fig. 6). Among the known components of the REST complex [9], [11], [12], [48], LSD1, BRG1, HDAC4, and HSPC1 (a component of PRC1) were bound by REST in 5 or more cells, frequently including ECC-1, H1 ESCs, HL-60, PANC-1 and SK-N-SH cells. EZH2 showed REST binding at its promoter in neurons but not other cells (see Table S2).
Finally, we examined the expression of and REST-binding at several miRNAs that have been reported to control neural development [74]–[77]. The expression levels of miRNAs were determined from small RNA-seq datasets. Of the 939 miRNAs annotated in miRBase [38] 134 of them (14%) were bound by REST in at least one of the 16 cell types and 39 of them (4.2%) were differentially bound in neurons and non-neuronal cell types. 32 of the 39 REST-bound miRNAs in neurons (82%) were bound either only in neurons (n = 20) or at a different genomic position in the non-neuronal cells (n = 12). Five miRNAs have been extensively studied for their interaction with REST and their roles in promoting neurogenesis: miR-9 [9], [78], [79], miR-124 [9], [79], miR-132 [9], [79], [80], miR-135b [78] and miR-212 [80], many of which directly target neuronal genes. Interestingly, each of these miRNAs showed a different pattern of REST binding in neurons and non-neuronal cells (Fig. S7), with miRs 124-3, and 132/212 targeted by REST in all cell types but neurons, with miRs 9-3, 124-1, 124-2, and 135b targeted at a position in neurons different from other cells, and miR-9-2 bound only in neurons. Compared to their expression in H1 ESCs and GM12878 cells, all of these miRNAs were more highly expressed in neurons (Table 3), suggesting that REST may play a role in activating instead of repressing these miRNAs in neurons. In support of this hypothesis, the expression of miR-9 was indeed downregulated in mouse neuronal stem cells upon conditional knockdown of REST expression [81].
In this study, we have characterized genome-wide REST occupancy across multiple cell types and related that to other transcription factor binding and transcriptional outcome. With REST ChIP-seq data from 16 different human cells, we made an attempt to estimate the number of potential in vivo REST binding sites in the human genome. Although we called a total of ∼62,000 REST peaks, they were highly redundant. 63% of the total bindings in the 15 non-neuronal cell types could be identified using data from just three cell types with the most peaks (Fig. S8). Starting from the cell type (H1 ESC) with the most peaks, we computed the number of new REST peaks that were identified by including data from additional cells. As shown in Fig. S8, the number of additional peaks added with each cell line decreases and approaches a plateau at about 7 cell types. The extrapolation of our data indicates that there are ∼18,000 potential REST-bound regions in the human genome. This estimation, however, excludes neurons, because our analysis shows that REST binding in neurons is quite distinct from non-neuronal cell types and much is yet to be learnt for the diversity of REST bindings among neuronal subtypes [7], [30]. We should note that of the genomic sites previously predicted to contain cRE1 motifs [13], almost all (234/235) of the most highly confident ones were included in our REST peaks and additionally 85% (669/783) of those lower confidence peaks in non-repetitive regions were also covered by our REST peaks.
Our findings indicate that the REST cistrome in differentiated human neurons is dramatically different from those in non-neuronal cells. This difference, as well as differences in the dynamics of REST targeting, was also observed in previous studies of REST occupancy at promoters during the development of a variety of mouse neural lineages using microarrays [27], [30], [31]. The limitation of the current study is that the samples used for ChIP assays contained a mixture of neuronal subtypes, which may have differences in REST occupancy. The maturation status of the cultured neurons could be another important factor. From that perspective, the differences between the neuronal REST cistrome and that of non-neuronal cells may actually be larger than what we report here, since the limited overlapping could be due to the presence of a small number of neuronal progenitor cells in the ChIP samples. We should mention that RNA-seq data for our analysis of gene expression in neurons were not derived from the same sample used for ChIP-seq experiment: H1 ESC derived neurons for ChIP and iPSC derived neurons for RNA-seq. The differentiation protocols, however, were essentially the same and promoted the generation of GABAergic and Glutamatergic neurons, and both samples were harvested after four weeks of culture. The enrichment of neurons in our cell culture was supported by the high expression of neuronal and synapses genes [82]: (TUBB3, 431 FPKM; MAP2, 141 FPKM; MAPT, 86 FPKM; SYN1,16 FPKM; DLG4, 41 FPKM), low expression of genes marking astrocyte and oligodendrocytic glia [82] (GFAP, 6 FPKM; MBP, 0.5 FPKM; OLIG2, 0.7 FPKM), as well as low expression of NSC/NPC markers [83]. We should point out that REST expression in Glutamatergic and GABAergic neurons has been demonstrated previously by co-staining of REST and neuronal specific marks (see Fig. S2 in ref [30]) and recently mRNA and protein expression in neurons in the prefrontal cortex has been reported [84]. In addition, 21% and 25% of REST promoter targets previously found in mouse GABAergic and Glutamatergic neurons [30] were also identified in our neuron ChIP-seq data, respectively. We do not think the sample “mismatch” undermines our finding that REST targets exhibited active expression in neurons but lower expression in non-neuronal cells, because results from our analysis of histone modification ChIP-seq data collected for another neuronal sample (derived from H9-ESC line) also support our conclusion. Nevertheless, it will be important to revisit this issue when the RNA-seq data for H1-derived neurons becomes available from the ENCODE project.
The lack of RE1 motif enrichment in the neuronal REST peaks is quite surprising and intriguing. Although motif analysis identified GGAAA/TA as a potential alternative binding sequence, we do not think it represents true REST recognition motif in neurons, because it is a rather common motif and present in 9% of randomly selected genomic sequences. Instead, a more likely scenario is that REST loses its direct interaction with neuronal chromatin and becomes a co-factor for other transcription factors. Our study showed that small RNAs were enriched in the neuronal REST peaks, but more direct experimental assays are needed to investigate if small RNAs or long non-coding RNAs have a role in altering REST interaction with chromatin and its targets. REST4 has been associated with active gene expression in neurons. Although we did not uncover evidence that REST4 was the dominant isoform in neurons, we did observe that REST4 transcript was increased 3-fold in the transition of day-14 to day-27 neurons (data not shown). Furthermore, our RNA-seq data indicated that SRRM4, shown to promote alternative splicing to generate REST4 transcript [85], was transcribed only in neurons, with 14.0 FPKM in neurons but <0.95 FPKM in all non-neuronal samples. This suggests that REST4 could be the dominant REST protein product in our neuron cultures. Nevertheless, REST4-bound chromatin sites would still be expected to be enriched with variants of RE1 motifs. Therefore, it remains a mystery how REST changes its DNA recognition specificity in neurons. Only with further study can we address if REST interaction with the neuronal genome is mediated by other factors, like in the case of recently reported cellular dependent chromatin interaction of TCF7L2 [86]. We need to point out that the lack of RE1 motif enrichment has also been reported in a previous analysis of REST regulation in mouse neurons [30]. We should also emphasize that two ChIP-seq replicates from the H1-derived neurons showed consistent REST binding and the same lab produced all the REST ChIP-seq data with the same antibody, except those from CD4+ T cells. Thus, we do not believe the unique features of neuronal REST cistrome is due to any experimental issues.
Our study sheds new insight into REST regulation of many genes that have critical roles in neuronal development and function (Fig. 6), including miR-9 and miR-124 (Fig. S7). These miRNAs contribute to neural differentiation, neural fate determination and cell cycle exit through the repression of a number of neural transcription factors including TLX, FOXG1, HES-1 (all repressed by miR-9), REST, and some of its cofactors (Fig. 6). TLX is critical for maintaining neural progenitor cells (NPCs) in their undifferentiated state [87]. HES-1 is required for NSC homeostasis/maintenance [88], as its repression accelerates, while its overexpression inhibits, neurogenesis [89]. FoxG1 maintains NPC self-renewal [90] and suppresses the formation of early-born neurons [91]. Our analysis showed that REST bound to HES1 and TLX in nearly all cell types but neurons, while REST occupied FOXG1 only in neurons. There is more than one locus encoding miR-9 and miR-124 in the human genome. The neuronal expression of miR-9 and miR-124 was likely promoted by the differential promoter REST binding in neurons, as well as certain loci not bound by REST altogether (Fig. S7). Together, these suggest that REST is involved in miR-9 and miR-124 transcription and consequently that their repression of neuronal factors that are critical for maintaining the fate of neural stem cells. Simultaneously, REST also directly represses those factors in non-neuronal cells. For two transcription factors that promote neuronal differentiation, NEUROD1 and BDNF, expression is also controlled by REST; both exhibit similar regulation by REST in all of our analyzed cell types except neurons (NEUROD1 lacks upstream REST binding in both neurons and T cells). Their expression in neurons, however, is also modulated by miR-124, likely to strike a balance between factors that activate and inhibit neurogenesis. Several other key neuronal factors are also regulated by the REST/miRNA regulatory circuitry (see Table S2), including BAF53a/b and EFNB1, as well as two (MYT1L and POU3F2) of the three genes sufficient to convert fibroblasts to neurons [92] and a transcription factor (NeuroD2) that speeds their miRNA-mediated conversion [93]. These results, along with previous findings in the literature, indicate that REST plays an important role in neurogenesis by both directly targeting key neuronal transcription factors and regulating the transcription of neuronal miRNAs. Together, the mini REST/miRNA regulatory network controls neurogenesis synergistically by fine-tuning the expression of individual components to maintain a balance, which is necessary for the proper development of multiple neuronal lineages and for maintaining some level of developmental plasticity.
We were intrigued to discover that REST bound to its own promoter in all 15 cell types but not in neurons. It may be that in neurons, the dynamics of REST production and degradation are different, for example, perhaps more post-translational control is taking place. It is known that REST is a protein with high turnover [10] and that negative auto-regulation speeds response times of transcriptional networks [94]. It could be that less steady-state transcription of REST leads to phases of lower levels of available REST despite the fact that REST expression levels are similar (ranging from 2.6–10.2 FPKM) in neurons and other cell types (Fig. S4). REST, itself, along with other members of the REST complex, such as SCP1, EZH2, CoREST (RCOR1) and MECP2, are all regulated by REST-bound and brain-specific miRNAs, including miR-9 and miR-124. Many of the REST complex components, miRNAs and REST targets are also regulated by CREB [95], a potential positive regulator of all the REST-regulated miRNAs. Therefore, all of these cofactors are controlled at a number of different levels through REST and its downstream miRNAs. It will be important to study how these interconnected regulatory factors are involved in the formation and function of the REST complexes, especially during neurogenesis, neuronal differentiation and maturation.
The protocols for ChIP-seq and ChIP-qPCR have been described [96]. In brief, human primary CD4+ T cells were purified from blood as previously described [97]. This cell type was chosen because it is an easily accessible primary cell type and known to express REST [3]. Twenty million cells were cross-linked by formaldehyde. Following sonication, chromatin fragments were immunoprecipitated with an anti-REST antibody [Millipore 07-579] and then prepared for sequencing. Sequencing was performed on an Illumina HiSeq 2000 machine. ChIP-seq reads were processed by the Illumina Analyzer Pipeline and aligned to the human genome (GRCh37/hg19) using Bowtie [98]. Unique reads mapped to a single genomic location (allowing up to three mismatches) were kept for peak identification. The primers used for qChIP analysis were listed in Table S9. The ENCODE ChIP-seq data (Table S1), including the one for H1-derived neurons, were collected with a REST antibody provided by Dr. David Anderson at Caltech. Sample information for the H1-derived neurons (including its purity) can be found at (http://genome.cse.ucsc.edu/ENCODE/protocols/cell/human/H1_Neurons_Round1.pdf).
Peaks were called using the SPP pipeline [35], following the guidelines of controlling Irreproducible Discover Rate (IDR) [99] by the ENCODE project [34], with a relatively strict IDR threshold (0.001). Where multiple ChIP-seq replicates were available, reads from all replicates were combined for peak calling. For T cell data, which did not have a replicate, we divided the ChIP-seq data into two halves to generate pseudoreplicates, and the IDR threshold was calibrated from analysis of the pooled-pseudoreplicated data of the other 15 cell types. We also filtered out REST peaks that mapped to genomic regions red-flagged by the ENCODE (both Duke mappability regions and ENCODE Dac mappability consensus regions) [34], as they were likely a result of experimental artifacts.
We used MEME 4.6.1 [36] to find enriched motifs within 200 bp sequences centered on the summits of all the REST peaks in the CD4+ T cell data. The resultant RE1 motifs were used to identify peaks with RE1 motifs by the program MAST in MEME suite, in all cases by scanning 200 bp sequences around peak summits and using default parameters.
The genes containing peaks were identified using in-house scripts. We used RefSeq [37] and miRBase [38] annotations from the UCSC Table Browser (http://genome.ucsc.edu), which provides the precursor forms of miRNAs; in cases of overlapping transcripts from the same gene we picked the longest one for both ChIP-seq and RNA-seq analysis, resulting in 23,010 genes and 928 miRNAs. The script assigned peaks to genes in the following step-wise manner: to promoter regions (−5 kb to +1 kb of TSSs (transcriptional start sites)), to exons, to introns, to distal regulatory regions (−50 kb of transcription starts to +50 kb of transcription ends). When mapping peaks to either promoter regions or distal regions, only the gene with the closest TSS was selected. A single base overlap was used as a rule for these assignments. A peak can be mapped to multiple genes, but only it is equidistant from the TSS of multiple genes, or if it is located to exons or introns shared by multiple genes. Notably, the definition of exonic and intronic REST binding was not applicable to miRNAs and their precursors, as they are short.
The merged and non-redundant REST peaks were compared to the peaks originally called for each cell type by the SPP program, and those overlapping with peaks from all cells were defined as “common peaks.” For the rest of the merged peaks, we computed a sequencing-depth-normalized maximal read coverage at the 300-bp surrounding the peak summit (averaged from all cells) of each peak in each cell type, and then inferred cell-specific REST binding based on significantly differential read coverage. Each ChIP-seq read was extended to 200 bp for this analysis. For each peak at each cell, we obtained a REST ChIP-enrichment score (Ej, j = 1,2, …, 16) that was determined from difference in maximal read coverage between REST-ChIP and input samples and normalized by a scaling factor that quantified the genome-wide background noise. The enrichment scores for all peaks were subject to quantile normalization across cell types before used for Z-score statistics. In the end, we defined cell-specific peaks as those having a high Z-score (>3) in one cell type but low Z-scores (<1) in the rest. Further information on this procedure and more details of the methodology development can be found in the Supplemental Methods (Text S1), where we also presented our exploration of several computational schemes to account for different chromatin structure, different immunoprecipitation efficiency/enrichment and other factors in ChIP-seq experiments of different cell types.
All RNA-seq data have been published (Table S5) except the one from neurons, which was collected from 27-day differentiating human neurons derived from induced pluripotent stem cells (iPSCs), which were made from a healthy male. The derivation of the iPSC line (iPSC-2), neuronal differentiation, and RNA-seq sample preparation have been described in previous publications [100], [101]. RNA-seq reads (from polyA+ RNAs) from replicates (when available) were merged and were subsequently aligned to the human genome (version hg19) by TopHat (v2) [60]. Transcripts from Refseq [37] (with micro- and sno-RNAs removed) were used to determine gene expression by the Cuffdiff tool in Cufflinks package (v2) [102], using options for correcting sequence bias and multiple hits, as well as the default geometric mean normalization. All expression comparisons were carried out by the Wilcoxon test and fold changes between groups were based on the medians of FPKMs, unless stated otherwise. Small RNA-seq data were downloaded from GSE24565 [72] and reads were aligned to the human genome by Bowtie (v2) [103], using the options: –local –sensitive-local –score-min G,0,2. We then counted the number of reads at individual REST peaks or miRNAs [38] using HTseq [73], which only uses uniquely mapped reads, with normalization by peak sizes and the number of total aligned reads to yield a RPKM (Reads Per Kb per Million mapped reads) value.
The ChIP-seq read density was calculated using the program seqMiner [45], which yielded an array that consisted of the maximal number of overlapping ChIP-seq reads (extended to 200-bp) in 300-bp bins from −3.15 kb to +3.15 kb of the REST peak summits. The enrichment of the sequencing-depth-normalized reads over those of input experiments for each cell type was calculated and the enrichment values were subject to both background normalization [104] as described above. This matrix of enrichment values were finally used to generate heatmaps in Figure 2 with the gplots package [105] in R.
All data are publicly available and they can be accessed in the Gene Expression Omnibus (see Table S1 and S5).
|
10.1371/journal.ppat.1002803 | A Cytotoxic Type III Secretion Effector of Vibrio parahaemolyticus Targets Vacuolar H+-ATPase Subunit c and Ruptures Host Cell Lysosomes | Vibrio parahaemolyticus is one of the human pathogenic vibrios. During the infection of mammalian cells, this pathogen exhibits cytotoxicity that is dependent on its type III secretion system (T3SS1). VepA, an effector protein secreted via the T3SS1, plays a major role in the T3SS1-dependent cytotoxicity of V. parahaemolyticus. However, the mechanism by which VepA is involved in T3SS1-dependent cytotoxicity is unknown. Here, we found that protein transfection of VepA into HeLa cells resulted in cell death, indicating that VepA alone is cytotoxic. The ectopic expression of VepA in yeast Saccharomyces cerevisiae interferes with yeast growth, indicating that VepA is also toxic in yeast. A yeast genome-wide screen identified the yeast gene VMA3 as essential for the growth inhibition of yeast by VepA. Although VMA3 encodes subunit c of the vacuolar H+-ATPase (V-ATPase), the toxicity of VepA was independent of the function of V-ATPases. In HeLa cells, knockdown of V-ATPase subunit c decreased VepA-mediated cytotoxicity. We also demonstrated that VepA interacted with V-ATPase subunit c, whereas a carboxyl-terminally truncated mutant of VepA (VepAΔC), which does not show toxicity, did not. During infection, lysosomal contents leaked into the cytosol, revealing that lysosomal membrane permeabilization occurred prior to cell lysis. In a cell-free system, VepA was sufficient to induce the release of cathepsin D from isolated lysosomes. Therefore, our data suggest that the bacterial effector VepA targets subunit c of V-ATPase and induces the rupture of host cell lysosomes and subsequent cell death.
| Vibrio parahaemolyticus is a bacterial pathogen that causes food-borne gastroenteritis and also wound infection and septicemia. It exhibits cytotoxicity that is dependent on its type III secretion system (T3SS1) during the infection of mammalian cells. Although an effector VepA plays a major role in the cytotoxicity, the mechanism was unknown. Here we show that VepA targets subunit c of the vacuolar H+-ATPase (V-ATPase) and induces the rupture of host cell lysosomes. We found that VepA alone is cytotoxic in HeLa cells and also toxic in yeast Saccharomyces cerevisiae. Using a yeast genome-wide screening, we identified yeast V-ATPase subunit c as essential for the toxicity of VepA to yeast. We also demonstrated that knockdown of V-ATPase subunit c decreased VepA-mediated cytotoxicity toward HeLa cells and that VepA interacted with subunit c of V-ATPase. During infection, lysosomal contents leaked into the cytosol prior to cell lysis, and VepA was necessary and sufficient for this leakage. Our data suggest that a bacterial effector VepA ruptures lysosomes, the “suicide bags” of host cells, by targeting the evolutionarily conserved V-ATPase, and elicits subsequent cell death.
| Vibrio parahaemolyticus, a Gram-negative marine bacterium, is a major food-borne pathogen that causes acute human gastroenteritis associated with the consumption of seafood. This pathogen also causes wound infections and septicemia in humans [1]–[3]. V. parahaemolyticus possesses virulence factors such as thermostable direct hemolysin (TDH) and two separate type III secretion systems (T3SSs), namely, T3SS1 and T3SS2 [4], [5]. T3SSs are protein export systems that enable bacteria to secrete and translocate proteins known as effectors into the cytoplasm of host cells. Translocated effectors modify host cell function and allow pathogens to promote infection and cause disease [6], [7]. V. parahaemolyticus T3SS1 is involved in the cytotoxicity to various mammalian cells, whereas T3SS2 is related to the enterotoxicity of this organism [8]–[10]. T3SS1-induced cell death occurs rapidly, within several hours after the inoculation of V. parahaemolyticus; is independent of caspases; and is associated with autophagy [11], [12]. The transcription of the T3SS1 genes of V. parahaemolyticus is regulated by a dual regulatory system consisting of the ExsACDE regulatory cascade and H-NS [13]. To date, VepA (VP1680/VopQ), VopS (VP1686) and VPA450 have been identified as T3SS1 effectors [14]–[18]. VopS functions as an AMPylator and contributes to cell rounding [17]. VPA450 acts as an inositol phosphatase and induces plasma membrane blebbing [18]. It has previously been shown that the deletion of vopS or vpa450 has little effect on T3SS1-dependent acute cytotoxicity, but a mutant strain of V. parahaemolyticus in which vepA was deleted (ΔvepA) lost its cytotoxicity, suggesting that VepA plays a major role in T3SS1-dependent cytotoxicity [14]–[16].
VepA is a 50-kDa protein and consists of 492 amino acids. Although VepA has no homology to known proteins, the amino-terminal 100 amino acids of VepA are required for secretion by T3SS1 [15]. VepA has also been reported to be involved in the activation of autophagy and mitogen-activated protein kinases (MAPKs) [16], [19]. However, the mechanism by which VepA is involved in acute cytotoxicity in host cells is unclear. To understand the mechanism underlying the T3SS1-dependent cytotoxicity of V. parahaemolyticus, elucidation of the function of VepA within host cells is important because VepA appears to play a critical role in cytotoxicity. In this study, we showed that VepA itself is a cytotoxic effector, and we screened for host factors essential for the cytotoxicity of VepA using yeast genome-wide analysis to elucidate the function of VepA.
To understand the function of VepA, we first examined the expression of green fluorescent protein (GFP) in 293T cells transfected with pEGFP-VepA. However, GFP-VepA expression was not detected (Figure 1A), raising the possibility that VepA itself is quite toxic in cells. To evaluate whether VepA itself is cytotoxic, we introduced purified VepA into HeLa cells by protein transfection using the hemagglutinating virus of Japan (HVJ) envelope vector. The delivery of VepA into cells caused a significant decrease in cell viability, in contrast to the delivery of glutathione-S-transferase (GST), bovine serum albumin (BSA) or HVJ alone (Figure 1B). VepA did not affect cell viability in the absence of the HVJ vector. These results suggest that VepA itself is cytotoxic and effective only inside cells, not outside cells. By contrast, the delivery of a truncated VepA protein lacking the C-terminal 92 amino acids (1–400, VepAΔC) into HeLa cells did not affect the viability of the cells (Figure 1B). Although complementation of the wild-type vepA gene in the ΔvepA strain (POR-3ΔvepA/vepA) rescued the infection-mediated cytotoxicity, complementation with vepAΔC (POR-3ΔvepA/vepAΔC) did not (Figure 1C). We also confirmed that VepAΔC is secreted from POR-3ΔvepA/vepAΔC (Figure 1D), excluding the possibility that VepAΔC is not expressed or not secreted in V. parahaemolyticus. GFP-VepAΔC was successfully expressed in 293T cells, in contrast to GFP-VepA (Figure 1A), suggesting that VepAΔC is stable in cells. Taken together, these results indicate that VepAΔC loses the cytotoxicity.
To investigate the mechanism underlying VepA-dependent cytotoxicity, we tested whether autophagy and the MAPK signaling pathway were required for the cytotoxicity induced by V. parahaemolyticus. Inhibiting autophagy or MAPK with short interfering RNAs (siRNAs) or inhibitors did not affect the cytotoxicity (Figure S1), indicating that the contributions of autophagy and the MAPK signaling pathway to cytotoxicity are negligible. The cell-permeable pan-caspase inhibitor carbobenzoxy-valyl-alanyl-aspartyl-[O-methyl]-fluoromethylketone (Z-VAD-FMK) also did not affect the cytotoxicity, a result that is consistent with those of a previous study which the V. parahaemolyticus strain NY-4 was used [12].
Saccharomyces cerevisiae has been widely used as a model system to study eukaryotic cells. An increasing number of reports have shown that the expression of bacterial effectors inhibits yeast growth, and this inhibition is implicated in the activity of effectors that affect cellular processes conserved among eukaryotic cells [20]. To determine whether VepA can inhibit yeast growth, we transformed S. cerevisiae BY4730 with the p426 expression plasmid encoding VepA and expressed VepA under the control of the GAL1 promoter. The ectopic expression of VepA inhibited the growth of yeast (Figure 2A), indicating that VepA is also toxic in yeast. By contrast, VepAΔC, which lacks cytotoxicity (Figure 1A, 1B and 1C), was less toxic in yeast, suggesting a correlation between the cytotoxic effects of VepA in HeLa cells and its toxicity in yeast.
Next, we used a yeast knockout (YKO) strain library [21] to screen for host genes that are essential for the toxicity of VepA (Figure 2B). To express VepA in non-essential gene mutant yeast strains and screen for clones that are able to grow in the presence of VepA, we transformed 56 pools of YKO strains (one pool typically includes 95 strains) with the p426 expression plasmid encoding VepA and plated the yeast onto SC plates lacking uracil and containing galactose (SC-Ura+Gal). To examine the plating efficiency, the transformants were also plated onto SC plates lacking uracil and containing glucose (SC-Ura+Glc), yielding at least 1,000 colonies (giving >10×coverage). Using this genome-wide screen, we found that the Δvma3 strain was insensitive to the toxicity of VepA (Figure 2C). No MAPK- or autophagy-related gene mutants were identified in this screen, a result that is consistent with the cytotoxicity analysis presented in Figure S1. The expression of VepA in the Δvma3 strain was confirmed under inducing conditions (Figure 2D). The ectopic expression of VopT and VopP, which have been reported to inhibit yeast growth [9] (Figure 2A), was toxic to the Δvma3 strain (Figure 2C), thus excluding the possibility that the Δvma3 strain was non-specifically insensitive to bacterial effectors. Complementation of the VMA3 gene in the Δvma3 strain restored the susceptibility to the toxicity of VepA (Figure S2A). VMA3 encodes subunit c of V-ATPase. V-ATPases, which are highly conserved in eukaryotic cells, are composed of two domains: a peripheral, catalytic V1 domain and an integral V0 domain that serves as the basal body. V-ATPases serve as proton pumps to acidify intracellular compartments [22]. In yeast, the V1 and V0 domains contain eight (A, B, C, D, E, F, G and H) and six (a, c, c′, c″, d and e) subunits, respectively. Because the deletion of any subunit causes a functional deficiency in V-ATPases [23], we determined whether the function of the V-ATPase is involved in the VepA-mediated growth defects in yeast. VepA was expressed in the yeast strains containing mutations of each component of the V-ATPase, and their growth was observed. With the exception of Δvma3, none of the yeast strains with mutations in the V-ATPase subunits were able to grow in the presence of VepA (Figure 2E), thus indicating that the function of the V-ATPase is dispensable for the toxicity of VepA in yeast.
Next, we investigated the involvement of ATP6V0C, the human ortholog of yeast Vma3p, in VepA-mediated cytotoxicity. HeLa cells were transfected with siRNAs to knockdown ATP6V0C or ATP6V1A, subunit A of human V-ATPase. The transfected cells were infected with V. parahaemolyticus strain POR-3 [14], and the level of cytotoxicity was evaluated. The knockdown of ATP6V0C reduced infection-mediated cytotoxicity, but the knockdown of ATP6V1A did not (Figure 3A). The knockdown of expression of each protein was confirmed by immunoblotting (Figure 3B). We also determined whether the knockdown of ATP6V0C affected the translocation of VepA in cells. HeLa cells treated with siRNAs were infected with POR-3, ΔT3SS1 or ΔvepA strain for 3 h. The cell lysates were then analyzed by immunoblotting. The amount of VepA that was associated with the cells was not different between the control siRNA-treated cells and the ATP6V0C siRNA-treated cells (Figure 3C), indicating that the knockdown of ATP6V0C did not affect the translocation of VepA into cells. Moreover, pre-treatment with the pharmacological V-ATPase inhibitors, bafilomycin A and concanamycin A [22] at 100 nM, a concentration that is sufficient to prevent vesicular acidification as validated by LysoTracker staining (data not shown), did not prevent infection-mediated cytotoxicity (Figure 3D). These results indicate that ATP6V0C is involved in VepA-mediated cytotoxicity and that the function of the V-ATPase is not required for cytotoxicity, which is consistent with the yeast study described above (Figure 2E).
Next, we characterized the localization of VepA in V. parahaemolyticus-infected HeLa cells by biochemical subcellular fractionation. HeLa cells were infected with V. parahaemolyticus strains for 3 h to avoid the severe cytotoxicity that occurred at 4 h and fractionated into cytosolic, membrane/organelle, nuclear and cytoskeleton fractions. Each of the fractions was then analyzed by immunoblotting with an anti-VepA antibody. VepA was localized in the membrane/organelle fractions of cells infected with POR-3 or POR-3ΔvepA/vepA (Figure S3A). Furthermore, we fractionated the infected cells by ultracentrifugation (Figure S3B). VepA was mainly detected in the upper fractions, predominantly fraction 1, of cells infected with POR-3 or POR-3ΔvepA/vepA. This distribution of VepA is similar to that of V-ATPase (ATP6V1A and ATP6V0D1), which suggests that VepA may be associated with V-ATPases in infected cells.
We next investigated whether VepA interacts with ATP6V0C. Biotinylated VepA (b-VepA) (Figure S4) was adsorbed onto streptavidin beads and incubated with lysates of 293T cells expressing ATP6V0C-Flag. Immunoblot analysis showed that ATP6V0C-Flag co-precipitated with b-VepA-immobilized beads (Figure 4A), indicating that VepA interacts with ATP6V0C. To further search for endogenous proteins that associate with VepA, proteins bound to VepA were pulled down from lysates of RAW264.7 cells, which are susceptible to T3SS1-induced cytotoxicity, and thus, have a potential for high expression of target molecules for VepA [10], and visualized by silver staining of SDS-PAGE gels (Figure 4B). We found a protein with a molecular weight of approximately 16 kDa that specifically associated with VepA. This protein band was excised, analyzed by liquid chromatography-tandem mass spectrometry (LC-MS/MS) and identified as ATP6V0C, as expected. By contrast, VepAΔC did not interact with Flag-tagged or endogenous ATP6V0C (Figure 4A and 4B). In addition, in cells infected with POR-3ΔvepA/vepAΔC, the subcellular distribution of VepAΔC was different from that of VepA, and the protein was not enriched in fraction 1, which contains lysosomes (Figure S3B). ATP6V0C is an integral membrane protein that consists of 155 amino acids and is predicted to possess four transmembrane domains and two cytoplasmic loops [24] (Figure S5A). VepA is injected into the cytoplasm of host cells by the T3SS and therefore might target the cytoplasmic loops of ATP6V0C. To test this possibility, 293T cells expressing the cytoplasmic loops (Vc1: ATP6V0C 31–80 or Vc2: ATP6V0C 111–155) were infected with V. parahaemolyticus. Interestingly, the expression of Vc1 significantly reduced infection-mediated cytotoxicity (Figure S5B). In addition, Vc1 interacted with VepA in pull-down assays (Figure S5C). Taken together, our data indicate that ATP6V0C is a primary cellular target for VepA.
V-ATPases are expressed on the membranes of various intracellular organelles, where they transport H+ across the membrane to generate and maintain acidic compartments. The lysosome is a major acidic compartment that contains various hydrolytic enzymes and functions as a digestive apparatus. The leakage of degradative contents from lysosomes into the cytosol is known to induce cell death, and the type of cell death that occurs is thought to be dependent on the extent of lysosomal damage, i.e., partial and moderate ruptures cause apoptosis, whereas more severe damage leads to necrosis [25], [26]. To investigate the integrity of lysosomes in infected cells, we used acridine orange (AO), which is a fluorochrome stain used for vital staining of lysosomes that exhibits red fluorescence when concentrated in lysosomes and green fluorescence when diffused in the cytosol [27]. AO-loaded HeLa cells were infected with V. parahaemolyticus. Enhanced green fluorescence was observed in cells infected with POR-3 or ΔvepA/vepA but not in cells infected with POR-3ΔvepA or ΔT3SS1, both of which are deficient in VepA (Figure 5A–I). Thus, the relocation of AO to the cytosol was VepA dependent.
To examine the extent of lysosomal leakage, we determined whether cathepsin D (CatD), a lysosomal aspartyl protease, translocates from lysosomes to the cytosol. CatD was detected in the cytosolic fractions of cells infected with POR-3 or POR-3ΔvepA/vepA but not in cells infected with VepA-deficient strains or POR-3ΔvepA/vepAΔC, indicating that CatD was also released into the cytosol during infection in a VepA-dependent manner (Figure 5K). Furthermore, the knockdown of ATP6V0C partially reduced the release of CatD into the cytosol during infection (Figure 5L). We also treated AO-loaded HeLa cells with TDH, a Vibrio exotoxin. Although TDH does not contribute to the cytotoxicity of V. parahaemolyticus during infection, TDH attacks plasma membranes and functions as a pore-forming toxin that leads to cell death when HeLa cells are exposed to a high concentration of purified TDH [10], [28]. In contrast to infection with V. parahaemolyticus, treatment with TDH did not induce the relocation of AO to the cytosol (Figure S6A–D). Taken together, these results suggest that VepA-induced lysosomal rupture precedes plasma membrane disruption. In addition, pre-treatment of cells with U18666A or deferoxamine (DFO), both of which increase lysosomal membrane stability [29], [30], reduced infection-mediated cytotoxicity (Figure S6E and S6F).
We next investigated the involvement of VepA in infection-mediated lysosomal rupture. Because VepA interacts with V-ATPase subunit c, which is highly expressed on lysosomal membranes, it is possible that VepA directly affects the integrity of lysosomes. We therefore examined whether VepA could directly rupture lysosomes in a cell-free system. Isolated lysosomes were incubated with VepA or VepAΔC. After precipitating the lysosomes, the release of CatD from the lysosomes into the supernatant was examined. For lysosomes treated with VepA, detectable levels of CatD were released into the supernatant. By contrast, the release of CatD was not observed in the supernatants from lysosomes treated with VepAΔC (Figure 5M). Therefore, these results suggest that the cytotoxic T3SS effector VepA alone is able to induce lysosomal rupture and the leakage of contents from lysosomes. Finally, we examined the effect of knockdown of ATP6V0C on VepA-induced lysosomal rupture in the cell-free system. Lysosomes that were isolated from control siRNA- or ATP6V0C siRNA-treated cells, were treated with VepA. For lysosomes from ATP6V0C-silenced cells, CatD release was partially decreased (Figure 5N), indicating that the knockdown of ATP6V0C reduces VepA-induced lysosomal rupture. Thus, these results indicate that ATP6V0C is involved in VepA-induced lysosomal rupture.
Pathogens manipulate host cell death to facilitate their ability to cause infections [31], [32]. The bacterial pathogen V. parahaemolyticus elicits T3SS1-dependent non-apoptotic and caspase-independent cell death during infection (Figure S1) [11], [12]. Among the effectors translocated by the T3SS1 of V. parahaemolyticus, VepA is thought to play an important role in V. parahaemolyticus-induced cytotoxicity because the ΔvepA strain has lost the majority of the cytotoxicity of the wild-type [14], [15]. In this study, we showed that VepA itself is cytotoxic and acts only inside cells, not outside cells (Figure 1). It is reasonable for VepA to be cytotoxic only inside cells because it is injected into the cytoplasm of host cells by T3SS. Among the vibrios, V. alginolyticus, V. harveyi and V. tubiashii possess T3SSs that are highly homologous to V. parahaemolyticus T3SS1 [8]. The V. alginolyticus T3SS exhibits cytotoxicity in mammalian and fish cells [33], [34]. Because VepA is also conserved among these species, it is possible that VepA homologs are involved in the cytotoxicity of other Vibrio species.
Our yeast genome-wide screen revealed that subunit c of V-ATPase is indispensable for the toxicity of VepA in yeast (Figure 2). In HeLa cells, knockdown of ATP6V0C with siRNA decreased VepA-mediated cytotoxicity significantly but not completely (Figure 3), presumably due to the insufficient knockdown efficiency of ATP6V0C, as validated in Figure 3B. The expression of GFP-Vc1, a cytoplasmic loop of ATP6V0C, also reduced the cytotoxicity significantly but only partially (Figure S5B). GFP-fused cytoplasmic loops exhibited a diffuse cytoplasmic pattern and no lysosomal localization (data not shown), which may result in insufficient competition with ATP6V0C. Our pull-down assays suggested that VepA prefers Vc1, but Vc2 was also weakly associated with VepA (Figure S5C). An alternative explanation for the residual cytotoxicity is that VepA may recognize not only the cytoplasmic loop 1 of ATP6V0C but also the more complicated structure of ATP6V0C. Thus, although we cannot exclude the possibility that VepA has other unknown cellular targets that may also be involved in cytotoxicity at this stage, VepA-mediated cytotoxicity in HeLa cells at least partially requires subunit c of V-ATPase, supporting the validity of our yeast genome-wide screen.
V-ATPase is composed of multisubunits, all of which are required for the proton transport activity. The V-ATPase complex is divided into the two domains, V1 and V0 domains, which can assemble independently [22]. Subunit c of V-ATPase, a component of the V0 domain, is also known as ductin, which is thought to form a channel by itself [24]. Therefore, even in the absence of other ATPase components, it is possible that molecular complexes containing subunit c of V-ATPase are expressed in the membrane, although these complexes may not function as ATPases. In yeast, in the assembly of the V0 domain, deletions in the V0 subunits have been reported to result in the failure of the V0 domain to localize to vacuolar membranes [35], [36]. However, in our yeast toxicity assays, presented in Figure 2, VepA was shown to be toxic to the V0 subunit mutants except for Δvma3. Therefore, although it should be noted that VepA was ectopically overexpressed in yeast in these experiments, we cannot completely exclude the possibility that VepA may target subunit c of V-ATPase, which is also located in compartments other than vacuolar membranes. This possibility will be explored in the future.
Biochemical cell fractionation revealed the distribution of VepA is almost identical to that of V-ATPases (Figure S3). V-ATPases are expressed in acidic organelles and at plasma membranes [22]. Indeed, the distribution of V-ATPases was similar to that of lysosomes and plasma membranes in Figure S3B. By contrast, the distribution of VepAΔC, which did not interact with ATP6V0C, was different from that of wild-type VepA, suggesting that VepA is associated with these organelles via the interaction with subunit c of V-ATPases. V-ATPases are evolutionally conserved among eukaryotic cells [23]. A bacterial virulence factor that targets such a broadly distributed molecule among species is an efficient way to ensure a wide range of host species susceptibility. The Pseudomonas aeruginosa pigment pyocyanin and Legionella pneumophila SidK inactivate V-ATPase [37], [38]. Pyocyanin has been implicated in chronic P. aeruginosa infection in cystic fibrosis [39]. However, some yeast V-ATPase mutant strains are more sensitive to pyocyanin than the wild-type strain, a property that is distinct from the toxicity of VepA, which is not toxic to Δvma3. SidK, a type IV secreted effector, targets subunit A of V-ATPase, which contains the ATP hydrolytic site, and inhibits ATP hydrolysis to prevent the proton pump function of V-ATPase and phagosomal acidification [38]. By contrast, our data indicate that, although VepA targeted subunit c of V-ATPase, the function of V-ATPase is dispensable for the toxicity of VepA (Figure 2 and Figure 3). Unlike VepA, which showed severe and acute cytotoxic activity within 4 h after infection (Figure 1 and S6C), SidK appeared not to be cytotoxic during the early infection period because macrophages loaded with the SidK protein survive longer than 24 h [38]. This difference may reflect the difference between the infection strategies of the two pathogens: V. parahaemolyticus is an extracellular pathogen that causes acute infection [5], whereas L. pneumophila is an intracellular pathogen that survives and replicates in phagosomes and therefore needs to avoid phagosomal acidification [40].
Infection with V. parahaemolyticus caused VepA injection-dependent leakage of the lysosomal contents (Figure 5). Lysosomes are described as “suicide bags”, because they contain numerous hydrolases [41]. Lysosomotropic agents such as H2O2 and sphingosine induce lysosomal membrane permeabilization and subsequent cell death [42]. Lysosomal rupture-induced cell death is thought to be dependent on the extent of lysosomal damage; partial and moderate rupture causes apoptosis, whereas more severe damage leads to necrosis [25], [26]. V. parahaemolyticus infection-mediated leakage of lysosomal contents was observed not only with a small-molecule dye AO but also with the protein CatD, suggesting that extensive lysosomal membrane permeabilization occurs within hours of infection. Notably, our cell-free assay showed that VepA is sufficient to reproduce infection-mediated lysosomal rupture, whereas VepAΔC, which is deficient for the association with ATP6V0C, could not induce the leakage of lysosomal contents. Moreover, the knockdown of ATP6V0C reduced VepA-induced lysosomal rupture (Figure 5L and 5N), indicating that ATP6V0C is required of VepA-induced lysosomal rupture. Thus, our data provide the first example of a bacterial T3SS effector that ruptures lysosomes directly to induce cell death.
Various bacterial effectors are known to have enzymatic activities such as proteolytic processing and post-translational modification [7]. However, despite experimental efforts, we could not detect the processing or modification of ATP6V0C by VepA in this study. Although it is not yet known how the interaction of VepA with ATP6V0C leads to destabilization of lysosomal membranes, ATP6V0C may serve as a scaffold for VepA, facilitating the access of VepA to lysosomal membranes and subsequent lysosomal rupture. Alternatively, the association of VepA with ATP6V0C may destabilize the ATP6V0C complex, which may result in lysosomal destabilization. Two lysosome stabilizers, U18666A and DFO, partially inhibited T3SS1-dependent cytotoxicity (Figure S6E and S6F). U18666A inhibits transport of cholesterol from lysosomes to the ER, which causes the accumulation of cholesterol in lysosomes [29]. DFO is known to protect against H2O2-induced lysosome rupture by chelating intralysosomal iron [30]. In contrast to those direct effects on lysosomal stabilization, the incubation of heat shock protein 70 (Hsp70) with cells, which is reported to stabilize lysosomes by an indirect effect that stimulates acid sphingomyelinase activity [43], failed to prevent T3SS1-dependent cytotoxicity (data not shown). Thus, these results suggest that VepA physically destabilizes lysosomal membranes.
The cell death induced by V. parahaemolyticus has also been reported to be associated with autophagy [11]. Although autophagy is well known to promote cell survival in response to various stimuli [44], recent studies have indicated that autophagy also plays a role as an executor of cell death in some aspects [45], [46]. However, the role of autophagy induced by V. parahaemolyticus in cell death is unclear. A previous report showed that V. parahaemolyticus does not activate autophagy in Atg5-/- murine embryonic fibroblasts, indicating that V. parahaemolyticus-induced autophagy is dependent on ATG5 [16]. The results of this study were consistent with those of a prior study, demonstrating that ATG5 depletion by siRNA inhibited V. parahaemolyticus-induced autophagy in HeLa cells (Figure S1E). However, ATG5 depletion did not affect T3SS1-dependent cytotoxicity (Figure S1C). Moreover, Δatg5 and Δatg8 yeast strains, both of which are deficient in the autophagic process, were not resistant to VepA (Figure S2B). Indeed, although there is a link between autophagy and lysosomal biogenesis [47], our results indicate that autophagy does not contribute to the cell death induced by V. parahaemolyticus. Thus, V. parahaemolyticus induces “cell death with autophagy” but not “cell death by autophagy” [48]. It is thus possible that autophagy may play a protective role against cytotoxicity of V. parahaemolyticus. It has also been reported that VepA is linked to the activation of MAPK signaling when cells are infected with V. parahaemolyticus [19]. MAPK cascades are activated by various stimuli including cellular stress [49]. A recent report has shown that deficiency in the tumor susceptibility gene 101 leads to lysosomal distention as well as induction of autophagy and MAPK activation [50]. Thus, it is possible that lysosomal stress may be linked to induction of stress response, such as activation of MAPK and induction of autophagy.
In conclusion, we demonstrated that VepA targets ATP6V0C and ruptures lysosomes. Although the mechanism of lysosomal rupture by VepA warrants further exploration, it is intriguing that a pathogenic bacterium induces cell death by causing the host cell to leak its own dangerous content, which is akin to striking the Achilles' heel of the host cell.
The Vibrio parahaemolyticus strains POR-3 (RIMD2210633ΔtdhASΔvcrD2, which is T3SS2 deficient) and POR-3ΔvcrD1 (which lacks both T3SS1 and T3SS2) were described previously [9], [14]. The POR-3ΔvepA strain was created as described previously [14]. To complement the mutant with vepA or vepAΔC (1–400), the pSA-19CS-MCS vector was used as described previously [14]. Saccharomyces cerevisiae BY4730 (MATa leu2Δ0 met15Δ0 ura3Δ0) was obtained from Open Biosystems. The plasmids and oligonucleotide primers used in this study are listed in Table S1.
Bafilomycin A and concanamycin A (both used at 100 nM) were purchased from Sigma. Acridine orange was purchased from Invitrogen. U18666A, deferoxamine, SP600125, U0126 and SB20358 were purchased from Sigma. The pan-caspase inhibitor Z-VAD-FMK was from Medical & Biological Laboratories. The following antibodies were used: antibodies against Lamp-1, CD49b, Grb78, GM130, nucleoporin-62, Bcl-2 and Hsp90 (BD Biosciences); antibodies against β-actin, Flag and poly-histidine (Sigma); anti-ATP6V1A (Abnova); anti-ATP6V0D1 (Abcam); anti-CatD (Cell Signaling Technology); HRP-conjugated streptavidin (Pierce); anti-ATP6V0C (Millipore); HRP-conjugated anti-GFP (Miltenyi Biotech); and anti-ATG5 and anti-LC3 (Medical & Biological Laboratories). The anti-VepA antibody has been described elsewhere [15].
HeLa, RAW264.7 and 293T cells were maintained in DMEM (Sigma) containing 10% FBS (Sigma) at 37°C in 5% CO2. For infections, bacteria were used to challenge HeLa cells at a multiplicity of infection of 10 [10]. The cytotoxicity assay was performed using the CytoTox96 Non-Radioactive Cytotoxicity Assay Kit (Promega) as previously described [9].
Yeast was transformed using the Frozen-EX Transformation Kit II (Zymo Research). Transformants were grown on SC plates lacking uracil and containing 2% glucose (SC-Ura+Glc) at 30°C for 48–72 h. Growing colonies were picked and cultured in SD media lacking uracil and containing 2% glucose (SD-Ura+Glc). Yeast was washed once with 0.67% yeast nitrogen broth without amino acids and adjusted to an optical density of 1 at 600 nm. SC plates lacking uracil and containing glucose or galactose (SC-Ura+Gal) were spotted with 5-µl aliquots of 10-fold serial dilutions and then incubated at 30°C for 72 h. For the complementation of VMA3 in Δvma3, a LEU2 plasmid pRS415, encoding VMA3 along with 500 bp upstream and downstream of VMA3 was used.
His-tagged VepA or VepAΔC protein was purified as described previously [15]. To construct pATP6V0C-Flag, cDNA for ATP6V0C with a Flag-tag and a stop codon at the C-terminus was inserted into pEGFP-N1, yielding the ATP6V0C-Flag construct without GFP.
For DNA transfection, 293T cells were used. To transfect the VepA constructs, 293T cells were seeded on collagen-coated coverslips in 6-well plates and grown for 24 h. The cells were then transfected with pEGFP-C1, pEGFP-VepA or pEGFP-VepAΔC using Lipofectamine 2000 (Invitrogen). The cells were fixed with 3% paraformaldehyde and permeabilized with 0.2% Triton X-100 (TX-100) 24 h post-transfection and then stained with rhodamine-phalloidin. The coverslips were analyzed by fluorescence microscopy using a Biozero BZ-8100 microscope (Keyence). To infect cells expressing the ATP6V0C constructs, 293T cells seeded in 96-well plates were transfected with pEGFP-C1, pEGFP-Vc1 or pEGFP-Vc2 using Lipofectamine LTX (Invitrogen). After 24 h, the cells were infected with POR-3 at a multiplicity of infection of 10 for 3 h and then assessed using the cytotoxicity assay.
Proteins were introduced into cells using GenomONE-Neo (Ishihara Sangyo) according to the manufacturer's instructions. Briefly, 10 µl of HVJ-envelope was mixed with 2.5 µl of Reagent A and incubated for 5 min on ice. The suspension was mixed with 5 µg of protein and then with 1.5 µl of Reagent B. After centrifugation at 10,000 g for 5 min, the supernatant was removed. The pellets were resuspended in 15 µl of Reagent Buffer, followed by the addition of 2.5 µl of Reagent C to complete the preparation of the HVJ-liposome-including proteins. A one-eighth aliquot of the envelope was added to cells in a 96-well plate. The cells were incubated at 37°C for 1.5 h and then washed with phosphate-buffered saline (PBS). The medium was replaced with fresh medium, and the cells were incubated at 37°C for 2.5 h, and cytotoxicity was determined by assaying the cellular dehydrogenase activity using Cell Counting Kit-8 (Dojindo).
Yeast MATa haploid non-essential gene knockout strains from Open Biosystems [21] were grown in YPD medium in 96-well plates at 30°C. All of the strains from a single plate were then pooled into a single culture, and transformed with p426-VepA as described above. Transformants were plated on SC-Ura+Gal and incubated at 30°C for 3 days. To determine the plating efficiency, transformants were also plated on SC-Ura+Glc. Colonies grown on galactose plates were picked and incubated in SD-Ura+Glc. Chromosomal DNA was purified by phenol/chloroform extraction with glass beads as described elsewhere [51]. Barcode-tag sequence regions of the extracted yeast DNA were amplified by polymerase chain reaction. Then, amplicons were purified using ExoSAP-IT (GE healthcare) and sequenced to verify the barcode-tag. We eliminated mutants that were only recovered in a single clone to exclude false positives. For the second selection, the identified mutants were individually transformed with p426-VepA, and their growth on SC-Ura+Gal was examined. A flowchart summarizing the protocol for the yeast genome-wide screen used in this study is shown in Figure 2B.
Ambion's Silencer Select Validated siRNAs and Silencer Select Pre-designed siRNAs were used to knockdown ATP6V0C and ATP6V1A, respectively. HeLa cells were reverse-transfected with 20 nM Silencer Negative Control#2 siRNA or with two independent siRNAs targeting ATP6V0C (V0C#1; s80, V0C#2; s81) or ATP6V1A (V1A#1; s1791, V1A#2; s1792) using the siPORT NeoFX Transfection Agent (Ambion). Then, 96 h after siRNA transfection, the cells were infected with V. parahaemolyticus strain POR-3 as described above. To determine the knockdown efficiency, immunoblot analysis was performed using anti-ATP6V1A and anti-ATP6V0C antibodies. To deplete ATG5, FlexiTube siRNAs (Qiagen) were used. HeLa cells were transfected with AllStars Negative Control siRNAs or two separate siRNAs targeting ATG5 (ATG5_2; #1, ATG5_6; #2) using HiPerFect Transfection Reagent (Qiagen). Then, 72 h post-transfection, the cells were infected with strain POR-3 for 4 h, and cytotoxicity was evaluated.
For pull-down assays, 293T cells expressing ATP6V0C-Flag or RAW264.7 cells were lysed with RIPA buffer containing a protease inhibitor cocktail (Sigma). VepA and VepAΔC were biotinylated using EZ-Link Sulfo-NHS-SS-Biotinylation kits (Pierce). Biotinylated proteins were captured with Streptavidin Sepharose beads (GE healthcare). Then, VepA- or VepAΔC-immobilized beads were incubated with cell lysates on a rotary shaker for 2 h at 4°C. The beads were washed with RIPA buffer five times, and resuspended in SDS-loading buffer to elute bound proteins from the beads. Eluates were subjected to SDS-PAGE and immunoblot analysis or LC-MS/MS analysis as described previously [52]. Proteins were identified by a database search using Mascot (Matrix Science).
Subcellular fractionation was performed using the ProteoExtract Subcellular Proteome Extraction Kit (Calbiochem) according to the manufacturer's instructions. For organelle fractionations, we used the Lysosome Enrichment Kit for Tissue and Cultured Cells (Pierce) according to the manufacturer's instructions. Briefly, the lysates from infected cells were applied to a 15–30% OptiPrep density gradient. After ultra-centrifugation at 150,000 g for 2 h at 4°C in a SW55 rotor (Beckman Coulter), 800-µl fractions were collected from the top of the tube. Each fraction was diluted with PBS and centrifuged at 20,000 g for 30 min at 4°C. After removing the supernatants, the pellets were dissolved with SDS-loading buffer and subjected to SDS-PAGE and immunoblot analysis. To prepare the cytosolic fraction, HeLa cells were solubilized with saponin buffer (0.1% saponin and protease inhibitors in PBS) for 15 min on ice. After centrifugation at 15,000 g for 10 min, the supernatants were collected. The cytosolic fraction was free of membranes as verified by immunoblots using an anti-Lamp-1 antibody. Lysis refers to treating the cells with 1% TX-100 to induce the total release of CatD as a positive control.
HeLa cells grown on 35-mm glass bottom dishes were treated with 2 µg ml−1 AO at 37°C for 15 min. After washing with PBS, the medium was replaced with fresh medium, followed by infection or exposure to TDH. Fluorescence micrographs of AO-stained cells were obtained using the same fluorescence intensity and exposure time. The intensity of green fluorescence was quantitated with a BZ Analyzer II (Keyence).
The cell-free lysosome-enriched fraction was prepared as described [53]. Briefly, HeLa cells were homogenized in ice-cold TS buffer (10 mM Tris-HCl, pH 7.5, 250 mM sucrose and protease inhibitors). Then, the cells were centrifuged at 1,000 g to pellet the nuclei and cell debris. The supernatants were centrifuged at 3,000 g. The supernatants were further centrifuged at 17,000 g, and the resulting pellets, which contained the lysosomes, were resuspended with PBS and used as a cell-free lysosomal preparation. Then, 50-µl aliquots of prepared lysosomes (50 µg) were exposed to 10 µg ml−1 VepA or VepAΔC for 2 h at 37°C. The reaction mixture was centrifuged at 20,000 g for 30 min at 4°C to separate the pellets, which contained the lysosomes, from the supernatants, which contained the material that leaked from the lysosomes. Both fractions were subjected to SDS-PAGE and immunoblot analyses using anti-CatD and anti-Lamp-1 antibodies.
To induce autophagy, HeLa cells were starved in Earle's balanced salt solution for 6 h or infected with V. parahaemolyticus POR-3 or ΔvepA for 3 h. For the nutrient-rich condition, cells were cultured in DMEM containing 10% FBS. To measure autophagy, the conversion of LC3-I to LC3-II was monitored by immunoblotting as previously described [54].
Statistical analysis was performed using the two-tailed Student's t-test.
|
10.1371/journal.pntd.0006498 | Regulation of midgut cell proliferation impacts Aedes aegypti susceptibility to dengue virus | Aedes aegypti is the vector of some of the most important vector-borne diseases like dengue, chikungunya, zika and yellow fever, affecting millions of people worldwide. The cellular processes that follow a blood meal in the mosquito midgut are directly associated with pathogen transmission. We studied the homeostatic response of the midgut against oxidative stress, as well as bacterial and dengue virus (DENV) infections, focusing on the proliferative ability of the intestinal stem cells (ISC). Inhibition of the peritrophic matrix (PM) formation led to an increase in reactive oxygen species (ROS) production by the epithelial cells in response to contact with the resident microbiota, suggesting that maintenance of low levels of ROS in the intestinal lumen is key to keep ISCs division in balance. We show that dengue virus infection induces midgut cell division in both DENV susceptible (Rockefeller) and refractory (Orlando) mosquito strains. However, the susceptible strain delays the activation of the regeneration process compared with the refractory strain. Impairment of the Delta/Notch signaling, by silencing the Notch ligand Delta using RNAi, significantly increased the susceptibility of the refractory strains to DENV infection of the midgut. We propose that this cell replenishment is essential to control viral infection in the mosquito. Our study demonstrates that the intestinal epithelium of the blood fed mosquito is able to respond and defend against different challenges, including virus infection. In addition, we provide unprecedented evidence that the activation of a cellular regenerative program in the midgut is important for the determination of the mosquito vectorial competence.
| Aedes mosquitoes are important vectors of arboviruses, representing a major threat to public health. While feeding on blood, mosquitoes address the challenges of digestion and preservation of midgut homeostasis. Damaged or senescent cells must be constantly replaced by new cells to maintain midgut epithelial integrity. In this study, we show that the intestinal stem cells (ISCs) of blood-fed mosquitoes are able to respond to abiotic and biotic challenges. Exposing midgut cells to different types of stress, such as the inhibition of the peritrophic matrix formation, changes in the midgut redox state, or infection with entomopathogenic bacteria or viruses, resulted in an increased number of mitotic cells in blood-fed mosquitoes. Mosquito strains with different susceptibilities to DENV infection presented different time course of cell regeneration in response to viral infection. Knockdown of the Notch pathway in a refractory mosquito strain limited cell division after infection with DENV and resulted in increased mosquito susceptibility to the virus. Conversely, inducing midgut cell proliferation made a susceptible strain more resistant to viral infection. Therefore, the effectiveness of midgut cellular renewal during viral infection proved to be an important factor in vector competence. These findings can contribute to the understanding of virus-host interactions and help to develop more successful strategies of vector control.
| The mosquito Aedes aegypti is a vector of several human pathogens, such as flaviviruses, including yellow fever (YFV), dengue (DENV) and zika (ZIKV), and thus this mosquito exerts an enormous public health burden worldwide [1,2]. During the transmission cycle, these insects feed on volumes of blood that are 2–3 times their weight, and the digestion of this large meal results in several potentially damaging conditions [3]. The digestion of blood meal requires intense proteolytic activity in the midgut and results in the formation of potentially toxic concentrations of heme, iron, amino acids and ammonia [4]. The midgut is also the first site of interaction with potential pathogens, including viruses, and supports a dramatic increase in intestinal microbiota after blood feeding [5,6]. To overcome these challenges, the ingestion of a blood meal is followed by several physiological processes, such as formation of a peritrophic matrix (PM) [7,8] and down-regulation of reactive oxygen species (ROS) production. In addition, the midgut epithelium is the first barrier that viruses must cross in the mosquito to achieve a successful viral cycle (reviewed in [9]). Thus, in order to ensure epithelial integrity and the maintenance of midgut homeostasis, the midgut epithelium must fine tune key cellular mechanisms, including cell proliferation and differentiation.
In both vertebrate and invertebrate animals, the gut epithelia have a similar basic cellular composition: absorptive enterocytes (ECs) that represent the majority of the differentiated cells and are interspersed with hormone-producing enteroendocrine cells (ee). The intestinal stem cells (ISCs) and enteroblasts (EB) account for the progenitor cells, responsible for replenishing the differentiated cells that are lost due to damage or aging [10–14]. In A. aegypti, description of the different cellular types and functions started with identification and basic characterization of absorptive (ECs) and non-absorptive cells (ISC, EB, and enteroendocrine cells) [15]. To date, the study of division properties of the ISCs in this vector species remains limited to the description of the division process during metamorphosis [16].
Several conserved signaling pathways are known to be involved in midgut tissue renewal and differentiation. Comparative genomic analysis of some of these pathways has been done between Drosophila melanogaster and vector mosquitoes [17,18], but functional studies in Aedes, under the context of tissue regeneration, are still necessary. Notably, the Notch signaling pathway regulates cell differentiation in the midgut of both mammals and D. melanogaster. In this fruit fly, loss of function of Notch is attributed to the increase of intestinal cell proliferation and tumor formation [19]. However, it has already been shown that depletion of Notch in D. melanogaster ISCs also leads to stem cell loss and premature ee cell formation [20]. Accordingly, disruption of Notch signaling in mice has resulted in decreased cell proliferation coupled with secretory cell hyperplasia, whereas hyperactivation of Notch signaling results in expanded proliferation with increased numbers of absorptive enterocytes [21], as also observed in D. melanogaster [20].
In the fruit fly, the ingestion of cytotoxic agents, such as dextran sodium sulfate (DSS), bleomycin or paraquat, or infection by pathogenic bacteria can stimulates cell turnover, increasing the midgut ISC mitotic index [18,22]. Similar to that, it has been recently shown that cell damage produced by ingestion of several stressors also induced intestinal cell proliferation in sugar-fed Aedes albopictus [23]. Likewise, viral infections can trigger cellular responses, such as apoptosis or autophagy, in different infection models [24–27]. However, the interplay between intestinal cell proliferation and pathogen transmission has been a neglected subject in the literature.
In this study, we have characterized the dynamics of A. aegypti intestinal epithelium proliferation during blood meal digestion in response to oxidative stress, bacterial infections, and viral infections. We have also shown that two mosquito strains with different DENV susceptibilities [28] presented differences in cell mitotic rates after viral infection. Finally, our results indicate for the first time that the ability to replenish midgut cells by modulation of cell renewal involves the Delta-Notch signaling and is a key factor that influences A. aegypti competence to transmit DENV. We show that the cell proliferation rates influences mosquito infection and vector competence for DENV.
Aedes aegypti adult females acquire DENV and other arboviruses during the blood feedings that are needed to complete the reproductive cycle of the mosquito. To characterize the epithelial adaptation to this event, we first evaluated the cellular response to the blood meal itself. Upon ingestion, the blood induces dramatic changes in the Red strain mosquito midgut at a chemical, microbiological and physiological level. We attempted to dissect each of these challenges, to understand the delicate balance of the factors that play a role in the intestinal micro-environment in which the arbovirus has to thrive in order to pass to the salivary gland and be transmitted.
The tissue homeostasis of the midgut depends on the ability to replenish the damaged cells, and this depends on the presence of ISCs. Due to the lack of specific markers for progenitor cells for A. aegypti, we used morphological and physiological parameters to define the presence of ISCs in the adult females. Progenitor cells are well characterized for their basal positioning and being diploid, different to the apical localization of differentiated cells and the polyploidy of enterocytes. Both cell types were clearly distinctive, as well as the peritrophic matrix, in the midgut epithelium of blood-fed adult females (Fig 1A). The further characterization of ISC’s was performed with phospho-histone 3 antibodies, to specifically mark cells undergoing mitosis. In Fig 1B, it can be observed the two monolayers of the A. aegypti midgut, where ECs are clearly distinguishable and the PH3+ cell is found, with nuclei corresponding to the diploid size, located basally. Clearly, not every ISC present in the tissue is going to be found undergoing mitosis, but the presence of PH3+ cells, undoubtedly characterizes such cells as ISCs.
To evaluate the homeostatic cell proliferation of the Aedes aegypti midgut, we observed the number of cells undergoing mitosis in adult females. After a blood meal, the midgut epithelium showed a lower number of cells undergoing mitosis (phospho-histone 3 positive; PH3+) compared with that of sugar-fed insects (Fig 1C and 1D). To test if this decrease in mitotic cells was due to progenitor cell impairment, we fed insects with blood supplemented with the pro-oxidant compound paraquat. The midgut epithelium responded to an oxidative challenge by increasing mitosis (Fig 1C and 1D), indicating that the intestinal stem cells maintained the ability to divide and replenish damage cells after an insult at blood-fed conditions.
A hallmark of blood digestion is the formation of the peritrophic matrix (PM), a chitin and protein-rich non-cellular layer secreted by the midgut epithelium [7,8]. The mosquito type-I PM surrounds the blood bolus, limiting a direct contact between the epithelium, the blood meal and the indigenous microbiota, thereby playing a similar function as the vertebrate digestive mucous layer. Ingestion of blood contaminated with bacteria allows close contact of these microorganisms to the midgut epithelium before PM formation, which is completed formed only a few hours (14 to 24 hours) after a blood meal [7]. In fact, oral infection with sub-lethal concentrations of the non-pathogenic Serratia marcescens or the entomopathogenic Pseudomonas entomophila bacteria resulted in a significant increase in mitosis of the epithelium cells (Fig 2A and 2B). The increased cell turnover was also observed when heat-killed P. entomophila was provided through the blood, indicating that molecules derived from these entomopathogenic bacteria are sufficient to trigger the cell proliferation program, not necessarily requiring tissue infection (Fig 2B). In this case, tissue damage may at least partially be attributed to the lack of cell membrane integrity promoted by Monalysin, a pore-forming protein produced by P. entomophila [29].
Supplementation of blood with diflubenzuron (DFB), a chitin synthesis inhibitor [30], leads to the inhibition of PM production, exposing the gut epithelium directly to the luminal content (S1 Fig). Consequently, DFB administration resulted in elevated numbers of mitotic cells (Fig 2C). The co-ingestion of antibiotics completely abolished this effect of DFB on cell proliferation (Fig 2C), demonstrating that in the absence of the microbiota, the lack of the peritrophic matrix did not result in elevated mitosis. These results indicate that not only oral infection with pathogenic bacteria, but also the proliferation of the resident microbiota (by inhibition of PM in this case), in contact with the epithelium, can trigger the midgut proliferative program.
Exposure of D. melanogaster enterocytes to bacteria results in ROS production as a microbiota control mechanism. However, the oxidative species produced as a result of bacterial presence can also cause damage to the midgut cells [31–34]. When mosquitoes were fed with blood supplemented with DFB together with the antioxidant ascorbate (ASC), the mitosis levels dropped significantly (Fig 2C). The ROS production by the midgut epithelium was assessed by fluorescence microscopy using the fluorescent oxidant-sensing probe dihydroethidium (DHE). As shown in Fig 2D and 2E, the midguts of DFB-fed mosquitoes exhibited a high fluorescence signal, indicating an intense production of ROS. The intensity of the fluorescence signal of the DFB-treated midguts was significantly reduced upon ascorbate supplementation of the blood meal. Similarly, the suppression of microbiota with antibiotics dramatically reduced ROS levels. These results suggest a mechanism linking PM impairment to ISC proliferation, indicating that the direct exposure of the midgut epithelium to microbiota activates the production of ROS as part of an immune response.
The role of epithelial tissue regeneration of the midgut upon viral infection has not been investigated in mosquitoes. Thus, we decided to evaluate the gut regeneration pattern of two mosquito strains that are known to exhibit different susceptibilities to DENV infection [28]. In basal conditions, i.e. sugar fed, all the strains used in this study presented no difference in the number of cells under mitosis (S2 Fig). However, after 24 hours of taking a non-infected blood meal (day 1), the DENV refractory Orlando (Orl) strain presented a higher number of mitotic cells compared with the susceptible Rockefeller (Rock) strain (Fig 3A and 3B), indicating that the refractory strain is naturally more proliferative than the susceptible one under these conditions. In the following days, both strains showed similar time course profiles of mitotic activity. Upon ingestion of DENV-infected blood, the refractory Orlando strain showed an increase of mitotic cells, peaking at the second day post blood meal (Fig 3C). Subsequently, these midguts showed low numbers of cells in mitosis throughout the remaining course of infection, reaching a similar number as non-infected midguts. In contrast, the susceptible Rockefeller strain showed a delayed regenerative response, only reaching the maximum rate at five days after infection (Fig 3C). These results suggest that the midgut cells of refractory mosquitoes are able to respond more promptly to the early events of infection.
To test whether the differences in gut homeostatic responses between the two strains could be a determinant of refractoriness/susceptibility, we disturbed the homeostatic condition of ISCs by silencing delta expression. The Notch ligand Delta (Dl) is an upstream component of the Notch pathway that is involved in cell division and differentiation. The delta gene is expressed in adult ISC cells. Thus, accumulation of Delta is used as a marker of ISCs in D. melanogaster [19]. Furthermore, Delta expression is induced by infection in the D. melanogaster midgut [35]. The efficiency and duration of Delta silencing by RNAi are shown in Fig 4A and S3 Fig, respectively. Silencing delta led to a significant reduction in mitosis in both mosquito strains (Fig 4B and 4C). Interestingly, silencing of delta did not have an effect on infection susceptibility in the Rockefeller strain (Fig 4D). In contrast, it significantly increased susceptibility of the Orlando strain to DENV infection, as observed by the increased viral titers in the delta–silenced refractory strain compared with the dsGFP-injected group (Fig 4D). Conversely, when the susceptible strain was pre-treated with DSS, a known inducer of midgut cell damage, and thereby ISC proliferation [18] and S4 Fig, a significant reduction was seen in both DENV infection intensity (Fig 4E) and prevalence (Fig 4F) in the midgut, compared with non-treated mosquitoes. Similar results were observed when DSS-treated Rock mosquitoes were infected with DENV4 isolates (S5 Fig). These data clearly indicate that the ability of midguts to respond at the cellular level, via regeneration of epithelial cells, modulates the success of viral infection of A. aegypti. Furthermore, these results show for the first time that the mosquito processes required to replenish damaged cells and control tissue homeostasis are determinants of vector competence.
Cell renewal is known to be the basis of midgut epithelial integrity in model animals such as fly and mice [12]. Given the importance of the midgut epithelium in mosquitoes, where this tissue is effectively the first barrier that arboviruses affront to complete the transmission cycle [9], we decided to address the question of how this epithelium replenish its cells during the different challenges of blood feeding and infection. Previous descriptive reports of epithelial cell structure, function and midgut remodeling during metamorphosis [15,16,36] have shed some light on this process in mosquitoes, suggesting that the cell types described in other organisms, such as D. melanogaster, are also found in A. aegypti. Amongst the fully differentiated cells, the enterocytes were clearly distinguishable by their large nuclei size, abundance and localization [10,11]. However, due to the current lack of mosquito specific markers for other differentiated and progenitor cells, like ee’s and EB’s, only recently these cells were identified in mosquitoes larvae [37]. Nonetheless, ISC hallmark capacity is to undergo mitosis, which can be marked using antibodies for phosphorylated histone 3. This allowed us to successfully identify the presence of ISC in the epithelium, and to quantify the number or cells dividing in the different conditions evaluated (Fig 1A and 1B).
In the life history of mosquitoes, blood feeding represents a dramatic change from a sugar diet to ingestion of a large protein-rich meal. This transition imposes challenges to midgut homeostasis that are not faced by non-hematophagous insects. Knowledge about the mechanisms involved in the maintenance of midgut cellular integrity and homeostasis upon blood feeding or stress conditions is limited not only for A. aegypti, but also for other important vectors. In this study, we show unique properties of the mosquito midgut, suggesting that the regulation of epithelial cell proliferation is tightly regulated to allow proper handling of both chemical and biological sources of stress, including DENV infection, that occur during and after blood digestion. Based on these findings, we suggest that this regulation of midgut homeostasis is an important determinant of viral infection dynamics in the vector gut.
In A. aegypti, the maximal digestion rate is attained 24 hours after a blood meal [38]. Despite the dramatic increase of the microbiota, approximately 1000 times the levels before a meal [5], mosquitoes seem to maintain midgut epithelial cell turnover controlled (Fig 1C and 1D). One explanation for this is the physical separation between the bolus and the epithelium by the PM. The PM is a thick extracellular layer composed mostly of chitin fibrils and glycoproteins that is gradually formed from 12–24 hours after a blood meal and surrounds the blood bolus, creating a physical separation from the midgut epithelium [7,8]. To preserve homeostasis, the PM establishes a selective barrier, permeable to nutrients and digestive enzymes but acting as a first line of defense against harmful agents. We show here that when the midgut epithelium was exposed to pathogenic bacteria ingested with the blood meal, thus before PM formation, there was a marked increase of mitosis (Fig 2B). More importantly, inhibition of the PM formation also resulted in elevated mitotic cell counts (Fig 2C). Treating insects with antibiotics abolished the mitosis upregulation promoted by chitin synthesis inhibition, further demonstrating that the contact of the blood bolus itself was not the determining factor to the increase mitotic cell numbers, but instead, the consequent exposure of the gut epithelium to the indigenous bacterial microbiota present in the lumen was the predominant event that elicited this response. In this way, the compartmentalization of the bolus may allow the enterocytes to minimize their exposure to deleterious agents, and it results in reduced need to shed and replenish damaged cells.
ROS production by midgut cells represents a major innate immunity effector mechanism that is involved in the control of the microbiota. However, ROS can also damage host cells, and thus, a proper balance between ROS production and microbial suppression is essential for the health of the host itself [31–34,39]. Here, we show that production of ROS was activated when PM formation was blocked and that this effect can be prevented by antibiotics (Fig 2D). Therefore, we propose that the signaling mechanism that leads to increased mitosis after exposure to indigenous bacteria is the production of ROS by the intestinal cells, as a defensive, yet possibly damaging, response (Fig 2).
The midgut epithelial cells are the first to support viral replication within the mosquito vector and several studies have addressed the immune response of the mosquito to the virus [40]. Additionally, it is well-established that changes in ROS production in the midgut impact not only innate immunity responses against bacteria, but can also affect the mosquito ability to transmit human pathogens [5,41–44]. Despite this comprehensive knowledge about infection-related processes that occur within midgut cells, little is known about the cell turnover prior to and after infection. It was our intention to evaluate if this natural process of the midgut epithelium was different between mosquito strains with different degrees of susceptibility to DENV. Rockefeller (Rock) and Orlando (Orl) strains are susceptible and refractory strains respectively; however, under normal (sugar fed) conditions, they possess similar levels of mitotic cells (S2 Fig). Interestingly, the Orl strain possesses higher levels of mitosis than the Rock strain 24 hours after the blood meal (Fig 3A and 3B). This increased number of mitotic cells, is restricted to this specific time window, as 48 hours after the feeding, the numbers are no longer significantly different. This fact becomes relevant when the timeline is superposed to the timeline of the initial viral infection [45]. This becomes more apparent, when the numbers of mitotic cells on the susceptible Rock strain increase after 5 days, in a consistent timeline to the establishment of a successful infection with higher levels of infected cells, which is not observed in Orl strain that constrains the infection. In day 7, when the viruses normally leave the midgut to infect other tissues [45], the mitotic rate is reduced to levels compared of non-infected sugar-fed midguts in both strains (Fig 3C). Transcriptomic analyses of mosquito strains with different degrees of susceptibility to DENV revealed that some genes associated with cellular proliferation, growth and death are differentially expressed in refractory strains, upon DENV infection [46–49]. However, this has not been directly associated to midgut regeneration in these studies. In addition, the increased expression and activation of a variety of apoptotic cascade components in the midgut after viral infections implicate apoptosis as part of the A. aegypti defense against arboviruses [24,25,27]. Altogether, these studies pointed to the significant importance of cell replenishing in the midgut epithelium to vector competence. Because of that, we decided to target the Notch pathway through RNAi; to disturb the normal regenerative process of the epithelium. Amongst the proteins involved in this pathway, the ligand Delta was an excellent candidate for RNAi because it is upstream of the Notch signaling pathway and is considered a marker of ISC [19]. Induction of RNAi by injection of dsDelta in adult females led to the silencing of the Notch ligand Delta and resulted on reduced cell division (Fig 4B and 4C), as previously reported by Guo and Ohlstein (2015) in D. melanogaster and by VanDussen et al (2012), in mice. As knockdown of Delta resulted on increased DENV2 viral titers in refractory strain (Fig 4D), this suggested that cell regeneration is also a contributing factor to the modulation of viral infection and consequently to refractoriness. In addition to this result, we pre-treated mosquitoes of the susceptible strain (Rockefeller) with DSS, to induce cell division. Likewise, we found that the increase in mitosis was able to expand refractoriness of these mosquitoes. Our data shows for the first time that the ability to replenish the epithelial differentiated cells, by ISC engagement in tissue regeneration, is an important aspect of the mosquito’s antiviral response in these strains. Furthermore, these results revealed that the involvement of the Notch signaling pathway in midgut cell proliferation is also conserved in A. aegypti. Additional work is required to further determine the involvement of the other cell types and to detail the mechanism by which Delta-Notch signaling interferes in midgut cell proliferation in the midgut of A. aegypti. Very recently, it has been shown that both Delta and Notch transcriptions were induced in midgut of DENV2-primered mosquitoes [50], suggesting that this pathway is important to the vector defense against DENV infection. The role of other pathways previously shown to regulate progenitor cell and differentiation in D. melanogaster and mammalians, such as the Hippo, JAK-STAT and other pathways, may also reveal key connections between intestinal cell replenishment and vectorial competence. The development of specific markers for each A. aegypti epithelial cell type would allow the evaluation of the fate of the new cells produced after ISC division, which could also give important insights on the entire process of midgut regeneration.
The first 24–48 h after ingestion of virus infected blood are considered the most critical for determining vector competence of a given mosquito (reviewed in [51]). Accordingly, we propose that the mitotic events in the early stages of infection (e.g., 24 h after viral ingestion) occur when the number of infected cells is still low and the capacity to eliminate damaged cells prevents viral spreading, and therefore must be effective to limit the infection. The number of mitotic cells of the refractory strain midgut at this initial time point is higher than in the susceptible strain, implicating this as a likely determinant for refractoriness (Fig 4A and 4B). The differences observed in the total number of mitotic cells and in the pattern of recovery between Rockefeller and Orlando strains may suggest more extensive damage in the midgut of the susceptible mosquitoes caused by virus infection. However, the correlation between viral infection progression, cell damage and regenerative responses in the early infection remains to be investigated. In addition, it is also of great importance to investigate the impact of midgut cell renewal on the cellular mechanisms that have been associated with the overcoming of the midgut escape barrier, leading to the dissemination of arboviruses and impacting the vector competence, such as disassembly of basal lamina [52], apoptosis [53] or midgut conduits [54].
In conclusion, our data suggest that the midgut infection by DENV is favored by delayed midgut renewal in a permissive mosquito strain and that refractoriness would be supported, at least partly, by the capacity to promptly activate the ISC division program. At the present time, dengue, chikungunya and zika viruses are widespread across the globe, and the understanding of the multiple factors affecting virus infection within the mosquito is crucial. The fact that faster cell renewal could be related to refractoriness adds up a new factor to be considered among the many determinants of vector competence and opens up the spectrum of the vector physiological events that are important when studying viral transmission. Future research is required to test if other DENV refractory field strains also possess differential tissue homeostatic properties and if a similar mechanism will occur in other arboviral infections. These findings reveal a new path towards a better understanding of vector competence, and may support the development of alternative strategies of virus transmission control.
Finally, these results highlight that the rate of midgut cell renewal should be taken into account when choosing mosquito strains for vector control strategies that use population replacement, such as SIT or Wolbachia based methodologies.
All experimental protocols and animal care were carried out in accordance to the institutional care and use committee (Comitê para Experimentação e Uso de Animais da Universidade Federal do Rio de Janeiro/CEUA-UFRJ) and the NIH Guide for the Care and Use of Laboratory Animals (ISBN 0–309-05377-3). The protocols were approved under the registry CEUA-UFRJ #155/13. All animal work at JHU was conducted in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health (NIH), USA. The protocols and procedures used in this study were approved by the Animal Care and Use Committee of the Johns Hopkins University (Permit Number: M006H300) and the Johns Hopkins School of Public Health Ethics Committee.
The Aedes aegypti (Red Eye strain) were raised at the insectary of UFRJ under a 12-hour light/dark cycle at 28°C and 70–80% relative humidity. The adults were maintained in a cage and given a solution of 10% sucrose ad libitum unless specified otherwise. The A. aegypti (Rockefeller and Orlando strains) were raised at the insectary of JHU under a 12-hourlight/dark cycle, at 27°C and 95% humidity. The adults were maintained in a cage and given a solution of 10% sucrose ad libitum. The adult females were dissected at different times after blood feeding for the experiments.
The mosquitoes were rendered free of cultivable bacteria by maintaining them on a 10% sucrose solution with penicillin (100 u/mL), and, streptomycin (100 μg/mL) from the first day post-eclosion until the time of dissection post blood feeding.
The A. aegypti mosquitoes from the Red Eye strain (four- to seven-days-old) were artificially fed with heparinized rabbit blood. The feeding was performed using water-jacketed artificial feeders maintained at 37°C and sealed with parafilm membranes. The insects were starved for 4–8 hours prior to the feeding. Unfed mosquitoes were removed from the cages in all the experiments.
The oxidative challenge was provided by addition of 500 μM of paraquat (ChemService, West Chester, PA, USA) to the blood meal. As an antioxidant treatment, 50mM of ascorbic acid (neutralized to pH 7.0 with NaOH) was also added to blood. The mosquitoes were orally infected by Serratia marcescens BS 303 strain or Pseudomonas entomophila L48 strain at a concentration of 105 bacteria/mL of blood. Briefly, overnight cultures were used either live or after heat inactivation. Inactivation of P. entomophila was done by heating at 98°C for 1 hour. Live and heat-killed bacteria were all pelleted after OD600 measurements to achieve final concentration of 105 bacteria/mL of blood. The media supernatant was discarded and the cell pellet was resuspended in blood previous to the mosquito feeding. The compound diflubenzuron (DFB) (0.4 g/L), a well-known chitin synthesis inhibitor, was added to the blood meal to prevent the peritrophic matrix establishment [30].
To stimulate ISC proliferation and midgut regeneration [18], the mosquitoes were fed with 1% DSS (dextran sulfate sodium salt 6.5–10 kDa, Sigma, St. Louis, MO, USA) dissolved in 10% sucrose for 2 days before infection. Twelve hours prior to infection, the DSS-sucrose solution was substituted with a 10% sucrose solution to remove residual DSS from the midgut content. The control mosquitoes were fed with 10% sucrose only. The infection with DENV was carried out as described in the following sections.
The quantification of mitosis in whole midgut tissues was performed by PH3 labeling as described elsewhere [55]. Briefly, female adult mosquitoes were dissected in PBS. Midguts were fixed in PBS with 4% paraformaldehyde for 30 minutes at room temperature. Samples were washed in PBS for 2 times of 10 minutes each. Then the tissues were permeabilized in PBS with 0.1% X-100 (for 15 min at room temperature) and blocked in a blocking solution containing PBS, 0.1% Tween 20, 2.5% BSA and 10% normal goat serum for at least 30 min at room temperature. All samples were incubated with primary antibody mouse anti-PH3 (1:500, Merck Millipore, Darmstadt, Germany). After washing 3 times of 20 minutes each in washing solution (PBS, 0.1% Tween 20, 0.25% BSA), samples were incubated with secondary goat anti-mouse antibody conjugated with Alexa Fluor 488 or 546 (Thermo Fisher Scientific, MA, USA) for at least 1 hour at room temperature at a dilution of 1:2000. DNA was visualized with DAPI (1mg/ml, Sigma), diluted 1:1000. The gut images were acquired in a Zeiss Observer Z1 with a Zeiss Axio Cam MrM Zeiss, and the data were analyzed using the AxioVision version 4.8 software (Carl Zeiss AG, Germany). Representative images were acquired using a Leica SP5 confocal laser-scanning inverted microscope with a 20X objective lens. Images were processes using Las X software.
Midguts from insects that were fed on naive blood or blood with DFB were dissected 24 h after feeding and fixed in 4% paraformaldehyde for 3 h. All of the midguts were kept on PBS-15% of sucrose for 12 h and then in 30% sucrose for 30 h. After a 24-h infiltration in OCT, serial microtome 14-lm-thick transverse sections were obtained and collected on slides that were subsequently labeled with the lectin WGA (Wheat Germ Agglutinin; a lectin that is highly specific for N-acetylglucosamine polymers) coupled to fluorescein isothiocyanate (FITC). The slides were washed 3 times in PBS buffer containing 2 mg/mL BSA (PBSB). The samples were then incubated in 50mM NH4Cl/PBS for 30 min; in 3% BSA, 0.3% Triton X-100 PBS for 1 h; and in PBSB solution with 100 mg/mL WGA-FITC (EY Laboratories) for 40 min. The slides were then washed three times with PBSB and mounted with Vectrashield with DAPI mounting medium (Vector laboratories). The sections were acquired in an Olympus IX81 microscope and a CellR MT20E Imaging Station equipped with an IX2-UCB controller and an ORCAR2 C10600 CCD camera (Hammamatsu). Image processing was performed with the Xcellence RT version 1.2 Software.
Midguts from insects that were fed on blood alone or blood with DENV-2 were dissected 5 days after feeding and fixed in 4% paraformaldehyde using the same protocol as for mitotic cell quantification. After the secondary antibody incubation washes, 30 min incubation with phalloidin 1:100 (1uL) in 98uL blocking solution, along with the DAPI (1:100) was done at room temperature protected from light. Samples were washed twice, for 5 minutes (stationary, room temperature, protected from light) in 0.5mL washing solution and then onto slides with VectaShield. Images (z-stack of 0.7 μm slides) were taken on a Zeiss LSM700 laser scanning confocal microscope at the Department of Cell Biology at JHU with a 20X objective lens and processed using Zeiss Zen Black Edition software.
The mosquito midguts were dissected in PBS 24h after feeding and incubated with 50μM of dihydroethidium (hydroethidine; DHE; Invitrogen) diluted in Leibovitz-15 media supplemented with 5% fetal bovine serum for 20 min at room temperature in the dark. The incubation media was gently removed and replaced with a fresh dye-free media. The midguts were positioned on a glass slide, and the oxidized DHE-fluorescence was observed by a Zeiss Observer Z1 with a Zeiss Axio Cam MrM Zeiss using a Zeiss-15 filter set (excitation BP 546/12; beam splitter FT 580; emission LP 590) (Carl Zeiss AG, Germany) [5,56].
For the qPCR assays, the RNA was extracted from the midgut using TRIzol (Invitrogen, CA, USA) according to the manufacturer’s protocol. The complementary DNA was synthesized using the High-Capacity cDNA Reverse transcription kit (Applied Biosystems, CA, USA). The qPCR was performed with the StepOnePlus Real Time PCR System (Applied Biosystems, CA, USA) using the Power SYBR-green PCR master MIX (Applied Biosystems, CA, USA). The Comparative Ct method [57,58] was used to compare the changes in the gene expression levels. The A. aegypti ribosomal S7 gene was used as an endogenous control [59]. The oligonucleotide sequences used in the qPCR assays were S7 (AAEL009496-RA): S7_F: GGGACAAATCGGCCAGGCTATC and S7_R: TCGTGGACGCTTCTGCTTGTTG; Delta (AAEL011396), Delta_Fwd: AAGGCAACTGTATCGGAGCG and Delta_Rev: TATGACATCGCCAAACGTGC.
Two- to three-day old mosquito females (Rockefeller and Orlando) were cold anesthetized and 69 nL of 3 μg/μL dsRNA solution was injected into the thorax. Three days after injection, the mosquitoes were infected with DENV. Mosquito midguts were collected after 24h for real time PCR and after 5 days for mitosis assay or DENV infection analysis. The HiScribe T7 in vitro transcription kit (New England Biolabs) was used to synthesize the dsRNA. The unrelated dsGFP was used as a control, and the silencing efficiency was confirmed through qPCR. To generate dsDelta, the following oligonucleotides (containing the T7 polymerase-binding site) were used:
The DENV-2 (New Guinea C strain) was propagated for 6 days in C6/36 cells maintained in complete MEM media supplemented with 10% fetal bovine serum, 1% penicillin/streptomycin, 1% non-essential amino acids and 1% L-glutamine. The virus titer was determined by plaque assay as 107 PFU/mL [60]. The females were infected through a blood meal containing: one volume of virus, one volume of human red blood cells (commercial human blood was centrifuged and the plasma removed), 10% human serum and 10% 10 mM ATP. Unfed mosquitoes were removed from the cages. The midguts were dissected at 5 days post-blood meal and stored individually in DMEM at -80°C until used.
For DENV-4 (Boa Vista 1981 strain) propagation, the virus was cultivated 6 days in C6/36 cells maintained in Leibovitz-15 media supplemented with 5% fetal bovine serum, 1% non-essential amino acids,1% penicillin/streptomycin and triptose (2.9 g/L) [61]. The virus titer was determined by plaque assay as 107 PFU/mL. The females that were pre-treated with DSS or regular sucrose (control) were infected using one volume of rabbit red blood cells and one volume of DENV-4. The midguts were dissected at 7 days after infection and stored individually in DMEM at -80°C until used.
The plaque assay was performed as previously described [28]. The BHK-21 cells were cultured in complete DMEM media, supplemented with 10% fetal bovine serum, 1% penicillin/streptomycin and 1% L-glutamine. One day before the assay, the cells were plated into 24 wells plates at 70–80% confluence. The midguts were homogenized using a homogenizer (Bullet Blender, Next, Advance) with 0.5mm glass beads. Serial dilutions (10 folds) were performed, and each one was inoculated in a single well. The plates were gently rocked for 15 min at RT and then incubated for 45 min at 37°C and 5% CO2. Finally, an overlay of DMEM containing 0.8% methylcellulose and 2% FBS, was added in each well, and the plates were incubated for 5 days. To fix and stain the plates, the culture media was discarded and a solution of 1:1 (v:v) methanol and acetone and 1% crystal violet was used. The plaque-forming units (PFU) was counted and corrected by the dilution factor.
Unpaired Student’s t-tests were applied where comparisons were made between two treatments or two different mosquito strains, as indicated in the figure legends. Mann-Whitney U-tests were used for infection intensity and chi-square tests were performed to determine the significance of infection prevalence analysis. All statistical analyses were performed using GraphPad 5 Prism Software (La Jolla, United States).
|
10.1371/journal.pgen.1002187 | Adaptations to Endosymbiosis in a Cnidarian-Dinoflagellate Association: Differential Gene Expression and Specific Gene Duplications | Trophic endosymbiosis between anthozoans and photosynthetic dinoflagellates forms the key foundation of reef ecosystems. Dysfunction and collapse of symbiosis lead to bleaching (symbiont expulsion), which is responsible for the severe worldwide decline of coral reefs. Molecular signals are central to the stability of this partnership and are therefore closely related to coral health. To decipher inter-partner signaling, we developed genomic resources (cDNA library and microarrays) from the symbiotic sea anemone Anemonia viridis. Here we describe differential expression between symbiotic (also called zooxanthellate anemones) or aposymbiotic (also called bleached) A. viridis specimens, using microarray hybridizations and qPCR experiments. We mapped, for the first time, transcript abundance separately in the epidermal cell layer and the gastrodermal cells that host photosynthetic symbionts. Transcriptomic profiles showed large inter-individual variability, indicating that aposymbiosis could be induced by different pathways. We defined a restricted subset of 39 common genes that are characteristic of the symbiotic or aposymbiotic states. We demonstrated that transcription of many genes belonging to this set is specifically enhanced in the symbiotic cells (gastroderm). A model is proposed where the aposymbiotic and therefore heterotrophic state triggers vesicular trafficking, whereas the symbiotic and therefore autotrophic state favors metabolic exchanges between host and symbiont. Several genetic pathways were investigated in more detail: i) a key vitamin K–dependant process involved in the dinoflagellate-cnidarian recognition; ii) two cnidarian tissue-specific carbonic anhydrases involved in the carbon transfer from the environment to the intracellular symbionts; iii) host collagen synthesis, mostly supported by the symbiotic tissue. Further, we identified specific gene duplications and showed that the cnidarian-specific isoform was also up-regulated both in the symbiotic state and in the gastroderm. Our results thus offer new insight into the inter-partner signaling required for the physiological mechanisms of the symbiosis that is crucial for coral health.
| The global decline of reef-building corals is of particular concern: they are among the most threatened marine ecosystems. Coral reefs have been decimated over the last several decades, leading to a rapid loss of reef biodiversity and the loss of the resources that humans derive from them. Coral health depends mainly on the trophic intimate association between the animal host and the photosynthetic symbionts (dinoflagellates) residing within its cells. Insights into the physiological mechanisms that underlie healthy as well as stressed (or bleached) corals are thus critical to predict whether they will be able to adapt to and survive climate change. We used a transcriptomic approach to decipher the molecular dialogues between a cnidarian host and its dinoflagellate symbionts. We identified in the host new key pathways that contribute to the symbiotic interaction, as well as specific genetic adaptations.
| The mutualistic symbiosis of anthozoans (Cnidaria), such as corals and sea anemones, with their intracellular dinoflagellate symbionts, mostly of the genus Symbiodinium, forms both trophic and structural foundation of coral reef ecosystems. Anthozoans have a very simple body plan and are composed of two tissue layers, the epidermis and the gastroderm (also called ectoderm and endoderm, respectively [1], Figure S1A). They host their unicellular symbionts, also called zooxanthellae, inside vacuoles (symbiosomes) within the gastrodermal cell layer. Safely localized inside the animal host cells, the photosynthetic symbionts fix large quantities of carbon dioxide. Most of the reduced organic carbon produced is then translocated to the host as mobile compounds, such as glycerol, lipids and amino acids [2]. In return, the host provides suitable conditions for symbiont photosynthesis: inorganic nitrogen, phosphorus and inorganic carbon, as well as a favorable high light environment [2].
This partnership generates many constraints, however, resulting in physiological and cellular adaptations (for review, [3], [4]). For example, the presence of photosynthetic zooxanthellae within the gastrodermal cells requires the host to transport inorganic carbon from the surrounding seawater to the symbionts, a process in which carbonic anhydrases (CAs) are central [5], [6]. To prevent possible cellular damage resulting from photosynthesis-induced hyperoxia, both partners also express a high diversity of antioxidant enzymes, including catalases, peroxidases and superoxide dismutases (SODs), [7]–[9]. SOD and CA isoforms specific to symbiotic anthozoans have been demonstrated [8], [10].
Environmental perturbations such as an increase in seawater temperature may induce dysfunction and collapse of the symbiosis, leading to zooxanthellae loss or so-called “bleaching”, and this phenomenon has led to severe worldwide decline of coral reefs [11]. The establishment and maintenance of this partnership must therefore be dependent on intimate molecular communications between the partners, including recognition and tolerance of symbionts, as well as adaptations for mutual transport and exchange of nutritional resources. A small number of candidate genes for this molecular dialogue has been proposed, including CAs and the cell adhesion protein Sym32 [10], [12], [13]. These have so far been examined by targeted protein analyses (for review, see [14]). Large-scale gene expression studies have tended to highlight a modulation of the host transcriptome, in particular genes involved in cell adhesion, lipid metabolism, cell cycle regulation, or cell death [15]–[17]. While most transcriptomic approaches have been performed in symbiotic cnidarians under thermal stress, imposed in order to understand the molecular and cellular basis of the early events leading to bleaching [18], [19], we rather focused our present experimental approach on two stable states: symbiotic and aposymbiotic individuals.
How do anthozoans maintain a stable partnership with their photosynthetic dinoflagellate symbionts? To decipher the molecular dialogue inferred by the presence of the dinoflagellate symbionts (Symbiodinium clade temperate A) within the sea anemone Anemonia viridis, we compared transcriptomes of symbiotic and aposymbiotic specimens using a symbiosis-dedicated microarray. This oligonucleotide microarray (2,000 features) was developed from the A. viridis 40,000 EST collection [20] and is dedicated to genes potentially involved in symbiosis regulatory pathways.
Our two main goals were to identify gene expression patterns characteristic of: i) the symbiotic and aposymbiotic conditions, ii) the epidermis and gastroderm tissue layers. The main advantage of our model, A. viridis, is it allows efficient separation of the two tissue layers of the animal (epidermis and gastroderm) with minimal cross contamination [8]. Concerted DNA microarrays and quantitative RT-PCR (qPCR) analyses outlined characteristic gene expression signatures for the symbiotic and aposymbiotic states. Among newly identified genes, some appeared to result from anthozoan-specific gene duplications. Analyses of tissue-specific expression demonstrated that most of the host genes involved in symbiotic interactions are preferentially expressed in the gastroderm (i.e. the zooxanthellate tissue layer). We detailed several genetic regulatory pathways involved in dinoflagellate-cnidarian recognition, carbon transfer to the intracellular symbionts and mesoglea constitution. Finally, we propose a model where the aposymbiotic state triggers vesicular trafficking whereas the symbiotic state favors metabolic exchanges between host and symbionts.
The “bleached” phenotype is shared by many stressed symbiotic cnidarians and is the consequence of a massive upstream loss of symbionts and hence their photosynthetic pigments. Most approaches to quantify symbiont loss use manual zooxanthellae counts. Here, we developed a fast and accurate approach using real-time quantitative PCR on total genomic DNA extracts to quantify the relative number of nuclei and hence cells.
Symbiodinium (temperate A clade) nuclear Elongation Factor 2 (EF2), Sucrose Phosphate Synthase (SPS) and Ascorbate Peroxidase (APX) and the A. viridis nuclear Coatomer subunit gamma (COP-γ Regulator of Chromosome Condensation protein 2 (RCC2) and Niemann-Pick disease type C1 (NPC1) gene copy numbers were assessed for total genomic DNA extracts from 5 symbiotic anemones (Sy1–Sy5) and 6 aposymbiotic anemones (AS1–AS6) collected along the Mediterranean coasts around Nice, France (Figure S2) and kept in laboratory culture, as well as in vitro cultured Symbiodinium (CZ) and epidermal tissue fraction (Ep). Although gene locus number per nucleus is assumed to remain constant for a given species, dinoflagellates have been shown to undergo gene specific amplifications [21], [22]. In Amphidinium carterae, the EF2 gene is present in tandem repeat contrary to APX [21]. Whilst the 3 A. viridis gene loci were present in a nearly 1∶1 ratio in all specimens, in Symbiodinium the number of genes was relatively variable from one specimen to another despite the proximity of collection sites to one another. The SPS∶APX ratio was around 1∶1 in all specimens except for AS2 and AS6. EF2∶APX and EF2∶SPS ratios displayed strong polymorphism, likely due to EF2 variable gene amplification (Figure S3A). Gene copy numbers between A.viridis and its symbiont were also measured. Figure 1 shows the relative symbiont to host nucleus ratio for APX/COP-γ SPS/COP-γ and EF2/COP-γ (similar results were obtained with RCC2 and NPC1, Figure S3B). Whatever gene locus used, the pattern of symbiont to host cell ratio was reproducibly similar with ratios between aposymbiotic and symbiotic specimens ranging from 8.2×10−2 (AS4 versus Sy1) to over 10−4 fold (AS6 versus Sy5).
In order to identify gene expression patterns characteristic of symbiosis, the gene expression profiles of symbiotic and aposymbiotic anemones were compared using a symbiosis-dedicated oligoarray. 60-mer oligonucleotides were designed from 2,000 sequences putatively involved in symbiosis selected from the large clustered and annotated A. viridis EST collection [20]. Figure S1 gives an overview of the Gene Ontology (GO) functional annotations of the selected sequences.
Microarray experiments were performed between symbiotic and aposymbiotic anemones, using a dual-dye protocol, one condition being labeled with Cy3, and the second by Cy5. The experimental design of hybridizations is shown in Figure S4. Microarray results were analyzed according to two methods. First, in order to have an overall estimate of the genes differentially expressed between the two states, we treated the 11 individual specimens as two batch categories: symbiotic anemones (Sy1–Sy5) and aposymbiotic anemones (AS1–AS6). Statistical analysis was performed with the limmaGUI package [23] that defined for each gene in each experiment an average fold change (M, corresponding to a log2 ratio between the 2 experimental conditions) and a statistical value (called B, with positive values for the more significant variations). 58 and 78 genes were significantly up-regulated (|M|>0.59 and B>0) in the symbiotic (hereafter called “SY gene set”) and aposymbiotic states (hereafter called “APO gene set”), respectively (Figure 2A, Tables S1 and S2). Functional annotation of these 136 genes using GO terms and statistical analysis by Gossip (Fisher's exact test, p<0.05) showed that the terms “plastid”, “calcium ion binding”, and “protein modification process” were over-represented in the SY gene set, in contrast to “cytosol”, “cytoplasmic membrane-bounded vesicle” and “transcription”, which were over-represented in the APO gene set (Figure S1C).
Second, in order to gain insights about inter-individual variation in gene expression, we compared each individual symbiotic anemone (Sy1 to Sy5) with the batch of 6 aposymbiotic anemones (Apo), and inversely, compared each aposymbiotic anemone (AS1 to AS6) with the batch of 5 symbiotic anemones (Sy). This analysis (Figure 2A) showed unexpectedly high inter-individual variability of differential gene expression patterns within these 136 genes. Figure 2B shows the number of individual anemones for which a given gene is significantly differentially expressed between the SY and APO gene sets. For instance, only 8 SY genes were differentially expressed in all 11 anemones. Conversely, one APO gene (M = −0.63; B = 1.15) encoding for Interferon Regulatory Factor 1 was significantly differentially expressed in only 3 out of 11 anemones. To obtain an indication of the source of variability, we clustered the different anemones according to their expression. Different clustering parameters were tested and showed similar cluster trees; the most representative is shown underneath the heatmap (Figure 2A). Whatever the parameters used, specimens Sy4 and Sy5, AS1 and AS2, and AS4 and AS5 always paired together. Remarkably, even if the anemones were maintained in the same culture conditions at least 3 months before sampling, such co-clustering seemed to correlate with their previous life history: different collection dates/areas for symbiotic specimens, and different bleaching causes for aposymbiotic sea anemones (see Material and Methods and Figure S2). Nonetheless, these variable transcriptomic profiles define the same stable symbiotic or aposymbiotic phenotype.
The “Kern” gene subset (named after the German word for nucleus) was where the remainder of the analysis was focused, and is defined as those genes of cnidarians origin only (excludes genes of unicellular or prokaryotic origin) found in the SY or APO gene set, and differentially expressed in at least 8 out of the 11 specimens. Table 1 lists the representative 19 and 20 Kern genes found in the SY and APO gene sets respectively. A blast homology search against the Nematostella vectensis genome database and the Uniprot generalist database, and searches for specific protein signatures (signal peptide, trans-membrane domains or others domains) allowed the genes to be named and functions and cellular localizations assigned (Table 1). Among the different functional categories found, cell adhesion proteins (8/39) were the most represented, a process indeed expected to play a key role in signaling events between partners. A total of 7 other genes were involved in metabolism (4 SY and 3 APO). The four SY genes were specifically involved in fatty acid metabolism, indicating that fatty acid metabolism is likely a preponderant metabolic pathway of the symbiotic condition. Although all genes listed in Table 1 would merit further investigation, only several selectively targeted genes are detailed in this study (see below).
In order to obtain support for the functional implication of some of the Kern genes, we monitored the expression of genes of interest directly after thermal stress, which is known to disrupt symbiosis. Anemones Sy3–5 were subjected to an 8°C temperature increase, and the expression of CA2-c, CA2-m and NPC2-D dropped by around 3 fold after 24 h and 48 h (Figure 3). This result is interesting, considering their up-regulation in symbiotic individuals. This immediate response to environmental stress precedes symbiosis breakdown and thus strengthens support for the possible role of these 3 genes in endosymbiosis in A. viridis.
As the two tissue layers can be separated in A. viridis, it is a powerful biological model for studying tissue-specific gene expression. We used two experimental designs to compare gastrodermal versus epidermal gene expression profiles (Figure S4). Firstly we directly hybridized the cDNAs of the two tissues against each other on the same array, and secondly we compared each tissue sample to the aposymbiotic AS6 reference and thus defined tissue expression profiles by transitivity. Only the genes with at least 1.50 fold increase in transcript abundance (M>0.59) in one tissue in both experiments were assigned as differentially expressed (Table 2, Tables S1 and S2). As expected, the expression of most Symbiodinium genes was restricted to the gastroderm. The Venn diagram in Figure 4 summarizes distribution of the 1,715 cnidarian-specific genes (prokaryote and zooxanthellae genes were excluded from the analysis) according to their gastrodermal (Ga), epidermal (Ep), and preferential symbiotic (SY) versus aposymbiotic (APO) expression. We further confirmed preferential expression of several genes in the different tissues using real-time quantitative PCR (Figure 5). Taken together, microarray and qPCR results showed that many more A. viridis genes were preferentially expressed in the gastroderm (71%) than in the epidermis (29%). More interestingly, among the 17 genes which were both preferentially expressed in a given tissue and differentially regulated under symbiotic/aposymbiotic status, the large majority (12) were in fact up-regulated within the gastroderm of symbiotic anemones (i.e. the zooxanthellate tissue). This suggests that the presence of symbionts directly modulates the gene expression of their hosting gastrodermal cells. In addition, out of these 12 genes, 9 belong to the Kern subset of genes, supporting the importance of this gene set in inter-partner communication and regulation (Figure 5 and Table 2). Hence, we have identified different categories of genes potentially involved in sequential aspects of symbiosis, and showed for the first time that the animal tissue-specific molecular response to the presence/absence of zooxanthellae was restricted almost entirely to preferential expression of genes within the compartment hosting the symbionts. These results support the crucial role of the gastroderm in this symbiotic interaction.
The gastroderm-specific gene expression disclosed interesting hallmarks. Carbonic anhydrases are primordial enzymes necessary for the transport of inorganic carbon through biological membranes. Two isoforms were identified in our library and had increased transcript abundance in zooxanthellate anemones. Based on their respective protein signatures and blast homology, one isoform (Av_CA2-c, Kern # 3) is cytoplasmic whereas the other isoform (Av_CA2-m, Kern # 5) probably localizes to the outer plasma membrane, since it contains a signal peptide sequence as well as a trans-membrane and GPI anchor domains at its NH2 and COOH termini, respectively. The cytoplasmic CA2 was equally distributed in the two compartments, but the membrane-anchored CA2 was principally expressed in the gastroderm (Figure 5). Another interesting finding on the tissue distribution of gene expression was that collagen biosynthesis was broadly under the control of the gastroderm. In cnidarians, fibrillar collagen genes were recently shown to be much more represented than first expected from bilaterian evolutionary comparisons, with 8 different genes present in the genome of N. vectensis [24]. In A. viridis, we found 13 different cDNAs corresponding to portions of the N. vectensis homologs. Additionally, we monitored the expression of 5 genes regulating the post-translational processing of collagen synthesis. Interestingly, 14 of these cDNAs showed increased transcript abundance in the gastroderm (Table 2). The mesoglea, the acellular layer between epidermis and gastroderm, which mostly consists of collagen, may partly originate from the gastroderm in A. viridis. However, none of these genes showed differential expression between aposymbiotic and symbiotic states, except for Prolyl-4-hydroxylase alpha. This enzyme is a key chaperone for the biosynthesis of collagen, catalyzing the hydroxylation of proline residues of procollagen chains necessary for their correct three dimensional folding [25]. Two different isoforms were identified in our sequence library and the expression of one of them (Kern # 14) was specifically enhanced in zooxanthellate anemones. Thus, in addition to showing that the gastroderm synthesized much of collagen fibers, our results infer that Symbiodinium may exert, directly or indirectly, post-translational control on collagen synthesis and modulate the formation of the mesoglea in A. viridis.
During the course of our analysis, we noticed that some of the genes involved in symbiosis (Kern genes) had multiple paralogs in our dataset. Such was the case with the previously described MERP gene family [20], but the same held true for other Kern genes including Niemann Pick type C2 (NPC2), Calumenin, Sym32 and C3 Complement (C3) families (Table 2). We conducted phylogenetic analyses on the NPC2, Calumenin and Sym32 gene families in order to assign an evolutionary origin for each member. A. viridis isoforms present in our dataset were compared to their homologs in N.vectensis and other representative eukaryotes with a complete genome sequence. Both maximum likelihood (Figure 6 and Figure 7A) and Bayesian (Figures S5, S6, S7) methods gave very similar trees.
In vertebrates, Calumenin belongs to the CREC gene family, which encompasses 5 members (CAB45, ERC-55, Reticulocalbin 1 and 3, and Calumenin) [26]. From phylogenetic analysis across metazoans (Figure 6A), 3 groups of homologs can be defined; the CAB45 homologs, the ERC55 homologs and the CALU (Calumenin, Reticulocalbin 1 and 3) homologs. Cnidarian homologs are found within these 3 groups, however additional cnidarian homologs (NvCALUa,c,f–h, AvCALUa,c) are also present outside the CALU and ERC55 groups, representing cnidarian specific CALU gene duplications.
NPC2 (Figure 6B) is a single copy gene in human, nematode, sea urchin, Placozoa and yeast and is independently duplicated in ascidian and fish. Two copies of the NPC2 gene were found in sea anemones. Phylogenetic analysis supports a gene duplication in Anthozoans, with NvNPC2a and AvNPC2a representing the orthologs of the chordate NPC2, and NvNPC2b and AvNPC2-D defining an anthozoan specific duplication.
In the symbiotic sea anemones A. viridis and Anthopleura elegantissima, Sym32 is composed of two adjacent FasI domains (Figure 7A). In A. viridis, we identified two FasI-containing proteins: Sym 32 and the related periostin gene (PN), which is conserved across metazoans (only cnidarians and human are shown). AvSym32 and AvPN are composed of 2 and 4 FasI domains, respectively. In N. Nematostella, there are also two comparable FasI-containing proteins; both are composed of 4 FasI domains. One corresponds to the PN homolog (NvPN) and the second, which we named Nv2Sym, is similar to a tandem duplication of the AvSym32 sequence with 2×2 FasI domains. In human, two related FasI-containing proteins are characterized: the cognate Periostin (HsPN) and the “Transforming growth factor-beta-induced protein ig-h3” precursor (HsBGH3), both with 4 FasI domains. As proteins did not have the same number of FasI domains, we conducted a phylogenetic analysis on the alignment of all single FasI domains in order to gain insight into the domain evolution of this protein family (Figure 7A and Figure S7B). Human and cnidarian PN genes likely evolved from a common ancestor while Hs_PN and Hs_BGH3 would have duplicated after the Cnidaria - Bilateria separation. Consequently, AvSym32 and Nv2Sym are cnidarian-specific genes. Whether a Sym32 version containing only 2 FasI domains is specific to symbiotic anthozoans remains to be clarified.
Comparison of the expression of the A. viridis CALU, NPC2 and sym32 homologs showed that the isoform which was more highly expressed in the symbiotic state was also preferentially expressed in the gastroderm, whereas the other members were ubiquitously expressed (Table 2). For instance, AvCalu-a (Kern # 1) was the most up regulated gene in the symbiotic condition and localized in the gastroderm, whereas expression of both the other isoforms (AvCalu-b and AvCalu-c) was neither different between symbiotic versus aposymbiotic conditions nor between the tissue layers (Table 2). The same held true for the two isoforms of NPC2 (AvNPC2-D (Kern # 2) versus AvNPC2-a) and for Sym32 (Kern # 4) versus the Periostin homologs (Table 2). Most interestingly, the isoform that was differentially expressed was always member of the Kern gene set.
Thus, analysis of N. vectensis gene families identified cnidarian-specific gene duplications (NvCALUa,c,f–h, NvNPC2b, Nv2sym) and A. viridis ortholog analysis showed that among the cnidarian-specific isoform, the Kern genes (AvCALUa, AvNPC2-D and AvSym32) were up-regulated both in symbiotic state and in the gastroderm.
Many anthozoans rely on photosynthetic endosymbionts (mostly Symbiodinium sp.) to grow in oligotrophic environments. The symbiotic relationship involves regulatory crosstalk between partners that allows the association to persist. This interpartner communication includes: i) the recognition of the partners, ii) the ability of symbionts to colonize host cells without being rejected by the host immune system, iii) the regulation of symbiont population, and iv) adaptations for mutual transport and exchange of nutritional resources [27]. In order to identify the genes potentially involved in the molecular dialog supporting this endosymbiosis, we used a microarray approach to compare the gene expression profiles of 11 A. viridis anemones representative of the symbiotic and the aposymbiotic states. We identified a subset of A. viridis genes, which we named “Kern”, characteristic of the symbiotic (SY) and aposymbiotic (APO) states in the sense that they were preferentially expressed in symbiotic or aposymbiotic anemones. Consequently, the Kern genes are potentially important candidate genes for the maintenance of the symbiosis process. Since in A. viridis, the two tissue layers can be separated, we were also able to assign tissue-specific gene expression. Many genes or gene functions that were expressed more highly in symbiotic anemones were also expressed more highly in the gastroderm, the symbiont hosting tissue. Functional annotation associated with sub-cellular localization of the Kern gene products allowed us to draw a map of the 39 genes differentially regulated in the symbiotic and aposymbiotic states (Figure 8).
One of the first conclusions from our present model is that many genes that are present in higher abundance in the aposymbiotic state are associated with active vesicle trafficking, both in the endocytotic and secretory pathways, as opposed to the many membrane-bound proteins associated with cellular recognition and adhesion coded by genes in higher abundance in the symbiotic state (Table 1 and Figure 8). Under aposymbiotic conditions, transcription of several proteins controlling trafficking was markedly enhanced. These include TRAPPC2 and TRAPPC10 (Kern # 22 and 23), which are members of the TRAPP complexes that regulate endoplasmic reticulum (ER) to post-golgi vesicular trafficking [28], thus attesting for more active de novo biosynthesis in the aposymbiotic anemone than in the symbiotic host. As part of the same line of evidence, the highly differentially expressed gene AP-2 sigma-1 (Kern # 38) codes for a key protein in the formation of clathrin-coated vesicle [29], implying increased endocytic trafficking. Among other aposymbiotic marker genes, HVCN1 and SLC30a (Kern # 26 & 36) are two ion channels shown to be present in phagosomes and late endosomal/lysosomal vesicles, respectively [30], [31]. Of note, our dedicated array comprises several oligonucleotide probes for various rab mRNAs (including rab1a, 2a, 7a, 8a and 11a). Rab proteins have been shown to control vesicular trafficking in various species including symbiotic anemones [32]–[34] where the symbionts have been shown to detour host membrane trafficking resulting in failure of host lysosomes fuse with the symbiosomal membrane [34]. However, in our array results, none of the Av_rab genes were differentially expressed. This does not preclude a role of these Rab proteins in A. viridis symbiosis, as expression regulation may be at the translational level. Interestingly, the fact that vesicular trafficking appears to be much more active in aposymbiotic A. viridis corroborates with heterotrophy, which mostly rely on predation, digestion and de novo synthesis, as opposed to autotrophic symbiotic anemones, in which 60% of carbon flux is provided by symbionts [35]. Numerous studies have characterized glycerol, lipids and amino acids as the major mobile compounds transferred from the symbiont (reviewed in [2], [4]). At the level of metabolism pathways, Phosphoenolpyruvate carboxykinase (PEPCK-C; Kern # 28) had enhanced expression in aposymbiotic conditions. This enzyme catalyzes the rate-controlling step of gluconeogenesis when pyruvate is used as the substrate. Two metabolic pathways can initiate gluconeogenesis: the PEPCK-C dependent pathway, which is fuelled by pyruvate (product of glycolysis), and the PEPCK-C independent pathway, which is fuelled by glycerol [36]. As glycerol is acquired from the symbiont, PEPCK-C is indeed expected to be down-regulated in the symbiotic state.
In zooxanthellate anemones, lipids are the second most abundant transferred mobile compounds from the symbiont [2], [4] therefore requiring up-regulation by the host of several key regulatory enzymes for fatty acid metabolism. Although mobile compounds have been identified using radioactive tracers [2], compound transfer pathways from the symbiont remain to be characterized. Four Kern genes were involved in lipid processing pathways, LSD-2 (lipid storage droplet), SDR12 (terpenoid metabolism), 2-hydroxyacyl-CoA lyase 1 and NPC2. These 4 genes were up-regulated in the symbiotic state (Kern # 2, 13, 15, 17 in Table 1). 2-hydroxyacyl-CoA lyase 1 is responsible for fatty acid alpha-oxidation, a modification specific to fatty acids synthesized from chloroplastic organisms [37]. NPC2 (AvNPC2-D, Kern # 2) is among the most up-regulated genes, substantiating another study which noted that a NPC2 transcript was up-regulated in symbiotic Aiptasia pulchella anemones [38]. NPC2 acts in synergy with NPC1 in the transport of sterols (cholesterol) using the late endosomal/lysosomal system [39]–[41]. Noticeably, NPC1 and NPC2 are preferentially expressed in the gastroderm (Table 2). Based on their tissue expression profile and symbiosis-related regulation, NPC1 and NPC2 appear to be major candidate genes for the transport in A. viridis of sterol compounds produced by Symbiodinium.
Another notable difference between aposymbiotic and symbiotic states is that aposymbiotic-specific Kern genes showed no tissue expression preference. This is in marked opposition to the numerous symbiosis-related Kern genes expressed principally in the gastroderm. Hence, expression of the latter is both limited to the symbiont-containing cells and correlated with the presence of Symbiodinium. Although the mechanisms underlying such control are yet to be characterized, several molecular dialogs between host and symbiont can be envisaged.
The Complement C3 (C3) is a precursor protein involved in adaptive immunity in vertebrates and also in cellular recognition, inflammatory processes and phagocytosis in invertebrates [42], [43]. C3 thus appears central to non-self response in metazoans and homologs have been identified throughout the metazoans, including cnidarians [42], [44]. In the coral Acropora millepora, one C3-like protein has been shown to localize in close association with the symbiosome in the gastroderm layer, supporting a role in recognition of Symbiodinium [44]. In sea anemones, there are three C3 isoforms in the non-symbiotic N. vectensis genome and a minimum of four different C3 isoforms were identified in the A. viridis EST dataset (PG and CS, personal communication). We monitored the expression of two of these: AvC3-1 and AvC3-2. Although AvC3-1 expression was mainly restricted to the gastroderm and may well represent the A. millepora functional homolog, no differential expression was observed between symbiotic and aposymbiotic anemones (validated by qPCR, Figure 5). On the other hand, AvC3-2 (Kern # 37) was expressed in both tissue layers, but was strongly repressed in the presence of symbionts. The functional divergence between the different C3 isoforms and their relative participation in Symbiodinium tolerance remains to be determined. However, based on their respective expression profiles in A. viridis, it is conceivable that recognition of and response to Symbiodinium may be carried out by different C3 paralogs.
Another recognition process that we uncovered in our experiment is related to the vitamin K-dependent (VKD) γ-carboxylation of Sym32, under the control of the Calumenin protein. In human, Calumenin proteins are composed of 6 to 7 Ca2+ binding EF hand domains. They are principally present in the Endoplasmic Reticulum (ER) due to specific ER retention motifs at their COOH termini [26], [45], [46]. Their differences in function are unclear, but they are associated with Ca2+ dependent processes, especially with post-translational VKD γ-carboxylation of several proteins [26]. The increase of intracellular vitamin K (a cofactor produced from plant) activates two proteins conserved throughout metazoans: Vitamin K1 2,3-Epoxide Reductase (VKOR) and γ-Carboxylase. The latter recognizes carboxylase recognition sites (CRS) in newly synthesized proteins and adds a CO2 group to adjacent glutamic acid (Glu) residues resulting in the production of γ-carboxyglutamic acid (Gla)-containing proteins [47], [48]. VKD proteins have been mainly explored in mammals and include bone marrow proteins such as Osteocalcin, Matrix Gla Protein and Periostin (PN) which are involved in bone formation [49]. The vitamin K cycle is itself under the negative control of Calumenin [26], [46] (Figure 7B). In A. viridis, the Calumenin homolog AvCALU-a (Kern # 1) was the most up-regulated gene of the symbiotic state detected in the microarray experiments (Table 1). It is also preferentially expressed in gastrodermal cells. It contains an ER retention motif and is thus expected to down-regulate the VKD γ-carboxylation in the ER of zooxanthellate A. viridis cells.
Sym32 was first characterized by Weis and colleagues as a symbiosis-specific protein that is over-represented at the perisymbiotic membrane of the zooxanthellate anemone A. elegantissima (AeSym32), but also present at the surface of gastrodermal vesicles in the aposymbiotic state [13], [50]. In A. viridis, the Sym32 ortholog (AvSym32, Kern # 4) is up-regulated by more than 40 fold in the symbiotic state and is principally expressed in the gastroderm (Figure 5, Table 1 and Table 2). Both AeSym32 and AvSym32 are composed of two fasciclin I (FasI) domains (Figure 7), functionally associated with cell-cell recognition and thus potentially involved in anemone-zooxanthellae interaction [13], [51]. Importantly, in both sea anemones, the first FasI domain of Sym32 is highly similar in sequence to the first FasI domain of PN, and contains a sequence motif very similar to the CRS domain of PN characterized in vertebrates [49]. Sym32 is thus an implicit substrate for the γ-Carboxylase in sea anemones. As part of the same line of evidence, a 2D immunoblot of zooxanthellate A. elegantissima extract hybridized with an anti-Sym32 antibody showed the presence of two spots of 32 KDa, one at PI = 7.9 and the other at PI = 8.2 [51]. Such a PI difference for the same protein could very well correspond to difference between Glu and Gla containing Sym32 proteins since γ-carboxyglutamic acid decreases the PI of a protein (e.g. [52]). However, analysis of freshly isolated Symbiodinium (FIZ, where most of the peri-symbiosomal membrane - of host origin - remains attached around the isolated Symbiodinium cell) of A. elegantissima showed the disappearance of the PI = 7.9 spot in favor of the PI = 8.2 spot [13]. These experiments infer that the Glu-Sym32 protein would localize to the perisymbiotic membrane. Since γ-carboxyglutamic acid residues have been shown to modify the three dimensional structure and Ca2+ binding affinity, Glu-Sym32 and Gla-Sym32 would show different ligand properties [53]. Moreover, up-regulation of Calumenin shown in our work suggests an inhibition of γ-carboxylation which would favor the production of Glu-Sym32 in zooxanthellate A. viridis cells (Figure 7B). Thus, the vitamin K-dependent γ-carboxylation of Sym32, with its effect on interpartner recognition and the symbiotic process, is definitively a pathway that should be investigated.
The photosynthetic symbionts are separated from the surrounding seawater by several host membranes: membranes of the epidermal cell layer, the collagenous basal membrane, gastrodermal cells, and perisymbiotic vesicles. The main source of inorganic carbon (Ci) for photosynthesis is seawater bicarbonate (HCO3−), which implies transport of exogenous inorganic carbon through these layers of animal tissue [5], [54]. In seawater (pH 8.2), most inorganic carbon is in the form of HCO3−, a form that needs carrier-mediation to cross membranes, and that is not readily converted to CO2 in the absence of enzymatic action [55], [56]. The currently accepted model for external Ci uptake by the host involves an H+-ATPase acidifying the boundary layer where bicarbonate is converted to CO2 by an external (likely membrane-bound) carbonic anhydrase isoform [3]. The uncharged CO2 molecule then diffuses into the epidermal cell following the concentration gradient created by the extrusion of H+ in the external medium. Once in the animal cytoplasm, CO2 is equilibrated with HCO3− according to the intracellular pH by another CA isoform, which prevents back-diffusion of CO2 (for review see [3]). The mechanism of transport of Ci through the other membranes to the symbionts is currently debated (for reviews, see [5], [57], [58]). However, previous works have highlighted the role of a CA localized on the perisymbiotic [59] or algal membrane [60].
According to this model, carbonic anhydrases are crucial enzymes for carbon supply to symbiont photosynthesis. In the present study, the expression of two different carbonic anhydrases (Av_CA2-c and Av_CA2-m) was monitored. Both show highly enhanced expression in symbiotic specimens compared with aposymbiotic ones (4 and 2.9 fold, respectively), suggesting that both isoforms could be involved in the symbiosis/metabolic exchanges between partners. This result is consistent with the relevant work of Weis and her collaborators [6], [59], [61], showing that enzyme activity and transcript quantity are higher in symbiotic than in non-symbiotic specimens of sea anemones A. pulchella and A. elegantissima.
AvCA2-c isoform, the cytosolic isoform, is equally expressed in both tissue layers. We suggest that AvCA2-c catalyzes the intra-cellular reversible hydration/dehydration of CO2 into HCO3− to facilitate the transport of CO2 through the membranes and cells (Figure 8). AvCA2-m is a membrane-bound isoform specifically expressed in the gastrodermal layer of anemones where symbionts are located. This membrane-bound isoform can be located either on the plasma membrane of the gastrodermal cells, or on the perisymbiotic membrane surrounding the symbiont. In the first case, AvCA2-m would favor the transfer of CO2 from one cell layer to another by preventing back-diffusion of CO2 through membranes. In the second case, this isoform would catalyze the final conversion of HCO3− into CO2 for photosynthetic needs. It should be noted that, since in N. vectensis at least six different isoforms of carbonic anhydrases have been identified (AM and D. Zoccola, personal communication), other A. viridis CA isoforms are expected to contribute in the Ci transport from seawater to Symbiodinium.
The previous model of Ci transport assumed that CO2 crosses membranes by diffusion through concentration gradients between both sides of the cell plasma membrane. For most of the past century, gas molecules such as CO2 were presumed to cross biological membranes merely by diffusing through the lipid phase. This view was challenged recently by studies demonstrating the permeability of water channel aquaporins and certain Rh-family members to CO2 [62]–[64]. The physiological function of those channels for CO2 transport seems to be particularly important since, for instance, one Rh protein (RhAG) accounts for up to 50% of the CO2 transport of human red blood cells [65]. In A. viridis, we identified two RhAG isoforms, one of which (Av_RhAG1) is preferentially expressed in the gastroderm of zooxanthellate anemones (Table 2). We suggest that these proteins could have a role within the holobiont to facilitate CO2 uptake, possibly in conjunction with the membrane-bound CA2. In human erythrocytes, RhAG and CA2 are part of the same Band 3 multiprotein complex involved in anion exchanges [66], [67]. It is worth mentioning that a role of Av_RhAG1 in NH4 transport is equally valid, as in the cnidarian-dinoflagellate symbiosis, ammonium resulting from host metabolism is not excreted into the surrounding water but is immediately re-assimilated by the algae, either through diffusion or by a transporter, and then recycled [4], [68].
Within the set of Kern genes up-regulated in the symbiotic condition, we noticed that some of the genes with a proposed function in symbiosis, including Sym32, Calumenin, and NPC2, were expressed as 2 or more related copies. Phylogenetic analysis allowed the conclusion to be made that there were cnidarian-specific gene duplications. Remarkably, our expression results in A. viridis showed that these cnidarian-specific duplicates were both preferentially expressed in the gastroderm (hosting zooxanthellae) and in the symbiotic condition.
Such cnidarian-specific gene duplications could correlate with the amenability of various cnidarians to have accepted photosynthetic endosymbionts during evolution. However, these gene duplications are not restricted to symbiotic anemones (e.g. N. vectensis). Their selective advantage with regard to symbiosis therefore remains to be determined, together with the origin of endosymbiosis in cnidarians: was photosynthetic symbiosis acquired or lost in various branches of the phylum? Moreover, at the transcriptional level, we have shown for several cases that expression of one isoform among the duplicated gene copies was specifically up-regulated both in the gastroderm and in the presence of symbionts. Thus, we suggest that these neofunctionalizations would be associated with the physiological constraints of endosymbiosis and would be tuned to the presence of zooxanthellae by transcriptional control. It is not known how such control is exerted or whether such gene regulation is restricted to symbiotic cnidarians. Comparative tissue expression of the different orthologs in non-symbiotic cnidarians, such as N. vectensis, would provide insight into the adaptive origin of symbiosis.
One of our most unexpected results was that outside the subset of Kern genes, many differentially expressed genes were up- or down-regulated in only few individual anemones, despite the fact that anemones were in a fixed symbiotic or aposymbiotic state before sampling. Similar inter-individual variable expression profiles have already been quoted in two studies comparing the response of A. millepora corals to environmental changes [15], [69].
On the other hand, several groups studying symbiosis breakdown in diverse cnidarian species showed that different cellular mechanisms were observed during loss of zooxanthellae, including apoptosis, necrosis, exocytosis, or phagocytosis [11]. Such variability in the cellular processes involved may reflect distinct causes of bleaching, i.e. different responses to variable environmental changes. Indeed, bleaching can be caused by a multitude of environmental stressors including changes in seawater temperature, salinity, ultraviolet radiation, increased sedimentation, nutrients and pollutants [70]. In response to different stresses, the host-associated microbiota is greatly and specifically modified within the whole holobiont (i.e. the community composed of cnidarian host, dinoflagellates and associated microbes) [71].
In the case of the Mediterranean symbiotic cnidarian used here, Symbiodinium belongs to the same clade temperate A. However, our results on the gene copy numbers of only 3 genes (EF2, SPS and APX) showed unexpected high polymorphism in the Symbiodinium hosted by anemones collected from neighboring location. In addition, we showed that the abundance of zooxanthellae per host cell is variable within our set of symbiotic sea anemones. Thus, variation in the associated microbiota or in Symbiodinium sub-clades may have stable extended effects on the host gene expression profile. This underlines that, although the term “bleached” defines one visual phenotype resulting from various responses, it can result from diverse expression profiles, and probably from various conditions of aposymbiosis. Indeed, not all bleached cnidarians die.
A. viridis can be sampled at different time points without harm to the anemone, and is therefore an excellent model organism for kinetics experiments. Additionally, since sea anemones are non-calcifying anthozoans, they allow the study of symbiosis-associated processes outside the cross-regulatory pathways of mineralization found in corals. It would be of interest, using the same array technique, to monitor the kinetics of the specific gene expression response to one or combination of different environmental stressors. In addition to giving a transcriptional map of the gene-specific response to stress, it would highlight whether different bleaching expression profiles are fixed in time or whether they tend to stabilize at a unique Kern profile after an adaptation period. Moreover, the unique amenability of A. viridis to separate the ectoderm from the endoderm will permit us to partition the scope of the environmental response in the cellular layer hosting the symbionts (gastroderm) and in the tissue in direct contact to the milieu (epidermis).
Mediterranean sea anemone specimens, Anemonia viridis (Forskål, 1775), were collected in five locations on and around the French Riviera (Figure S2): Antibes (Salis and Croutons sites), Villefranche-sur-Mer, Monaco and Menton. A total of 5 symbiotic and 6 aposymbiotic specimens were used in this study.
Symbiotic specimens Sy1 and Sy3–5 were collected from Antibes Croutons (Figure S2), Sy2 was collected from Antibes Salis. These anemones were maintained for several months in seawater aquaria at 17.0±0.5°C with weekly water renewal. A metal halide lamp (HQI-TS 400 W, Philips) provided light at a constant saturating irradiance of 250 µmol m−2 s−1 on a 12/12 h light/dark cycle. Naturally-occurring aposymbiotic animals were sampled from the public aquaria of the Oceanographic Museum of Monaco and originated from different sites in Monaco (AS1 and AS2) or Menton (AS3). The three other aposymbiotic specimens were obtained after either a thermal stress-induced bleaching (AS4, AS5, collected in Antibes-Crouton) or a treatment with the catalase inhibitor aminotriazol (AS6, collected in Villefranche-sur-Mer, [7]). These latter 3 anemones were maintained in the dark (necessary to keep them bleached) for a minimum of 2 months. Both symbiotic and stress induced aposymbiotic anemones were fed twice a week with frozen Artemia salina.
Sy3, Sy4 and Sy5 specimens were later subjected to an 8°C-temperature increase (17 to 25°C) to assess imposed stress response. Tentacles were sampled over a 48 h kinetic period (t0, t24h and t48h) and then subjected to RT-qPCR experiments.
Specimen sampling was always at 10:00 (2 hours after light start) to avoid circadian effect of gene expression fluctuations.
Over 2,000 genes were selected from the A. viridis clustered and annotated EST collection [20] for their putative participation in symbiotic processes. Genes were selected by keyword search in their functional annotation for matches with the following terms: intracellular transport, metabolic processes (lipids, proteins), signal transduction, organelle organisation and biogenesis, response to stress, trans-membrane, apoptosis and cell death. Two thousand 60-mer oligonucleotides were designed (Eurogentec, Belgium) and printed on slides in triplicate (Eurogentec, Belgium). A luciferase oligonucleotide was also spotted as an external control. This A. viridis Oligo2K version 1.0 oligoarray is fully described under platform record GPL10546, stored in the Gene Expression Omnibus (GEO) at NCBI (http://www.ncbi.nlm.nih.gov/geo).
cDNA synthesis and labelling was performed from total RNA using the ChipShot Direct labelling and Clean-up system (Promega), according to the manufacturer's instructions. Five ng of a Lux mRNA exogenous standard was added to each mRNA sample before labelling. RNA quality was evaluated using the Agilent Bioanalyzer 2100 and quantified on a ND-1000 Spectrophotometer (NanoDrop). Then, 500 ng of each labeled sample were mixed in Hi-RPM hybridization buffer (Agilent), and incubated on the array overnight at 47°C. Slides were scanned after post-hybridization washes, using a GenePix 4200A scanner (Axon Instruments). Data acquisition and quality control were performed using Genepix Pro software.
Dye-swap experiments were performed for each hybridization condition. Two different sets of hybridizations were performed (Figure S3): i) symbiotic versus aposymbiotic specimens and ii) epidermis versus gastroderm tissues. The same AS6 extract was used as the sample reference in all experiments. Experimental data and associated microarray designs were deposited in the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) (http://www.ncbi.nlm.nih.gov/geo/) under SuperSeries record: GSE22375 and platform record GPL10546.
Background (positive offset of 50) was evaluated according as described in [72]. The data were normalized by the print-tip loess method (within-array normalization) and by quantile method (between-array normalization) using the LimmaGUI package from Bioconductor [23]. Means of ratios from all comparisons were calculated for each gene, and B test analysis was done using LimmaGUI. Differential gene expression was determined by the Bayesian statistical method (B value), with a cut-off of zero for the B value and a log ratio |M|>0.590 as significant. All normalized data sets were registered in the GEO database under the accession number GSE22375.
Cluster 3.0 software [73] was used to estimate the hierarchical clustering between individual anemone array results. The following parameters (complete, average, and centroid linkage) were tested and gave similar results.
Specific primers amplifying around 100 bp were designed using the software Primer3 [74]. The primer sequences used in this study are listed in Table S3. Amplicon specificity for either A. viridis or Symbiodinium was tested against epidermal (extracts without zooxanthellae) or CZ (zooxanthellae culture) extracts, respectively. Expected length of the amplicons was checked by agarose gel electrophoresis after regular PCR amplification. Primer efficiencies were determined using standard curve analysis with a 10-fold dilution series of pooled cDNA from both control and treated samples (data not shown), and ranged from 1.8 to 2. The qPCR products were sequenced (Macrogen Inc, Korea) and all matched the expected product identities.
cDNAs were prepared using SuperScriptII reverse transcriptase (Invitrogen) and a mixture of oligodT and random primers, according to the manufacturer's instructions. Transcript level quantification was performed using the SYBR green fluorescence method and a Light Cycler 480 (Roche). The PCR conditions were as follows: 1× SYBR green mix (LC480 SYBR Green Master Mix, Roche), 100 nM primers and 2.5 ng of cDNA in a total volume of 15 µl. Each sample was run in triplicate using the following PCR parameters: 94°C for 10 min, followed by 40 cycles of 15 s at 94°C, 20 s at 60°C and 15 s at 72°C, then a dissociation curve step (60 to 95°C) to confirm the absence of non-specific products. The dissociation curves showed a single amplification product and no primer dimers.
Several control genes were chosen based on the microarray results as a whole (most stably expressed genes, in all tested conditions), and the expression stability of ten putative control genes was evaluated using the GeNorm software [75]. A reliable normalization factor was calculated based on the expression level of the most stable control genes. The control genes finally selected in this study are RPLP0, RCC2, and COP-γ. Expression levels of target genes were normalized using the normalization factor described above and the results given as expression relative to the aposymbiotic specimen (AS6) value as calibrator (reference). The significance of the results was tested using t-tests (software Jump 5.1, Cary, USA). Results were considered statistically significant when P<0.05.
The relative abundance of zooxanthellae in A. viridis cells was quantified using a real-time quantitative qPCR method. For each organism, three nuclear genes were chosen from available datasets: COP-γ (Coatomer subunit gamma), RCC2 (Regulator of Chromosome Condensation protein 2) and NPC1 (Niemann Pick type C1) for A. viridis, and EF2 (Elongation factor 2), Sucrose phosphate synthase (SPS) and Ascorbate peroxidase (APX) for Symbiodinium clade temperate A. Primer design was performed using Primer3 and the same parameters described above; the primer sequences are given in the Table S3. Standard curves were generated with six logarithm dilutions of corresponding cloned sequences. Results are expressed as relative quantification of Symbiodinium nuclei (or nuclear genes) to A.viridis nuclei.
Sequencing of either A.viridis library clones or newly cloned cDNAs was performed by Macrogen Inc.
Signal peptides and Trans-membrane domains were predicted using SignalP and TMHMM, respectively, from the CBS prediction servers (www.cbs.dtu.dk/services). GPI anchors were predicted with PredGPI (http://gpcr.biocomp.unibo.it/predgpi). Other Domains were inferred from PFAM (pfam.sanger.ac.uk). Blast analysis tools used were from NCBI (blast.ncbi.nlm.nih.gov).
Functionnal annotation of the A. viridis EST dataset (GO annotations) has been performed using Blast2GO [20]. Statistical assessment of annotation differences between SY-genes (test group) and APO-genes (reference group) was performed using the Gossip package [76], that employs Fisher's Exact Test to estimate the significance of associations between two categorical variables.
Sequence alignments were performed using MultAlin [77] and ClustalW with Blosum62 default parameters. Alignments were optimized manually using the 2 computer generated alignments as a model. Using the segment of the sequence alignment conserved in all sequences (bordered by the bar above alignments in Figures S5, S6, S7), the best-fitted substitution model was evaluated using ProtTest [78]. Using parameters indicated in the Figure legend of each alignment tree, a Maximum Likelihood tree was determined using phyML and branches support were calculated using aLRT [79]. Alternatively, Bayesian analysis using MrBayes 3-1.2 (mrbayes.csit.fsu.edu) was conducted with the following settings: fixed rate amino acid model was set to mixed (prset aamodelpr = mixed) and proportion of invariable sites model was combined with the Gamma model (lset rates = invgamma).
|
10.1371/journal.pgen.1004094 | Nannochloropsis Genomes Reveal Evolution of Microalgal Oleaginous Traits | Oleaginous microalgae are promising feedstock for biofuels, yet the genetic diversity, origin and evolution of oleaginous traits remain largely unknown. Here we present a detailed phylogenomic analysis of five oleaginous Nannochloropsis species (a total of six strains) and one time-series transcriptome dataset for triacylglycerol (TAG) synthesis on one representative strain. Despite small genome sizes, high coding potential and relative paucity of mobile elements, the genomes feature small cores of ca. 2,700 protein-coding genes and a large pan-genome of >38,000 genes. The six genomes share key oleaginous traits, such as the enrichment of selected lipid biosynthesis genes and certain glycoside hydrolase genes that potentially shift carbon flux from chrysolaminaran to TAG synthesis. The eleven type II diacylglycerol acyltransferase genes (DGAT-2) in every strain, each expressed during TAG synthesis, likely originated from three ancient genomes, including the secondary endosymbiosis host and the engulfed green and red algae. Horizontal gene transfers were inferred in most lipid synthesis nodes with expanded gene doses and many glycoside hydrolase genes. Thus multiple genome pooling and horizontal genetic exchange, together with selective inheritance of lipid synthesis genes and species-specific gene loss, have led to the enormous genetic apparatus for oleaginousness and the wide genomic divergence among present-day Nannochloropsis. These findings have important implications in the screening and genetic engineering of microalgae for biofuels.
| Microalgae are promising feedstock for biofuels. However, the diversity, origin and evolution of oil-producing microalgal genomes in general, and those of their oleaginous traits in particular, remain poorly understood. We present five new genomes of the oleaginous microalgae Nannochloropsis spp. that allow genus-, species- and strain-level genomic comparison. With each Nannochloropsis genome encoding approximately 6,562–9,915 genes, a core genome of ca. 2,700 genes and a large pan-genome of >38,000 genes were found. The genomes share key genetic features such as gene dose expansion of selected nodes in lipid biosynthesis pathways. Evidence of horizontal gene transfers, primarily from bacteria, was found in most of these nodes. However, the eleven type II acyl-CoA:diacylglycerol acyltransferase genes (DGAT-2),the highest gene dose reported among known organisms, likely originated from three ancient genomes of the secondary endosymbiosis host and the engulfed green and red algae; they were strictly vertically inherited in each of the Nannochloropsis spp. Thus, multiple genome pooling and horizontal genetic exchange have underlain the enormous genetic makeup underlying TAG production in present-day Nannochloropsis.
| Microalgae represent a promising source of biomass feedstock for fuels and chemicals because many species possess the ability to grow rapidly and synthesize large amounts of storage neutral lipids in a form of triacylglycerol (TAG) from sunlight and carbon dioxide. They can be cultivated on non-arable land with non-potable water and waste streams (e.g., flue gases and wastewaters) and thus pose little competition to food crops while providing environmental benefits [1]. However, understanding of the divergence and evolution of oleaginous traits and the underlying evolutionary forces and molecular mechanisms in microalgae remains elusive [2].
Nannochloropsis is a genus of unicellular photosynthetic microalgae in the class Eustigmatophyceae, ranging in size from 2–5 µm and widely distributed in marine, fresh and brackish waters. They are of interest as a potential feedstock for fuels and high-value products because they tolerate broad enivronmental and culture conditions while growing rapidly and producing large amounts of TAG and eicosapentaenoic acid, a high-value polyunsaturated fatty acid [3]. A homologous recombination–based gene transformation system was recently established in Nannochloropsis [4], making trait improvement in this organism possible for overproduction of biomass or desirable products.
Here we present a comparative analysis of six genomes of oleaginous Nannochloropsis spp. that includes two N. oceanica strains (IMET1 and CCMP531) and one strain from each of four other recognized species: N. salina (CCMP537), N. gaditana (CCMP526, which was previously reported [5]), N. oculata (CCMP525) and N. granulata (CCMP529) (Figure 1A; Figure S1; Figure S2; Table S1A, S1B). Moreover, for N. oceanica IMET1, the diversity of transcripts was mapped to support gene prediction by sequencing cDNA libraries using 454-based long reads. Furthermore, transcript dynamics were measured via a two-condition (control condition and nitrogen starved condition), three time-point temporal series of transcriptomes during TAG accumulation using Illumina-based short-reads (Text S1). Integration of phenotypic, genomic and transcriptomic data across a Nannochloropsis phylogeny provided new insights into the molecular mechanisms driving the diversity and evolution of these oleaginous microalgae.
The genome sizes of the six oleaginous Nannochloropsis species and strains range from 25.38 to 32.07 Mb (Figure 1A; Table 1). For strain IMET1, the nuclear, chloroplast and mitochondria genomes are 31.36 Mb, 117.5 Kb and 38 Kb, respectively, totaling 31.5 Mb. Pulse-field gel electrophoresis on total IMET1 DNA confirmed the genome size and indicated the presence of 22 chromosomes (Figure S3A, S3B). For IMET1, 9,754, 126 and 35 protein-coding genes were predicted in the nuclear, chloroplast and mitochondrial genomes, respectively (Table 1). Among the nuclear genes, 93.4% (9,111) were covered by mRNA-Seq data (defined as >80% of the transcribed region mapped by at least 10 reads; Table S1C, S1D, Text S1).
These Nannochloropsis genomes are all relatively compact (Table S2; [5], [6]), much smaller than that of the model green microalga Chlamydomonas reinhardtii (121 Mb; [7]). The IMET1 genome features a higher coding potential (52.1%) than the diatom Thalassiosira pseudonana (32.7%; [8]), which has a similar genome size. Mobile elements can be prevalent in algae [e.g. T. pseudonana harbors 238 long terminal repeats (LTRs) totaling 1.56 Mb], but they are rather limited in IMET1, as only 26 LTRs (24.3 Kb in total), along with several DNA transposons (864 bp in total), are present in the genome without transposases (Table S2). The relative paucity of mobile elements appears to be one shared feature of the six Nannochloropsis strains (Table 1)
Genomic diversity and divergence defining microalgal genera, species or strains are largely unknown [9]. A whole-genome phylogeny of Nannochloropsis (Figure 1A) was constructed from 1,085 single-copy-orthologous groups identified from the six genomes, which is consistent with the 18S-based phylogeny (Figure S2). Among the five Nannochloropsis species, N. granulata and N. oculata have a recent common ancestor and are clustered with the two N. oceanica strains. Among the 1,085 single-copy orthologous groups, 628 (61.7%) exhibited congruent phylogenies with the whole-genome phylogeny. The mean Ka/Ks of 0.08 calculated from these candidate phylogenetic markers in the nuclear genomes was higher than in the chloroplast genomes (0.031) and in the mitochondrial genomes (0.064). Among these candidate markers, 25 genes exhibited sequence variations large enough to differentiate each of the species and strains (density of inter-species SNP at 20–40% and intra-species over 1%), but allowed for the design of consensus flanking PCR primers (Dataset S1). Those with the highest resolution included cytochrome P450, btaA, plastid ribosomal protein S1 and transaldolase etc., which represent novel phylogenetic markers that are more sensitive than 18S or ITS sequences (0.16% and 0.52% in intra-species SNP density, respectively) in strain-typing of Nannochloropsis.
Between any two genomes among the six Nannochloropsis strains, 35% of protein-coding genes (ranging from 2.6% between the two N. oceanica strains to 66.4% between IMET1 and N. salina CCMP537) were not found in the other genome on average, despite >98% similarity in full-length 18S rDNA. This places their inter-species genome divergence higher than the green algae studied and their intra-species divergence comparable to E. coli and yeast (Figure 1B). Therefore, the Nannochloropsis pan-genome, as defined by the six strains, consists of at least 38,000 protein-coding genes, along with a relatively small pool of Nannochloropsis core genes (e.g., 2,734 genes in IMET1) that are shared by the six strains (Figure 1C, Text S1). Most (93.2%) of these core genes have blast hits in NCBI non-redundant (NR) database, of which 94% were functionally annotated. The core genes mostly encode proteins involved in DNA, RNA, and protein synthesis and modification, transporters, signal transduction and central metabolic pathways (Figure 1D; for functional classification based on molecular function and cellular component, see Figure S4). The accessory genes, referring to those missing in at least one strain, mainly encode (i) central metabolism such as carbohydrate, lipid, energy, and nucleotide and amino acid metabolism (which are overlapped with the core genes), (ii) secondary metabolism and N-glycan biosynthesis (which are complementary to the core genes, and (iii) unknown functions (Figure S5).
There were 164–1,513 genes that were strain-specific among the six genomes. In contrast to the 2,734 Nannochloropsis core genes, of which 96.7% were supported by our mRNA-Seq reads, 11.0% (18) of the 164 IMET1-specific proteins lacked such supports, suggesting the possible presence of pseudogenes or false positives in gene prediction. Among the IMET1-specific genes with mRNA support, 94.5% were putative novel genes without any known homologs (Blast hits) in the NCBI NR database. It is possible that some of them might have horizontally transferred from unsequenced species. Among strain-specific genes with functional annotations, most were involved in responses to freezing in N. oculata CCMP525, N. granulata CCMP529 and N. oceanica CCMP531. In N. gaditana CCMP526, transporters were prevalent, while in N. salina CCMP537, no significant enrichment was found in any processes (Figure S6).
Correlation analysis revealed that the core and accessory genes exhibited different sequence and transcriptional features under the experimental conditions tested (Text S2). The accessory genes tend to be under lower purifying pressure while lower transcriptional levels (Text S2; Figure S7; Figure S8), supporting a link between sequence evolution and transcriptional activity [10], [11], [12].
To probe the link between the accessory genes and divergence of the genomes, protein-coding genes in the six Nannochloropsis were classified into different groups based on the number of strains in which they were present (thus those present in all the six strains were part of the Nannochloropsis core). The most prominent group included the genes shared by four of the strains, in which the majority (97.3%) were found in the phylogenetically closely related species, i.e., N. oceanica (two strains), N. granulata and N. oculata. The absence of these genes in the other two species explained the small number of Nannochloropsis core genes (Figure S9A). These genes might have been present in the common ancestors of heterokonts and later lost in N. salina and N. gaditana, as >60% of them were found in other heterokonts (e.g., diatoms, Ectocarpus and other non-photosynthetic heterokonts such as Phytophthora). The functions supported by these genes were similar to those of core genes, with oxidation-reduction, transmembrane transport and protein-related metabolism being dominant. This does not support the presence of functional bias in the gene loss events (Figure S9B).
To seek the cause of the structural divergence among the Nannochloropsis genomes, we clustered all encoded proteins based on their sequence similarity. Among sequenced plant and algal genomes, large paralogous groups are common, e.g., 217 F-box family protein genes in Arabidopsis [13] and 51 Class III guanylyl and adenylyl cyclase genes in Chlamydomonas [7]). However, Nannochloropsis spp. appear to have adopted a strategy in which paralogous groups are less biased in size, i.e., they formulate a large number of relatively small paralogous groups (Figure 2, Text S1). There are 4,263, 4,325, and 7,171 paralogous groups in Thalassiosira, Chlamydomonas, and N. oceanica IMET1, respectively. The top 15 largest paralogous groups in each Nannochloropsis genome range in size from two to seven genes, with a median value of three to four (Figure 2). For example, the largest paralogous group in IMET1 consists of 11 genes (mainly in metabolic process), which is in a sharp contrast with T. pseudonana (46 genes; protein modification process; [8]), Cyanidioschyzon merlae (23 genes; DNA metabolic process; [14]) and C. reinhardtii (150 genes; protein modification process; [7]). As genes from different origins might exhibit relatively low sequence conservation and thus fail to formulate a paralogous group, the reduced sizes of paralogous groups in the Nannochloropsis genomes might result from the integration of multiple genome resources, which is consistent with the proposal that heterokonts originated from multiple secondary endosymbiosis [15]. This observation also suggests that strain-specific gene sequence duplication was relatively rare in Nannochloropsis. On the other hand, it is also possible that Nannochloropsis spp. have adapted to their environment via a strategy of frugality in proteome structure, with paralogous protein-coding genes either emerging less frequently or many of them being lost.
Among the 8,992 homologous groups from the six Nannochloropsis genomes (by OrthoMCL [16]; based on amino acid sequence similarity), 1,731 included the genes from all six strains. However, 2,312, 1,515, 1,551 and 1,653 groups included genes from two, three, four and five of the strains, respectively, and thus were “mosaic” groups as they included genes from only a subset, but not all, of the strains. Furthermore, 230 groups were specific to one of the six strains, with 4 to 151 such groups in each strain. The large number of mosaic paralogous groups (7,031 or 78.2% in total) could explain the large size of the Nannochloropsis pan-genome, although the numbers of genes and gene groups could be over- or under-estimated due to the presence of alternative splice forms or artifacts of genome assembly.
Despite their high structural diversity, each of the six Nannochloropsis genomes exhibits functional features that underlie their oleaginous phenotypes. There is significantly higher gene enrichment for cellular lipid metabolism in each genome than in C. reinhardtii (Figure 3, Dataset S2). In all or most of the Nannochloropsis strains, the subcategories of lipid metabolism are enriched, including glycerolipid metabolism, phospholipid metabolism, lipopolysaccharide metabolism and lipid modification. Metabolic pathways enriched in Nannochloropsis also include organic acid metabolism, precursor generation and sulfur compound metabolism. Genes related to stress response, including responses to DNA damage stimulus, DNA repair and cold stress response, were also enriched in several Nannochloropsis strains. However, the number of genes involved in phosphorus metabolism and cellular macromolecule metabolism was significantly lower in each Nannochloropsis strain than in C. reinhardtii (Figure 3). Thus, the enrichment of gene doses in lipid metabolism pathways and stress response-related pathways appears to be a shared feature of Nannochloropsis genomes and likely underlies their advantageous oleaginous and environmental tolerance traits.
In the lipid biosynthesis pathway (the de novo biosynthesis of fatty acids and TAG), a prominent expansion in gene copy number in particular reaction nodes was observed as a shared feature among the six Nannochloropsis strains, despite a genome size only one-fourth of C. reinhardtii. Such enriched genes include those encoding ketoacyl-ACP synthase (KAS, four to five in each Nannochloropsis strain vs. three in C. reinhardtii), acyl-ACP thioesterase (acyl-ACP TE, five vs. one), long-chain fatty acyl-CoA synthetase (LC-FACS, 11–12 vs. seven), phosphatidic acid phosphatase (PAP, five vs. one), and the last two acyltransferases: lysophosphatidyl acyltransferase (LPAT, seven to eight vs. one) and diacylglycerol acyltransferase (DGAT) (Figure 4). Multiple copies of KAS proteins were found in each Nannochloropsis strain for the assembly of type II fatty acid synthases. In addition, six bacterial type I fatty acid synthase genes, each with several conserved functional domains, were identified (compared to only one in C. reinhardtii); phylogenetic analysis revealed that these genes are closely related to polyketide synthases (Figure S10), yet they might be involved in fatty acid synthesis [17].
Notably, such expansion in gene dose was not ubiquitous along the TAG pathway. For many of the nodes, the gene doses are comparable to those in C. reinhardtii (Figure 4B). These nodes include the acetyl-CoA carboxylase (ACCase), MCAT, KAR, HAD in fatty acid biosynthesis, GPAT in TAG assembly, and other membrane lipid biosynthesis-related enzymes (such as the MGD and DGD in galactolipid synthesis, SQD in sulfolipid synthesis, BtaA and BtaB in betaine lipid synthesis and EPT in phosphatidylethanolamine synthesis). The expansion of gene dose for the selective steps highlights their crucial roles in channeling carbon flux into TAG synthesis and might be considered a “genomic signature” of oleaginousness.
To probe the evolutionary forces expanding the TAG biosynthesis gene repertoire in Nannochloropsis, we carried out a phylogenomic analysis to investigate the horizontal gene transfer (HGT) events in N. oceanica IMET1 genome (Text S1; [18]). We identified 99 HGT candidates (Figure S11A; Dataset S3), accounting for 1.0% of nuclear genes. Among them, the most abundant functions encoded (in terms of GO Slim terms in biological process) included biosynthetic process, small molecule metabolism, cellular nitrogen compound metabolism and lipid metabolism (Figure S11).
HGT appeared to have played an important role in the evolution of oleaginousness loci in these organisms. Totally nine HGT candidates (15.3% of total lipid biosynthesis genes, much higher than average percentage of HGT presence in nuclear genome) were inferred in most of the nodes with increased gene doses, such as KAS, enoyl-ACP reductase (ENR), acyl-ACP TE, LC-FACS and PAP (Figure 4A, Figure S12, Figure S13). PAP catalyzes the Mg2+-dependent dephosphorylation of phosphatidic acid (PA) to yield diacylglycerol (DAG) and Pi. Both PA (via CDP-DAG) and DAG can enter phospholipid synthesis, and DAG is the direct precursor of TAG. Thus, PAP may control the direction of carbon flux and affect overall cellular lipid synthesis [19]. Five genes encoding PAP enzymes were found in each Nannochloropsis strain: three were conserved in eukaryotes, while the other two were clustered with the bacteria, indicating a bacterial HGT origin (either one HGT followed by gene duplication or multiple horizontal transfers; Figure S12F, Figure S13F). The two horizontally transferred PAP genes exhibited higher transcriptional levels than the eukaryotic ones. The presence of multiple prokaryotic PAP genes suggests complex mechanisms to regulate the substrate preference for the synthesis of various classes and species of lipids. Among the ENR genes in each Nannochloropsis strain, two likely originated by HGT from bacteria into the common ancestor of the six Nannochloropsis strains (Figure S12C, Figure S13C; suggested by the absence of other heterokonts in the bacterial ENR clade), which were then inherited by each of the Nannochloropsis strains.
The most prominent example of gene dose expansion is DGAT, which catalyzes the last step of TAG synthesis from DAG and acyl-CoA [20] and includes DGAT-1 and DGAT-2 [21]. There are 12–13 DGAT in each Nannochloropsis strain (one to two DGAT-1 and 11 DGAT-2), representing the highest dose among known genomes (Figure 5A). In contrast, only six and four DGAT are present in C. reinhardtii and the diatom T. thalassiosira, respectively, and even fewer in some other green algae and heterokonts (Figure 5A). In IMET1, all the DGAT-1 and DGAT-2 were transcriptionally active [FPKM (Fragments Per Kilobase of exon per Million mapped reads) >1.0]. Phylogenetic analysis of DGAT from selected bacteria, fungi, algae and higher plants revealed extraordinary evolutionary diversities of all 74 DGAT in the six Nannochloropsis strains (Figure S14). Several observations were apparent. (i) The partition of DGAT-1 and DGAT-2 might have occurred early, likely before the primary endosymbiosis event or even earlier. (ii) The copy number of DGAT-1 was lower (1–2) and less variable than that of DGAT-2, which is consistent in a wide range of organisms from bacteria to land plants. (iii) A similar degree of DGAT-2 dose expansion was observed in all six Nannochloropsis strains (Figure 4B). Moreover, for each of the 11 DGAT-2 identified in each strain, the orthologs in the other five strains were all identified and clustered into a phylogenetic group (Figure S14); the sequence identity between orthologous gene pairs was >98% between the two N. oceanica strains, >80% among N. oceanica, N. oculata and N. granulate, and >65% between N. oceanica IMET1 and the outmost N. gaditana. These results suggested the stable inheritance of DGAT-2 genes in Nannochloropsis evolution. In contrast, DGAT-1 might have experienced species-specific gene loss. For example, no counterparts of DGAT-1B in IMET1 were found in N. salina and N. gaditana despite a high degree of conservation of this gene in the other four strains. (iv) The 11 DGAT-2 genes in IMET1 exhibited relatively low intra-genome pairwise identity (averaging 18%), and each was grouped into a separate paralogous group with its orthologs from the other five Nannochloropsis strains, indicative of distinct and divergent phylogenetic origins of DGAT-2 in Nannochloropsis. Two of the DGAT-2 genes in IMET1 (DGAT-2F and DGAT-2D) exhibited a relatively high protein sequence similarity (identity at 51%), suggesting that the two genes might be derived from a gene duplication event in the Nannochloropsis lineage (Figure S14). However this individual case of suspected gene duplication cannot account for the expanded dose of DGAT-2 genes in IMET1.
To infer the origin of these genes, a comprehensive phylogenetic analysis was carried out among all species with genomes and ESTs available in several public databases (Text S1; [22]). DGAT-2C showed a phylogeny with strong affiliation with the red algae C. merolae and formed a sister group with those from other chromalveolates (Figure S15A, Figure S16A). The most plausible explanation for such strong links between Nannochloropsis and red algal DGAT is a red-algae derivation of DGAT-2C through endosymbiotic gene transfer (EGT) in the secondary endosymbiosis event (which was proposed as the evolutionary mechanism through which the common ancestor of chromalveolates acquired chloroplasts from a red algae-related endosymbiont [23]). On the other hand, four DGAT-2 genes, including DGAT-2A, DGAT-2B, DGAT-2I and DGAT-2G, are clustered with their counterparts from green algae as well as other chromalveolates (Figure S15B–E, Figure S16B–E), suggesting a green algal origin of these DGAT-2 genes. This is consistent with the hypothesis of a green algae related endosymbiont residing in the common ancestor of chromalveolates [15]. Three (DGAT-2A, DGAT-2I and DGAT-2C) of the above five red-lineage (red algae derived) and green-lineage (green algae derived) DGAT-2 genes were predicted to harbor chloroplast targeting signals, supporting their ancestral derivation from the endosymbionts. The higher dose of green- than red-lineage DGAT-2 in each of the Nannochloropsis strains suggests a more significant contribution of the green lineage to the oleaginous traits of modern Nannochloropsis. Furthermore, phylogenetic trees of the other six DGAT-2 genes did not exhibit unambiguous relationships with those from red or green algae and are thus referred to as “unresolved.” It is possible that several of these genes originated from the secondary host [24], as four (DGAT-2D, DGAT-2E, DGAT-2F and DGAT-2H) of the six genes were predicted to be located in the endoplasmic reticulum (ER) or cytosol. Thus, the observed sequence divergence of the 74 DGAT genes (eight type I and 66 type II) in the six Nannochloropsis genomes mainly resulted from their diverse origins from the red- or green-algae–related endosymbionts (through EGT) and the secondary host (Figure 5B).
Phylogenetic evidence also supports a green endosymbiont origin for one gene encoding MCAT (s00247.g6828) in fatty acid biosynthesis (Figure S15F, Figure S16F). Thus, the diverse evolutionary origin of the Nannochloropsis DGAT-2s and one of the other lipid synthesis genes has underlain their massive genetic pools and likely contributed to the extraordinary capacity for TAG synthesis in present-day strains.
In addition to their diverse origins, differentiation in selective pressure appeared to underlie the sequence divergence of DGAT and other members of lipid-related gene families. DGAT genes in Nannochloropsis were generally under strong purifying selective pressure (Ka/Ks typically under 0.1). However, higher Ka/Ks ratios were observed in the red-lineage DGAT-2C of 0.11 (Figure S14). No significant difference in the ratio was found between the green-lineage and secondary-host–originated DGAT-2. Furthermore, DGAT-2C with the highest Ka/Ks ratio was among the DGAT-2 genes with the lowest transcriptional level under normal growing conditions, while DGAT-2J with the lowest ratio was one of the most transcribed DGAT-2 (second only to DGAT-2A). These findings add further support to the negative correlation between transcriptional level and selective pressure in the evolution of Nannochloropsis genes.
Dramatic enrichment of glycoside hydrolase (GH) genes accompanied by a reduced pool of glycoside synthase genes (as compared to C. reinhardtii) was also observed in each of the Nannochloropsis genomes. C. reinhardtii harbors seven starch synthase genes for starch production and two 1,3-β-glucan synthase genes; however, each Nannochloropsis encodes just one 1,3-β-glucan synthase gene (which might convert glucose into the polysaccharide chrysolaminarin or laminarin), and no starch synthase genes were found. Conversely, 48–49 GH genes were found in each strain (Dataset S4), with a gene dose per Mb of genome 6–7 fold higher than that of C. reinhardtii (27 GH genes). These Nannochloropsis genes were from 13 different GH families, dominated by GH2 (13 members), GH9, GH3 and GH1 families with over four members. Surprisingly, there were only three genes in the GH16 family, which specifically hydrolyzes the glycosidic bond of 1,3-β-glucan, while GH16 was the dominant group in C. reinhardtii, with five members. In IMET1, 91.7% of the 48 GH genes were transcriptionally active at each of the time points under both N-replete and N-depleted culture conditions (FPKM>1.0; Text S1). Among them, 16 exhibited significant variations at the transcriptional level under N-depleted conditions (10 with increased transcription), including two members of the GH2_C family and one GH17 gene with a significant increase in transcription (fold-change >1.5) from 3 h and 6 h after the onset of N-depletion and one GH9 gene down-regulated under the same conditions. The monosaccharides released from GH that catalyzed hydrolysis of polysaccharides may be used in glycolysis to produce acetyl-CoA and ATP for fatty acid synthesis.
Among the 48 GH genes in each strain that were conserved among the six genomes, 16 were inherited from the common ancestor of heterokonts, as their homologs were found in the diatoms Phytophthora and Ectocarpus. Another five GH genes were likely to have originated from bacteria via HGT. Among these, three GH8 genes inferred to be horizontally acquired from cellulose-digesting Clostridium-like bacterium were absent in other sequenced unicellular algae. The remaining 27 GH genes were phylogenetically closest to homologs in animals, insects or multicellular fungi, such as the nine putative cellulase genes that were most similar to those in the nematode Pristionchus, indicating HGT events with donors being Nannochloropsis-like organisms [25]. In addition, N. granulata and N. salina each possessed one strain-specific GH gene that might have been introduced after their speciation.
Microalgae, which are primarily unicellular, aquatic and photosynthetic eukaryotes, are phylogenetically diverse. They are responsible for over 45% of our planet's annual net primary biomass [26]. The Nannochloropsis genomes studied here, one of the first such datasets for microalgae, reveal the nature and degree of genome divergence and dynamics at the strain, species and genus level. They could serve as an initial framework for genome-wide association studies, while the genome-derived nuclear gene markers should be useful for highly sensitive typing of strains.
The genomes of the six oleaginous Nannochloropsis strains presented here are of relatively small size and high coding potential and many fewer mobile elements compared to many previously sequenced microalgae [9]. The large size of the Nannochloropsis pan-genome can be partially traced to the large number of mosaic paralogous groups, which further suggests a significant degree of species-specific gene loss during Nannochloropsis evolution. On the other hand, the small core genome size and the large number of mosaic homologous gene clusters among the Nannochloropsis spp. suggest that, as one moves down the tree of life for stramenopiles, the number of shared genes reduces quickly and is replaced by lineage-specific gene gains and losses. The core genes generally exhibit lower Ka/Ks ratio but higher transcriptional levels than non-core genes, suggesting their roles in shaping the evolution of microalgal genes. Our findings, together with observations in yeasts [10], [27], revealed a link that is conserved in unicellular eukaryotes in terms of gene function, selective pressure, transcriptional level and gene essentiality.
Despite the high sequence diversity of protein-coding genes, the six Nannochloropsis genomes shared a genus-level oleaginousness signature that included enrichment of selective lipid biosynthesis genes and certain glycoside hydrolases that potentially shift carbon flux from storage carbohydrate to TAG synthesis. It is quite remarkable that these gene expansions have occurred despite their significant genome shrinkage relative to other microalgae such as C. reinhardtii. Different mechanisms have underlain the emergence of the signature. Multiple-genome pooling was particularly evident for the 11 DGAT-2 in each strain, which were all transcriptionally expressed during TAG synthesis and apparently originated from at least three ancient genomes: (i) the engulfed green algae, (ii) the engulfed red algae and (iii) the host cell in the secondary endosymbiosis.
Chromalveolates include both photosynthetic (e.g. diatoms and Eustigmatophyceae that include Nannochloropsis) and non-photosynthetic taxa (e.g., Phytophthora). The chromalveolate hypothesis suggests that the common ancestor of Chromalveolates originated via an eukaryotic host (i.e., the secondary host) engulfing a red alga (as the secondary plastid) in an ancient secondary endosymbiosis event [23]. The presence of a large number of “green genes” in the diatom nuclear genomes has been interpreted as evidence of a cryptic prasinophyte-like secondary endosymbiosis before the red algae intake [15]. Though confounded by potential sampling bias against red algae and artifacts in phylogenetic analysis [28], this hypothesis was supported by the 172 membrane transporter genes showing potential origins from green or red algae in a relatively strict phylogenomic analysis [22]. Moreover, genomes of the cryptophyte alga Guillardia theta and the chlorarachniophyte alga Bigelowiella natans also contain hundreds of genes with a phylogenetic affiliation to red or green algae [24]. Our search of DGAT-2 in publicly available red algae genomes (and ESTs) returned one DGAT-2 each from Cyanidioschyzon merolae, Galdieria sulphuraria [18] and Porphyridium purpureum [29]. The paucity of DGAT-2 in red algal genomes and the distinct features of these genes in the six Nannochloropsis genomes (the greatly expanded copy number, large pair-wise sequence divergence, rare gene duplication events, and absence of mobile elements or evidence for HGT in each of the DGAT-2 loci) suggested multiple-genome pooling as the cause for the massive DGAT pool in Nannochloropsis spp. These findings also provided additional support for the existence of a green algae–associated secondary endosymbiosis in the evolutionary history of chromalveolates.
Furthermore, among the six Nannochloropsis strains, the inheritance of each DGAT-2 was highly conserved in that no strain-specific duplications or losses were found for any DGAT-2 in each of the six strains, and the genes have been under strong negative selection. In contrast, diatoms such as T. pseudonana (believed to have also experienced the multiple secondary endosymbiosis [15]) encode many fewer DGAT-2; only four DGAT-2 were identified, and all were predicted to be from the green algae–related endosymbiont and the secondary host, with none from red lineage. The absence of Thalassiosira genes in certain gene-phylogeny clusters (e.g., the red-lineage DGAT-2C) in the diatom, in contrast to the presence of these genes in Nannochloropsis and many other heterokonts, suggests the loss of DGAT-2 in diatoms. Thus, such strict inheritance and stable maintenance of the large reservoir of DGAT-2 from multiple lineages seem to be Nannochloropsis-specific. It also suggests the essentiality of each DGAT-2 and its possible functional complementarity in the cell.
In addition, HGT primarily from bacteria were found in the majority of the gene dose-expanded lipid synthesis nodes and in many glycoside hydrolases. In the red alga Galdieria sulphuraria, 5% of protein-coding genes were acquired from bacteria and archaea via HGT, which forged its adaptation to a thermophilic and metal-rich environment [18]. The HGT events in Nannochloropsis likely reflected an organismal adaptation to a niche that favored oleaginousness and glycoside hydrolysis.
Therefore, the multiple-genome pooling and horizontal genetic exchange from bacteria, together with the selective inheritance of lipid synthesis genes and species-specific gene loss, might have underlain the enormous genetic apparatus for oleaginousness and led to the structural divergence and functional conservation observed among present-day Nannochloropsis. In many organisms, other mechanisms such as gene and genome duplications may play an important role in supplying new genetic materials for organismal adaptation [30] and have been frequently proposed as drivers of the emergence of particular traits in bacteria [31], [32], fungi [33], [34], plants [35] and animals [36]. Thus, the extraordinary origin and evolution of oleaginous traits in Nannochloropsis have important implications in the selection and genetic engineering of such traits in these and other microalgae of economic interest.
All genomic data for this study, including the assembled genomes and mRNA-Seq data, were deposited at NCBI. The BioProject accessions for assembled genomes were: PRJNA202418 for N. oceanica IMET1, PRJNA65107 for N. oculata CCMP525, PRJNA65111 for N. granulata CCMP529, PRJNA65113 for N. oceanica CCMP531 and PRJNA62503 for N. salina CCMP537. The mRNA-Seq data were deposited at SRA under SRP032930.
Five new Nannochloropsis genomes were sequenced in this work (Table 1; Table S1). For Nannochloropsis oceanica IMET1, both shotgun sequencing data and paired-end data with different pair distances from 454 Titanium and Illumina GAIIx were collected. Newbler (Roche) was used for initial assembly. Gap-filling and scaffold-building were performed with Illumina data, followed by manual manipulation and sorting of contigs. Genes were predicted by combining the ab initio predictions with predictions based on mRNA-Seq read alignments (387K aligned cDNA reads from a Roche 454 Sequencer) by AUGUSTUS (v2.5). For each of the other four Nannochloropsis strains (Table S1), paired GAIIx reads were assembled using Velvet with specified insert sizes. The previously published genome sequence of N. gaditana CCMP526 [5] was downloaded from http://Nannochloropsis.genomeprojectsolutions-databases.com/. Gene models of each of the six genomes were predicted using two different ab initio gene predictors (AUGUSTUS and GeneID) followed by a combination of gene models using EVidenceModeler (EVM) with a 1∶1 weight ratio. For all strains, predicted protein-coding genes were annotated via searching against three databases: the NCBI NR and KEGG databases by BlastP, and the Gene Ontology database by InterProScan. GO terms were mapped to the GO slim hierarchy proposed by the GO consortium by a customized script (http://www.bioenergychina.org/fg/d.wang_scripts/).
For collecting the transcriptomics datasets underlying TAG production, N. oceanica IMET1 was cultivated in f/2 liquid medium [37] with 4 mM NO3− under continuous light at 50 µmol photons m−2 s−1. Mid-logarithmic phase algal cells were inoculated in nitrogen-replete and nitrogen-depleted conditions, respectively. Total RNA were collected at 3, 6 and 24 h after each inoculation and pooled together for full-length cDNA sequencing in 454 Titanium. The data produced were subsequently used for gene prediction. Furthermore, total RNA from each of the aforementioned control (nitrogen-replete) and nitrogen-starvation conditions along the time points of 3, 6 and 24 h after the onset of nitrogen depletion (six samples under each condition) were loaded for mRNA-Seq in Illumina GAIIx.
Nannochloropsis core genes were identified as the intersections of the five “IMET1 pairwise cores,” which were obtained by searching IMET1 proteins via BlastP and tBlastN against the proteome and the genome, respectively, of each of the other five strains with an e-value cutoff of 1e-5 and a protein sequence identity cutoff of 80%. Paralogous groups among these six strains were identified by a Markov Clustering algorithm (OrthoMCL [16], v. 4) with an inflation index of 1.5. PAML (v. 4.4c) codon substitution models and likelihood ratio tests (codeml) were used to estimate the selective pressure. An identical method was applied in the establishment of paralogous groups among other model microalgae.
HGT candidates were inferred following the method in the genomic analysis of Galdieria sulphuraria (Text S1; [18]). Phylogenetic trees for each of the putative HGT genes in NEWICK format were deposited in Dataset S3. The phylogenetic tree for each HGT candidate was manually checked and only accepted when a clear pattern of HGT was observed in both Neighbor Joining (NJ) and Maximum Likelihood (ML) trees. To deduce the evolutionary origins of lipid biosynthesis-related genes, we first implemented the strategy described in Chan et al. [22] to build a comprehensive database and to construct the homologous groups for each lipid synthesis gene, except that we collected more recently published genomes and EST datasets updated in public databases, including genomes of the red algae G. sulphuraria [18], Chondrus crispus [38] and Porphyridium purpureum [29]. In the following phylogenetic analysis, phylogenies for the homologous group of each lipid synthesis gene were constructed in MEGA5 by both NJ and ML methods. A gene was inferred to be potentially derived from a green or red algae related secondary endosymbiont when the phylogeny was supported by both NJ and ML trees.
For a comprehensive and detailed description of the methods, please refer to Text S1.
|
10.1371/journal.pcbi.1000642 | A Bayesian Partition Method for Detecting Pleiotropic and Epistatic eQTL Modules | Studies of the relationship between DNA variation and gene expression variation, often referred to as “expression quantitative trait loci (eQTL) mapping”, have been conducted in many species and resulted in many significant findings. Because of the large number of genes and genetic markers in such analyses, it is extremely challenging to discover how a small number of eQTLs interact with each other to affect mRNA expression levels for a set of co-regulated genes. We present a Bayesian method to facilitate the task, in which co-expressed genes mapped to a common set of markers are treated as a module characterized by latent indicator variables. A Markov chain Monte Carlo algorithm is designed to search simultaneously for the module genes and their linked markers. We show by simulations that this method is more powerful for detecting true eQTLs and their target genes than traditional QTL mapping methods. We applied the procedure to a data set consisting of gene expression and genotypes for 112 segregants of S. cerevisiae. Our method identified modules containing genes mapped to previously reported eQTL hot spots, and dissected these large eQTL hot spots into several modules corresponding to possibly different biological functions or primary and secondary responses to regulatory perturbations. In addition, we identified nine modules associated with pairs of eQTLs, of which two have been previously reported. We demonstrated that one of the novel modules containing many daughter-cell expressed genes is regulated by AMN1 and BPH1. In conclusion, the Bayesian partition method which simultaneously considers all traits and all markers is more powerful for detecting both pleiotropic and epistatic effects based on both simulated and empirical data.
| Genome-wide association studies (GWAS) have yielded several causal genes for many human diseases. However, the mechanisms underlying how DNA variations affect disease phenotypes have not been well understood in many cases. Gene expression is intermediate between DNA and clinical endpoints. Linking DNA variation and gene expression variation, often referred to as “expression quantitative trait loci (eQTL) mapping”, has yielded clues of mechanisms and pathways by which DNA variations impact phenotypes. Because of the large number of genes and genetic markers in such analyses, it is extremely challenging to discover how a small number of eQTLs interact with each other to affect mRNA expression levels for a set of co-regulated genes. We present a Bayesian method to identify genetic interactions and more eQTLs by treating co-expressed genes as a module. Our method provides a tool to study genetic interactions in human disease models.
| Studies in the genetics of gene expression combine gene expression and genotype data in segregating populations to detect loci linked to variations in RNA levels. These loci are referred to as expression quantitative trait loci (eQTL). To date, eQTL studies have been pursued in a number of species ranging from yeast to mouse and human [1]–[3]. A common theme of these studies is to treat thousands of gene expression values as quantitative traits and conduct QTL mapping for all of them.
Most eQTL studies are based on linear regression models [4] in which each trait variable is regressed against each marker variable. The p-value of the regression slope is reported as a measure of significance for the association. In the context of multiple traits and markers, procedures such as false discovery rate (FDR) controls [5] can be used to quantify family-wise error rates. Despite the success of this type of regression approach, a number of challenging problems remain. First, these methods can not easily assess the joint effect of multiple markers beyond additive effects. Storey et al. [5] developed a step-wise regression method to find eQTL pairs, then Zou and Zeng improved it [6]. This procedure, however, tends to miss eQTL pairs with small marginal effects but a strong interaction effect. There are methods for detecting eptistatic effects without main marginal effects [7]–[8]. However, their applications are limited to a few clinical traits instead of thousands of expression traits due to computational constraints. Second, there are often strong correlations among expression levels for certain groups of genes, partially reflecting co-regulation of genes in biological pathways that may respond to common genetic loci and environmental perturbations [2], [9]–[11]. Previous findings of eQTL “hot spots”, i.e., loci affecting a larger number of expression traits than expected by chance, and their biological implications further enhance this notion and highlight the biological importance of finding such gene “modules”. Mapping genetic loci for multiple traits simultaneously is more powerful than mapping single traits at a time [12]. Although for a known small set of correlated traits, one can conduct QTL mapping for the principal components [13], this method becomes ineffective when the set size is moderately large or one has to enumerate all possible subsets. An alternative approach is to identify subsets of genes by a clustering method, and then fit mixture models to clusters of genes [14]. The eQTL mapping then depends on whether the distance metric used by the clustering method is appropriate, whether the method can find the right number of clusters.
We address these issues by modeling the joint distribution of all genes and all markers simultaneously. Under a Bayesian framework, we introduce three sets of latent indicator variables for genes, markers, and individuals, and then systematically infer the association between groups of genes and sets of markers. In this framework, correlated expression traits and their associated set of markers are treated as a module so as to account for epistatic interactions and pleiotropic effects. Parameters of interest are the partitions of genes and markers into modules, and the partition of individuals into different types that correspond to the relationships between expression levels and marker genotypes in a given module. A Markov chain Monte Carlo (MCMC) algorithm is designed to traverse the space of all possible partitions. Simulation studies show that the proposed method achieves significantly improved power in detecting eQTLs compared to traditional regression-based methods. A particular strength of our method is its ability to detect epistasis with high power when the marginal effects are weak, addressing a key weakness of all other eQTL mapping methods.
We applied our method to a previously described data set consisting of gene expression and genotypes data for 112 segregants from a cross between laboratory (BY) and wild (RM) strains of S. cerevisiae [15]. In addition to identifying several modules linked to single eQTLs that are consistent with previous reports [1],[11],[16], our method dissected large eQTL hot spots into different modules that correspond to different causal regulators or to primary and secondary responses to causal regulators. In addition, we detected nine modules under the control of two genetic loci. One of these modules corresponds to a previously verified result regarding the interaction between GPA1 and MAT [5],[16]. another is regulated by both ZAP1 expression and genotype, consistent with previously described results [17]. The other seven modules represent novel findings. Three of these appear to be artifacts of cross-hybridization in microarray experiments; while another exhibits strong epistatic interactions between two loci consisting of many daughter-cell expressed genes that we predict are under the regulation of AMN1 and BPH1.
We define a module as a set of gene expression traits (referred to as “genes” henceforth) and a set of genetic markers (e.g., SNPs) such that the variation of the gene expression traits is associated with the variation of the markers, as shown in Figure 1. This association between multiple genes and markers is characterized by a latent indicator variable, individual type, conditional on which the trait and marker variables are independent of each other. The individual type latent variable can be viewed as representing a certain combination of markers that induces changes in expressions of a certain set of genes across different individual types. In the simplest case with a single marker, the individual type could correspond to a dominant genetic model, as illustrated in Figure 2A. In this instance, our model is mathematically equivalent to the regression model (Figure 2B). In the case of two markers associated with gene expression traits, there could be two to nine individual types (various genotype combinations). Figure 2C illustrates a case with three individual types: 1) high expression values associated with red-colored genotype combinations, 2) medium expression values with blue-colored combinations, and 3) low expression values with green-colored combinations. The goal of the Bayesian partition method is to simultaneously partition genes and SNPs into modules. The details of the Bayesian partition model are described in the Methods section.
To test the effectiveness of our method, we simulated 120 individuals with 500 binary markers and 1000 expression traits in the context of inbred cross of haploid strains. There are eight modules (summarized in Table 1), each consisting of 40 genes, simulated from different epistasis models based on the linear regression framework, which is different from the posited Bayesian model in our analysis. The genotypic means and frequencies for the two loci used in the simulation are listed in Table 2. We repeated the simulation 100 times and analyzed the simulated data using two methods: (1) our Bayesian partition method using parallel tempering [18] with 15 temperature ladders, referred to as BP; (2) the two-stage regression method of Storey et al [5], referred to as SR. Details of the simulation and implementation of these two methods are described in the Supplemental Material. As shown from the receiver operating characteristic (ROC) curves in Figure 3, BP achieved a significantly higher power to detect eQTLs compared to SR. For example, allowing for 50 false positives, BP detected more than 500 (out of 640) true gene-marker pairs, whereas SR only detected ∼100 true pairs and became plateaued even with many more false positives allowed. There are likely two reasons for this. First, we modeled the co-regulated genes as a module so that information from all genes in a given module could be aggregated to improve the signal. Multiple trait mapping has proven to be more powerful than single trait mapping [12] in the regression framework. Second, we modeled epistatic interactions explicitly so that markers with weak marginal but strong interactive effects could be detected.
The contrast of the performances of these two methods is most prominent when the marginal effect is weak. For example, in modules B, D and H, the rate of true positive detections of SR never exceeded 5% even at the generous FDR threshold of 90%. In modules E, F, and G where the major marker explains more than 70% of the genetic variation, SR detected the major marker in nearly 50% of the simulations at the 50% FDR threshold, but not the minor marker. In contrast, BP performed superiorly and robustly in all eight modules. The module by module comparisons are detailed in the Supplemental Material Text S1 and shown in Supplementary Figure S1.
Figure 4 provides a graphical view of the BP result for another simulated dataset with 120 individuals, 1000 genes, and 500 markers. Four distinct modules, with 60, 60, 40, and 40 genes, and controlled by 3, 2, 1, and 2 markers, respectively (shown in Supplementary Table S1), are simulated similarly as in the previous example (more details in the Supplemental Material Text S1). The shape and height of a point represent the most probable module classification and the corresponding maximum posterior probability of a gene. We see that all of the “background” genes were correctly classified according to their highest posterior probabilities. Most genes in the four non-null modules were also correctly classified, other than a very few ones that were classified into the null module, most likely due to their weak signals. BP also correctly identified the truly associated markers of the four modules with high posterior probabilities (shown in Supplementary Table S2).
We applied our Bayesian method to a data set consisting of gene expression and genotypes for 112 segregants from a cross between laboratory (BY) and wild (RM) strains of S. cerevisiae [15] and detected 29 modules of genes and their associated markers (Methods). Among these 29 modules, 20 are linked to a single eQTL while the remaining nine are linked to two eQTLs. Three of the nine linking to two eQTLs give rise to significant epistatic interactions between the two loci. Twenty-six of the 29 modules significantly overlap (corrected p-value<0.05) with at least one of the 13 gene groups previously reported as mapping to eQTL hot spots [11]. We also tested each of these modules for enrichment using GO terms, a yeast knockout compendium [19], and transcription factor binding sites [20]. At p-value<0.05 after multiple testing correction, 21 modules have at least one GO term enrichment; 22 modules overlap with at least one knockout signature, and 13 modules are enriched for at least one transcription factor binding site. The result is summarized in Table 3 and a breakdown result is in Supplementary Table S3. In contrast, the LOD score distributions of transcripts at the associated markers under the “single-transcript-single-marker” model are shown in Supplementary Figure S2. Our Bayesian method identifies significantly more weak gene-marker associations than the simple model. These GO enrichments support the biological relevance of different modules detected by our method. Each module is described in detail in the Supplemental Material Text S1.
We have developed a Bayesian partition model for simultaneously mapping multiple eQTLs for multiple sets of co-regulated genes. Whereas conventional linkage analysis has been widely and successfully applied to the study of one or a small number of traits at a time, our module-based method is suitable for analyzing thousands of phenotypes simultaneously. Both simulation studies and empirical data examples demonstrated that our method is effective for detecting marker interactions, even when no marginal effects could be detected. These improvements in power are a direct result of accounting for the correlation among gene expression traits and assessing the joint effect of multiple eQTLs, including interactions, on these correlated gene sets.
One of the main advances in our approach is the introduction of the “individual type” as a latent variable to describe associations between gene expression traits and markers. The individual type latent variable can be interpreted as a classification of individuals according to a combination of phenotypes and genotypes. The underlying mathematical model for this dependence structure is represented as a chain in which the joint distribution for some set of markers influences a set of expression traits via a latent “individual type” variable. After integrating out this latent variable, we observe a direct relationship between the marker and gene expression sets, similar to what would have been obtained from a the traditional regression model in the single-marker, single-gene case (Figures 2A and 2B). However, the advantage over the standard regression in introducing the latent individual type variable is its enabling us to model epistatic interactions and pleiotropy simultaneously.
Linkage disequilibrium (LD) among adjacent markers is an important feature of the genetic marker data. For individuals produced by the laboratory crosses (e.g., F1 and F2 designs), the marker dependency can be modeled satisfactorily by a Markov chain. The BP model can easily entertain this modification of the background marker distribution, but the computation time required to run this modified model dramatically increases since we need a forward-summation-backward-sampling algorithm to update the marker indicators (see Supplemental Material Text S1 for details). Another ad hoc strategy to account for the marker correlations without directly modeling them was to first scan all markers and to enumerate those marker pairs with correlations exceeding a given threshold. Then, in the MCMC algorithm, we imposed a mutually exclusive condition for such pairs so that highly correlated marker pairs would not appear simultaneously in any module.
We compared the Markov model approach with the ad hoc strategy on a small simulated data sets and a subset of the real data (data not shown). The ad hoc strategy always provided nearly identical results to that of the Markov model with only a fraction of the computation cost. Note that there are also markers that are highly correlated but are not physically linked [26]. In such cases the Markov model actually worked less satisfactorily than the ad hoc approach.
Our method shares some similarities to other methods in the literature, but also shows clear distinctions. For example, Lee et al. [17] proposed to simultaneously partition the gene expression and genotype markers. However, their method requires strong priors on the potential regulators, while our method does not. Kendzioski et al. [14] proposed a mixture of markers model to find the eQTLs for multiple gene expression. However, their method separates the gene clustering and eQTL mapping steps, where they first use k-means clustering to identify subsets of genes, and then apply eQTL mapping to the clusters of genes. In addition, their method does not address the epistatic effects. In contrast, gene expression partition and eQTL mapping are modeled jointly in our Bayesian method, and we are able to effectively detect epistasis by using a comprehensive statistical model on both the gene expression and the markers. Our analysis of the yeast data identified 20 modules linked to one eQTL and 9 modules linked to two eQTLs, among which three giving rise to strong epistatic interactions between markers. Some of the modules linked to two eQTLs are consistent with previously reported results [5],[17], and we were able to identify more true positive hits along with fewer false positives than previously reported.
It is of note that our approach can also be applied to mammalian data and to other quantitative traits data with discrete genetic and environmental covariates. In typical mouse studies, about 2000 SNPs are genotyped and 25,000 transcripts are measured, among which about 8000 are significantly differentially expressed [2]. The computation time will be at a similar order of the yeast data analysis. In typical human studies, 650,000 SNPs are genotyped and 40,000 transcripts are measured. The computation time will dramatically increase. We may, however, restrict our attention to hundreds of SNPs identified as possibly associated with gene expression traits in a human cohort, or/and to fewer expression traits identified as being relevant to diseases of interest [27]–[28]. In this type of scenarios, the input datasets would be roughly equivalent to the yeast data set described herein. Many other such applications can be imagined,
We are also improving parallelization implementation. Hopefully, we will be able to appropriately generalize and improve the Bayesian model as well as the MCMC algorithm so that our method can be applied to complete mammalian and other large data sets.
A module is defined in the Results section as a set of gene expression traits (referred to as “genes” henceforth) and a set of genetic markers (e.g., SNPs) such that the mRNA expression variation of the genes is associated with the allelic variation of the markers. This association between multiple genes and markers is characterized by a latent indicator variable, individual type, conditional on which the trait and marker variables are independent of each other. The individual type latent variable can be viewed as representing a certain combination of markers that induces changes in expressions of a certain set of genes across different individual types.
To formally describe our model, consider a sample with N individuals. Each individual i is measured with G gene expression values denoted as and M marker genotypes denoted as . We assume that the observed data can be partitioned into D nontrivial modules plus a null component. The number of non-null modules, D, is pre-specified by the user and should reflect the user's prior belief in the higher level structure of the data. Every gene g or marker m belongs to one of the D nontrivial modules or the null module, determined by the gene indicator and the marker indicator . For each module , we further partition the N individuals into types denoted by the individual indicators for . Each module may have a different number of individual types as well as different ways of partitioning the N individuals. For example, with a single biallelic marker (alleles ‘A’ and ‘a’) in the module, the module may have two individual types corresponding to genotypes aa vs. Aa or AA (dominant model), or 3 individual types corresponding to genotypes aa, Aa and AA (additive model). We seek module partitions in which expression patterns are similar for all genes, and gene expression variations across different individuals can be explained by the individual types. A cartoon illustration of the partition model is shown in Figure 1.
We model the gene expression traits in module d by an ANOVA model so that each trait value is the sum of the gene effect (), the eQTL effect for individual type k (), the individual effect (), and an error term:where gene g is in module d, k is the individual type of i, and ri and αg are random effects, following independent Gaussian distributions with mean zero.
To account for epistasis, we model the joint distribution of all the associated markers of module d, , by a multinomial distribution, whose frequency vector is determined by the individual type k, i.e.,For example, if there are two markers in the module and each has three genotypes, then there are nine combinations of the marker patterns. Thus follows a 9-dimensional multinomial distribution.
For the null component, we assume that there is no association between the genes and the markers. The gene expression traits follow a normal distribution and the marker genotypes follow an independent multinomial distribution.
To avoid overfitting, we put an exponential prior on the indicator variables to penalize partitions with high complexity:where are the number of genes, markers and individual types in module d, and is the number of genotypes at each marker. We use conjugate priors on the continuous parameters, such as means and variances of the Gaussian distributions and frequency vectors of the multinomials, so that most of these parameters can be integrated out analytically to reduce the complexity of the posterior distribution.
The joint posterior distribution of all unknown variables is of the form:where β represents the set of left-over continuous parameters unable to be integrated out analytically. In order to make inference on the eQTL modules from this posterior distribution, we construct a Markov chain Monte Carlo method to traverse the joint space of all unknown parameters. Each Markov chain is randomly initialized, and uses the Gibbs sampler and the Metropolis-Hasting algorithm [18] to update the variables. We implement a split-merge algorithm, which is a special case of the reversible jump MCMC [29], to update the individual partitions globally. Parallel tempering [30] is used to help mixing the Markov chain. Further details of the modeling and sampling strategies can be found in the Supplemental Material Text S1.
Posterior probabilities are evaluated for each gene and candidate marker set to belong to each module based on the Monte Carlo samples. A threshold is then applied to the posterior probabilities to determine whether a particular gene and marker set should be included in a module.
We assembled genotypic and expression data from 112 segregants obtained from a previously described yeast cross between the BY and RM strains of S. cerevisiae [15]. Of the 5,740 genes represented on the microarrays in this study, we selected 3,662 informative genes as input into the partition algorithm following the same criteria as previously described [10]. We then transformed the gene expression values by first performing quantile normalization [31] to make the distribution of the log-expression ratios for each individual to be the same, and then normalizing each gene so that the mean expression level for each gene was 0 and the standard deviation was 1.
Given that genes in the data set have been previously mapped to 13 distinct eQTL hot spots [11] and that there can be multiple causal factors for a single eQTL hot spot, we set the number of starting modules for our MCMC algorithm to 35∼45 (3×13 plus a null model) to account for these previously identified groups, and to also allow for the detection of new groups as well. For the parallel tempering implementation, we used 30 temperature ladders with almost equal spacing so that the average acceptance probability for exchanges between adjacent chains was roughly 0.15–0.3. We ran MCMC sampling for 1,000,000 iterations in each chain, which took one week of 30 CPUs (accounting for 30 parallel temperature ladders of the MCMC algorithm) on a Linux cluster with 2GHz CPUs. The log posterior probability and its auto-correlation curve depicted in Figures S5C and S5D highlight that the Markov chain became stationary after a burn-in period. See Supplemental Material Text S1 for more details.
Because markers in the yeast data set are very densely distributed, adjacent markers are almost always highly correlated. After MCMC sampling, markers adjacent to the “truly” linked marker often diluted the posterior probability for the true marker-module linkage. Since a proper Markov chain model for unlinked markers is computationally too expensive to implement (see Supplemental Material Text S1), we employed a heuristic method to counter this problem. We first specified a window centered at each marker so that markers inside the window are in high LD with the marker at the center. The posterior probabilities of all markers in the window were summed up and regarded as the modified posterior probability of the central marker. The markers with peak probabilities exceeding the given threshold were selected and all other markers in the corresponding windows were masked out.
Although we did not explicitly model pleiotropic effects for markers (i.e., single markers were not allowed to be associated with expression traits in multiple modules), we reported several modules mapped to the same markers in the yeast data set (See Table 3 and discussions in the Supplemental Material Text S1). The reason for this apparent contradiction is due to the aforementioned moving window approach and the dense distribution of the markers. In other words, if marker m is truly linked to two modules, in computation its adjacent markers can serve as its surrogates so that a subset of these markers are mapped to module 1, and the remainders mapped to module 2. Then the use of the moving window method can restore the total probability back to marker m.
To test the robustness of our result with respect to the initial parameters, we ran our program using three different numbers of modules, , and , each having three independent runs. Samples from the run with the highest average posterior probability for each value of were used in the subsequent analyses. We chose 0.8 as the threshold for the posterior probabilities to determine the module membership for each gene and marker. We observed that more than 70% of the genes were consistently grouped together and mapped to the same markers (or null module) in all the runs with different D values. These genes and their associated markers formed the list of 29 modules.
|
10.1371/journal.pgen.1001292 | A Study of CNVs As Trait-Associated Polymorphisms and As Expression Quantitative Trait Loci | We conducted a comprehensive study of copy number variants (CNVs) well-tagged by SNPs (r2≥0.8) by analyzing their effect on gene expression and their association with disease susceptibility and other complex human traits. We tested whether these CNVs were more likely to be functional than frequency-matched SNPs as trait-associated loci or as expression quantitative trait loci (eQTLs) influencing phenotype by altering gene regulation. Our study found that CNV–tagging SNPs are significantly enriched for cis eQTLs; furthermore, we observed that trait associations from the NHGRI catalog show an overrepresentation of SNPs tagging CNVs relative to frequency-matched SNPs. We found that these SNPs tagging CNVs are more likely to affect multiple expression traits than frequency-matched variants. Given these findings on the functional relevance of CNVs, we created an online resource of expression-associated CNVs (eCNVs) using the most comprehensive population-based map of CNVs to inform future studies of complex traits. Although previous studies of common CNVs that can be typed on existing platforms and/or interrogated by SNPs in genome-wide association studies concluded that such CNVs appear unlikely to have a major role in the genetic basis of several complex diseases examined, our findings indicate that it would be premature to dismiss the possibility that even common CNVs may contribute to complex phenotypes and at least some common diseases.
| Despite the large number of SNPs found to be reproducibly associated with complex diseases, they collectively account for only a small proportion of the overall heritability to such traits. CNVs have thus been proposed to explain some of the missing heritability and to alter disease susceptibility. However, a recent study of the genetics of 8 common diseases involving 16,000 cases and 3,000 controls failed to identify any novel CNVs associated with disease and concluded that CNVs are unlikely to play a major role in their etiology. Studies we report here show that we must be careful not to dismiss the possibility that CNVs may indeed underlie some of the observed associations with complex disease. Our findings show that well-tagged CNVs are disproportionately more likely to be eQTLs, as well as cis-eQTLs, than frequency-matched SNPs; furthermore, reproducible trait associations, as represented in the NHGRI catalog, are enriched for well-tagged CNVs than frequency-matched SNPs. Because of these findings on the strong functional relevance of these CNVs, we created a database (available at http://www.scandb.org/) of expression associated CNVs to supplement our earlier studies of SNP eQTLs and to contribute to future studies of the genetics of complex traits.
| Although genome wide association studies (GWAS) have considerably advanced our knowledge of the genetics of complex traits, they fail to account for the bulk of the overall heritability [1]. Structural variation, including copy number variants (CNVs), may account for some of the missing heritability, but a comprehensive study of the contribution of these variants to complex traits genetics is generally lacking. One of the main challenges of integrating CNVs into GWAS is the reliability of their assay, but improvements in genotyping platforms and the development of algorithms for the analysis of CNVs increasingly facilitate the investigation of the role of CNVs in disease susceptibility [2]. Given the effect of adjacent CNVs on SNP intensity data, early SNP arrays were designed with the consequence that assays from many CNV loci were excluded, but a newer generation of arrays have been implemented that combine SNP assays and copy number probes optimized for copy-number measurement for target regions of known CNVs [3]. New algorithms are also facilitating higher resolution maps of common CNVs (those that segregate at an allele frequency >5%).
Partly due to these technological and analytic developments, the search for links between CNVs and disease susceptibility has been actively pursued in recent years. Structural variation has of course long been known to influence such Mendelian disorders as Williams-Beuren syndrome or Charcot-Marie Tooth neuropathy Type 1A, but more recently there has been widespread interest in the study of CNVs as mediators of more common complex diseases. The contribution of structural variation to certain neurodevelopmental disorders such as schizophrenia and autism spectrum disorder has been investigated by recent studies of de novo copy number variation [4]. A common 20 kb deletion upstream of IRGM in perfect linkage disequilibrium (r2 = 1.0) with the SNP most strongly associated with Crohn's disease in the region has recently been shown to alter IRGM regulation, which in turn is known to affect autophagy, consistent with the deletion polymorphism being the causal allele [5]. Similarly, a 45 kb deletion polymorphism in NEGR1 has been shown to be the likely causal variant for the reproducible association with body mass index in humans, demonstrating a neuronal influence on body weight regulation [6]. These discoveries and similar developments suggest that the interpretation of known disease associations continues to be fraught with challenges, as CNVs may be in linkage disequilibrium (LD) with associated variants, and that the identification of causal variants requires an approach that considers both SNPs and CNVs at associated loci.
The recent release [7] of a reference map of human CNVs in the HapMap populations promises to shed light on the role of these variants in disease pathogenesis. On the other hand, a recent comprehensive study [8] involving 16,000 cases of eight common diseases and 3000 shared controls strongly suggests that CNVs are likely to make a relatively minor contribution to disease susceptibility. In the present study, we approached the investigation of this important topic by exploring evidence of enrichment of CNVs among trait-associated loci in disease susceptibility and in disease classes, as represented by the NHGRI catalog of reproducible associations from genome-wide association scans [9], for certain classes of CNVs, and by conducting a comprehensive functional characterization of CNVs as expression quantitative trait loci (eQTLs). The former seeks to evaluate the genetic information contained in CNVs in regard to disease risk and other complex traits; the latter assesses the functional impact of CNVs on the transcriptome, through which they may contribute to or cause disease phenotypes and other complex traits [10].
Table 1 summarizes the CNVs included in our study. Because of the ease with which SNPs can be utilized in enrichment studies appropriately conditioned on minor allele frequency, we utilized tCNVs (CNVs tagged by SNPs) for many of these studies, but did not otherwise consider them to be a special class of CNVs.
At an expression p value threshold of 10−4, tCNVs from the recently released study (N = 3,432) of CNVs in 16,000 cases and 3,000 controls [8] were evaluated for their effect on gene expression. We first considered tCNVs that are tagged at r2≥0.80. The tag SNPs for these CNVs enable SNP-based analyses and simulation studies. We identified 1,714 such SNPs tagging CNVs (corresponding to 1,761 tCNVs), of which 532 are eQTLs at this threshold; collectively, these SNPs tagging CNVs were found to target 2,215 distinct transcripts. Restricting our focus to eQTLs that regulate transcript levels in cis (defined as within 4 MB of target transcript), we observed a highly significant cis eQTL enrichment (p<0.001) from a simulation procedure (N = 1000) used to empirically generate the null distribution using randomly generated sets of variants of the same set size and matching minor allele frequency distribution as the SNPs tagging CNVs (see Materials and Methods). These SNPs tagging CNVs are significantly more likely to be associated with altered expression of genes in their vicinity than frequency-matched SNPs (see Figure 1). Specifically, 24 such SNPs tagging CNVs are acting in cis, whereas the expected count is 10.7 (SD = 3.2). On the other hand, the observed count of trans-acting eQTLs is 508 while expected count is 488.9 (SD = 17.6).
We identified 621 SNPs tagging CNVs corresponding to 626 tCNVs that are perfectly tagged (r2 = 1.0). Of these 621, 185 predict the expression of at least one transcript, whereas, based on simulations, the expected count is 166.5 (SD = 9.9). This observation implies a statistically significant (p = 0.024) enrichment for eQTLs among this class of CNVs. Restricting our focus to eQTLs that regulate transcript levels locally, we observed 10 such cis-acting tCNVs. This observed count is, as before, a significant enrichment (p<0.001) for cis-acting regulators compared with expected count (see Materials and Methods) of 3.5 (SD = 1.9) (see Figure S1).
To determine whether our observations are dependent on the p value threshold used to define an eQTL, we explored evidence of eQTL enrichment among the tCNVs (r2≥0.80) at p<10−6. Simulations yielded an expected eQTL count of 19.7 (SD = 4.5) whereas observed count is 28, yielding a statistically significant enrichment (p = 0.036). Expected cis-acting eQTL count is 1.9 (SD = 1.4) while observed count is 8. These findings demonstrate that the observed enrichment holds robustly at lower expression thresholds. There are 2.5 times as many trans eQTLs as cis eQTLs among the tCNVs at this threshold.
Using Gene Ontology (GO) enrichment [11], we found the target genes of these expression associated CNVs (eCNVs) to show a highly significant overrepresentation for certain biological processes including immune response (p = 2.3e-4, adjusted p-value 0.034) and regulation of apoptosis (p = 8e-6, adjusted p-value 0.015). Relative to other genic CNVs, the observation on the enrichment of the target genes of eCNVs for response to stimuli continues to hold (p = 9.9e-3, adjusted p-value 0.05).
Notably, we identified 22 tCNVs (r2 = 1) that predict the transcript level of 10 or more genes. These CNVs are regulating the expression of distal transcripts. By performing simulations on frequency-matched SNPs as the tCNVs, we observed an enrichment (expected count = 14.7, SD = 3.8) for CNVs controlling the expression of 10 or more, often distal, transcripts (See Figure S1). We identified 2 CNVs CNVR4596.1 (chrom10:4698559–4700493) and CNVR1045.1 (chrom2:165722938–165725072) that predict the expression of at least 100 transcripts. CNVR4596.1 does not contain any genes; on the other hand, CNVR1045.1 overlaps with SCN3A. Additional CNVs CNVR3307.1 (chrom7:17780194–17781249), CNVR3500.1 (chrom7:97233946–97240537), and CNVR7062.1 (chrom17:27130696–27131638) predict the expression of at least 50 transcripts. The CNV region CNVR3307.1 contains several SNPs (e.g., rs2723520, rs1404418, and rs1404419) in strong LD, which individually predict the expression of more than 80 transcripts.
We identified a CNV (CNVR2845.27 at chrom6:32710664–32743652) in perfect LD with the replicated type 1 diabetes SNP rs9272346 predicting the expression of at least 20 transcripts, mostly in the HLA region; a CNV (CNVR2845.40 at chrom6:32735154–32737954) is tagged (r2 = 0.95) by the same type 1 diabetes SNP. A tCNV CNVR2845.46 is also in strong LD (r2 = 0.93) with a replicated multiple sclerosis SNP rs3135388 and predicts the expression of 20 transcripts (of which the largest expression association p value is ∼10−33).
Since the deletion/duplication of multiple regulatory SNPs may result in aberrant transcription or disease, we evaluated the presence of SNP eQTLs within CNVs. We identified 33 CNVs harboring 50 or more SNP eQTLs that predict the expression of a transcript (Table S1 lists these CNVs based on the number of SNP eQTLs within); these CNVs show an average SNP eQTL density of 6 per 10 kb.
We observed 1,306 CNVs (out of the 3,432 CNVs included in the WTCCC study) containing at least one SNP eQTL. Notably, CNVs that harbor a greater number of SNP eQTLs tend to be less well-tagged; this observation holds robustly after accounting for CNV minor allele frequency, which shows little correlation with the number of SNP eQTLs within the CNV (corr = 0.005). In particular, of the top 100 WTCCC CNVs ranked according to the number of SNP eQTLs within, only 31 are well-tagged (r2≥0.80) and 56 are tagged at r2<0.30.
Given the observed functional relevance of tCNVs as regulatory polymorphisms, we comprehensively and systematically characterized the recent genome-wide survey of CNVs in HapMap lymphoblastoid cell lines (LCLs) for their role in expression. We expanded our SCAN database [12], which to date serves results only on expression associated SNPs, to include eQTL mappings of HapMap CNVs to transcriptional expression (assayed by the Affymetrix exon array) in the full set of HapMap CEU (Caucasians from UT, USA) and YRI (Yoruba people from Ibadan, Nigeria) populations (See Materials and Methods). This online catalog of expression associated CNVs (eCNVs), publicly available at http://www.scandb.org, should serve as an important resource in helping to inform our understanding of complex traits.
Of the CNVs in HapMap [7] showing association with a transcript (at p value threshold of 10−4) in the CEU and YRI samples, 31.1% and 33.1% respectively are predicting the expression of 5 or more transcripts. This is a much higher proportion than is the case for the well-tagged CNVs in the WTCCC CNV study (12.8%).
Among trait-associated SNPs as represented in the NHGRI catalog, we found a significant overrepresentation (p = 0.01) for tCNVs (r2≥0.80). Figure 2 shows the distribution of the number of tag SNPs generated from 1000 random sets of SNPs matching the minor allele frequency of the trait-associated SNPs. Table 2 shows the observed overlap between the trait-associated SNPs and the tCNVs (r2≥0.80). Enrichment analysis of disease classes shows that the observed tCNV enrichment for reproducible trait associations holds for autoimmune disorders and in quantitative metabolic traits.
Three HLA region CNVs, namely CNVR2841.20, CNVR2845.14, and CNVR2845.46, were recently shown to be associated with Crohn's disease, rheumatoid arthritis, and type 1diabetes, respectively [8]. Our expression study confirms that the last two of these (tagged at r2>0.90) are eQTLs that regulate the expression of multiple genes in this region; in contrast, the first CNV shows no evidence for being an eQTL. A previously identified Crohn's disease associated locus CNVR2647.1 [5] on chromosome 5, 22 kb upstream of the IRGM gene, is a trans eQTL for SHISA4 (p = 3e-05). Two replicated SNPs associated with HDL cholesterol are tagging (r2≥0.99) nearby CNVs, CNVR3814.1 (chrom8:19898924–19899800) and CNVR5165.1 (chrom11:48557432–48560877). Our study shows that both CNVs are eQTLs, with CNVR5165.1 targeting, distally, the olfactory receptor gene OR6J1. The CNV CNVR3814.1 is also well-tagged (r2 = 0.913) by a SNP reproducibly associated with triglycerides and is an eQTL targeting ARF4, which has been shown, in mice, to be involved in metabolic disorders [13].
Some studies have observed a differential taggability for common CNVs; these loci, it is argued, are less likely to be in strong LD (r2>0.8) with flanking markers than are frequency-matched SNPs. Several explanations have been proposed to account for this observed differential taggability, such as lower SNP density in regions near CNVs and the higher mutation rates of certain CNVs (producing greater allelic diversity nearby) than those of SNPs [2], [14], [15]. This observed difference in the magnitude of LD between CNVs and SNPs relative to LD among SNPs impacts, through its effect on allelic association, our ability to assess the phenotypic influence of CNVs. On the other side of this controversy, the recent genome-wide study of an extensive catalog of CNVs in 16,000 cases of eight common diseases has argued that CNVs are generally well-tagged by SNPs. Indeed, among 2- and 3- class CNVs that passed QC and had MAF>10%, the study found that nearly 80% were tagged by SNPs at r2>0.80. Consequently, replication of association results for CNVs can be conducted in an independent sample set by the use of tag SNPs. In this study, we set out to conduct a study of CNVs by analyzing their effect on gene expression and their association with disease susceptibility and other traits. The CNVs that are well-tagged by SNPs, which we call tCNVs, facilitate SNP-based simulation studies to evaluate enrichment. We proceeded to test whether these CNVs were disproportionately more likely to be functional than frequency-matched SNPs, as trait-associated loci or, under the assumption that few trait-associated polymorphisms are likely to alter the composition of gene products, as eQTLs influencing phenotype by altering gene regulation. Our study found that CNV-tagging SNPs are enriched for cis eQTLs, and, furthermore, that reproducible trait associations show an overrepresentation of tCNVs relative to frequency-matched SNPs. While the tagged CNVs are particularly easy to investigate in enrichment studies, we found that the proportion of eQTLs (at p value threshold of 10−4) in the non-WTCCC CNVs (39%) is higher than in the well-tagged WTCCC CNVs (30%). Given these strong findings on the functional relevance of CNVs, we created a comprehensive online resource of expression associated CNVs in the HapMap populations to supplement our earlier studies on SNP eQTLs.
CNVs can affect phenotype in several ways [16]. Genes fully covered by CNVs may contribute to disease through a duplication or deletion event. Copy number variant breakpoints may disrupt the expression of genes that overlap CNVs. On the other hand, we have identified two CNVs at considerable distance (in trans) from their targets controlling transcript abundance as potential master regulators. Another CNV contains multiple regulatory elements which are each predicting the expression of at least 80 transcripts; the deletion of such important regulatory elements is likely to profoundly alter gene transcription. Importantly, we observed that 1,306 CNVs (out of the 3,432 CNVs included in the WTCCC study) harbor at least one SNP eQTL (defined at p value threshold of 10−4). Given our earlier observation that tCNVs are enriched for expression-associated CNVs (eCNVs), it is interesting to ask whether those CNVs not well-tagged by SNPs have interesting properties with respect to gene regulation. Of the CNVs that are tagged at r2<0.30, 44% harbor SNP eQTLs.
Previous studies [7], [15], [16] have reported that genes that undergo dosage differences due to the presence (proximity) of CNVs show an enrichment for genes involved in immune response and response to external biotic stimuli. We identified an olfactory receptor gene OR6J1 and a gene DAD1 (defender against cell death 1), both on chromosome 14, that are trans-regulated genes for the tCNV CNVR5165.1 on chromosome 11. We found an overrepresentation for target genes involved in the calibrated molecular response to stimulus (whether chemical stimulus or potential internal or invasive threat). Our novel observation in this regard is that previously reported enrichment for genes relevant for molecular-environmental interactions generalizes to the target genes for tCNVs as eCNVs.
The recent WTCCC CNV study concluded that common CNVs that can be typed on existing platforms are unlikely to have a major role in the genetic basis of complex diseases [8]. The same study reported that there was no enrichment of association signals among CNVs involving exonic deletions. Our findings recommend caution in assessments of the contribution of CNVs to the genetics of complex traits. Even under the assumption that most common CNVs are well tagged by SNPs and therefore interrogated by existing SNP GWAS, reproducible trait associations are enriched for these CNVs compared to random expectation. The prominence of such CNVs among reproducible trait associations with autoimmune disorders and metabolic traits suggests that these variants may indeed contribute to certain disease classes; alternatively, they may act in conjunction with other variants such as SNPs to confer susceptibility. Importantly, these CNVs are disproportionately more likely to predict transcript levels than frequency-matched SNPs, and they are more likely to affect many different gene expression traits as master regulatory polymorphisms. An important issue to address is whether the enrichment of tCNVs as eQTLs and as disease-associated SNPs are correlated. Note that the probability that a random SNP is found in the NHGRI catalog increases from 0.062% in the set of HapMap SNPs to about 0.5% (more than 5-fold) in the set of well-tagged CNVs that are not eQTLs and to 2% in the tCNVs that are eQTLs.
It should be noted that the WTCCC CNVs included in our study reflect certain limitations. Indeed, as the WTCCC study itself explicitly noted [8], a large proportion (nearly 60%) of the candidate list of putative CNVs could not be reliably assigned copy number classes from the combination of experimental assay and analytical approaches; it is estimated that only about half of these are not polymorphic [8]. Particularly, nearly 6,500 such putative polymorphisms were excluded from subsequent analyses, as they were called with a single copy number class. It is of course possible that our conclusions may not generalize to these CNVs. Furthermore, eQTL mapping in microarray-based studies in LCLs is likely to yield only a subset of the eQTLs that will be identified using more refined methods in a variety of human tissues. The present study also has little conclusive to say about low frequency variants. Despite these current limitations, the annotation system we implemented in SCAN should prove useful to other investigators and seeks to be as comprehensive as possible by providing functional information for the most extensive map of CNVs to date from the recent population-based genome-wide survey [7]. Since the MAF spectrum of the NHGRI catalog of trait-associated SNPs from published GWAS is quite different from that of the SNPs on the genotyping platforms used to conduct these GWAS, we performed our enrichment analyses while conditioning on the MAF distribution. There are other potential representational biases in the NHGRI catalog of reported variants that may have affected our studies. The enrichment in disease classes relative to other classes of traits, for example, may be the result of increased power (e.g., due to greater sample sizes) for these categories.
Since gene expression is intermediate to other complex phenotypes, a global view of the influence of CNVs on the transcriptome may lead to a better understanding of their role in disease susceptibility. While we identified, as perhaps expected, CNVs located in the HLA region associated with multiple expression phenotypes and with various autoimmune disorders, we also observed, strikingly, several non-HLA CNVs that regulate multiple transcripts, including 2 (one on chromosome 10 and the other on chromosome 2) associated with more than 100 expression traits. Using the most extensive population-based CNV map available [7], we observed a greater proportion of master regulatory CNVs than observed among the well-tagged CNVs. Collectively, all these findings reinforce the importance of considering all types of variation to elucidate the genetic architecture of complex traits.
The mRNA expression was assayed in HapMap lymphoblastoid cell lines (LCLs) in 87 CEU and 89 YRI using the Affymetrix GeneChip Human Exon 1.0 ST Array [18]. This dataset is available in the public repository Gene Expression Omnibus (Accession No: GSE9703). Measurements on the exon array are done both at the exon-level and at the gene-level. The exon array includes 1.4 million probesets and profiles nearly 13,000 transcript clusters. The transcript clusters include a set of probesets containing all known exons and 5′ and 3′ UTR regions in the genome so that probes interrogate entire gene regions and not just 3′ UTRs (in contrast to other arrays).
The results of our eQTL studies, previously on single base polymorphisms [10], [17], [18], in HapMap LCLs have been made publicly available in an online database SCAN [12]. To evaluate the influence of CNVs on the transcriptome, we performed linear regression on over 13,000 transcript clusters with reliable expression – namely, the log2 transformed normalized expression intensity is greater than 6 for at least 80% of the samples – and the CNVs assayed in HapMap LCLs. We downloaded genotype information for these CNVs from the recently released reference catalog of HapMap CNVs [7]. All eQTL analyses were done in both CEU and YRI samples. We extended the SCAN database to include a catalog of expression-associated CNVs in the HapMap populations.
The National Human Genome Research Institute (NHGRI) curates the results of published GWAS and makes them publicly available in an online catalog of SNP-trait associations. This catalog provides a useful resource for the investigation of genomic characteristics of trait-associated SNPs and of the role of common variants in common disease etiology [9]. The version we utilized for our present study was retrieved on June 29, 2009. SNP-trait associations included in the catalog are those with p values<1.0×10−5, and generally only one SNP within a gene or a high LD region is included unless there is evidence of an independent association. To evaluate whether our observation on the significant enrichment among tagged CNVs for trait-associated SNPs is driven by certain traits or disease classes, we considered separately various disease types, including neuropsychiatric disorders, cancers, and autoimmune disorders.
We conducted simulations to evaluate enrichment for eQTLs among the tCNVs included in our study. To empirically generate the null distribution of no enrichment, we randomly generated sets of SNPs of matching minor allele frequency as the original list of SNPs tagging the tCNVs, as previously described [10]. To enable us to perform simulations conditional on MAF, we constructed non-overlapping MAF bins, each of width 0.05, using the MAFs of the SNPs in the HapMap CEU samples. The null sets were drawn from the combined platform SNPs (Affymetrix 6.0 and Illumina 1M) as well as from the entire set of HapMap CEU SNPs. The observed count is then compared to the empirically generated distribution to get an empirical p value for the enrichment.
A survey of 3,432 WTCCC CNVs [8] shows that their minor allele frequency distribution differs markedly from the distribution of reproducible trait-associated SNPs as represented in the NGHRI catalog (See Figure S2). Of these, nearly 80% with minor allele frequency at least 10% are well tagged (r2≥0.8). Roughly 89% of the tagged CNVs (r2≥0.8) are less than 10 kb in length, and 26% are less than 1 kb long (See ). Table S2 shows the median length (in log10) of the WTCCC CNVs.
The 3,432 CNVs included in this study passed the WTCCC quality control filters out of the 4,326 that were called with multiple classes [8].
|
10.1371/journal.pgen.1004657 | The Proprotein Convertase KPC-1/Furin Controls Branching and Self-avoidance of Sensory Dendrites in Caenorhabditis elegans | Animals sample their environment through sensory neurons with often elaborately branched endings named dendritic arbors. In a genetic screen for genes involved in the development of the highly arborized somatosensory PVD neuron in C. elegans, we have identified mutations in kpc-1, which encodes the homolog of the proprotein convertase furin. We show that kpc-1/furin is necessary to promote the formation of higher order dendritic branches in PVD and to ensure self-avoidance of sister branches, but is likely not required during maintenance of dendritic arbors. A reporter for kpc-1/furin is expressed in neurons (including PVD) and kpc-1/furin can function cell-autonomously in PVD neurons to control patterning of dendritic arbors. Moreover, we show that kpc-1/furin also regulates the development of other neurons in all major neuronal classes in C. elegans, including aspects of branching and extension of neurites as well as cell positioning. Our data suggest that these developmental functions require proteolytic activity of KPC-1/furin. Recently, the skin-derived MNR-1/menorin and the neural cell adhesion molecule SAX-7/L1CAM have been shown to act as a tripartite complex with the leucine rich transmembrane receptor DMA-1 on PVD mechanosensory to orchestrate the patterning of dendritic branches. Genetic analyses show that kpc-1/furin functions in a pathway with MNR-1/menorin, SAX-7/L1CAM and DMA-1 to control dendritic branch formation and extension of PVD neurons. We propose that KPC-1/furin acts in concert with the ‘menorin’ pathway to control branching and growth of somatosensory dendrites in PVD.
| Sensory neurons receive input from other neurons or sample their environment through elaborate structures termed dendritic trees. The correct patterning of dendritic trees is crucial for the proper function of the nervous system, and ample evidence points to the involvement of dendritic defects in a wide range of neuropsychiatric diseases. However, we still do not understand fully how this process is regulated at the molecular level. We discovered an important role for the protein-processing enzyme KPC-1/furin in the development of touch-sensitive dendritic trees in the roundworm C. elegans. Animals lacking this enzyme show multiple defects in the size, shape and number of these dendritic branches as well as other neurons. We further show that the gene encoding KPC-1 is expressed widely in the nervous system and that it is required within the branching neuron to exert its function on dendritic growth. Finally, we reveal a genetic connection between KPC-1 function and genes of the menorin pathway, which was recently discovered to also play an essential role in dendrite development. Thus, our findings add new insight into the molecular understanding of dendrite formation.
| Multicellular organisms sense their environment through specialized nerve cells termed sensory neurons. The diversity in shape and structure that dendrites of sensory neurons (hereafter named ‘dendritic arbors’) assume reflects the variety of stimuli they receive [1]. Work over the past two decades has established that dendritic arbor development of sensory neurons relies on conserved molecular mechanisms [2], [3]. However, dendritic arbors do not only exist in the periphery, but also in the central nervous system where they function to receive synaptic input from other neurons. The molecules and mechanisms that orchestrate dendritic arbor development, e.g. cell adhesion molecules and molecular motor proteins, are often similar between the peripheral and central nervous system [see ref. 3 for these and more examples]. Importantly, failed dendrite development has been linked to neurological diseases of the central nervous system, ranging from Autism Spectrum Disorders (ASD) to Schizophrenia [4], [5].
The understanding of sensory dendrite morphogenesis has been greatly advanced through studies in the fly Drosophila melanogaster (reviewed in [3]). These studies revealed the importance of transcriptional cascades, cytoskeletal proteins, the secretory pathway, microtubular transport, and the basement membrane for development of the so-called dendritic arborization (da) neurons in flies [3], [6]–[8]. In the nematode Caenorhabditis elegans, the PVD mechanosensory neurons have been used to study development of dendritic arbors owing to their stereotypic branching patterns (Figure 1A,B) [9], [10]. A PVD cell on each side of the animal sends a primary dendrite both anteriorly and posteriorly that branches off perpendicular higher order branches in a stereotypical fashion (Figure 1A, B). The resulting structures resemble candelabra and have hence been named menorahs [11]. Like fly da neurons, transcriptional cascades and microtubule based motor proteins have been shown to play a role during PVD dendritic arbor formation suggesting that basic principles of dendritic arbor formation are evolutionarily conserved [12]–[14]. Recently, a tripartite complex consisting of the extracellular molecule MNR-1/menorin, the neural cell adhesion molecule SAX-7/L1CAM both of which act from the hypodermis (skin) and the leucine rich repeat DMA-1/LRR transmembrane receptor on PVD dendrites have been shown to be critical for PVD development [15], [16].
Here we report the identification of the furin homolog kpc-1 in C. elegans as a factor that acts in concert with the ‘menorin’ pathway to shape sensory dendrite development. Furin is a serine protease of the proprotein convertase family, that following autocatalytic activation cleaves proteins at characteristic dibasic motifs (reviewed in [17]–[19]). We show that in C. elegans a reporter for kpc-1/furin is expressed broadly in the nervous system and is necessary for multiple neuronal positioning, neurite extension and branching events, including PVD dendritic growth and self-avoidance. Moreover, kpc-1 can act cell autonomously to shape PVD dendrites in a manner that is dependent on the ‘menorin’ pathway.
In a screen for genes involved in PVD development [16] we also identified mutant alleles of the proprotein convertase kpc-1, which encodes the C. elegans furin homolog based on sequence similarity and domain organization [20], [21]. Mutant animals show no obvious morphological or defects in body size. However, they do display severely defective dendritic arbors in both PVD and the analogous anterior FLP neurons (Figure 1B–E). The identified recessive alleles include two nonsense (dz177, dz182) and a missense allele (dz185) (Figure 1F, Figure S1). We also obtained a deletion (gk8) and two additional missense alleles (gk333538, gk779937) (Figure 1F,G, Figure S1). The deletion and nonsense alleles truncate or delete the catalytic domain of KPC-1/furin, respectively, and their phenotypes seem indistinguishable in severity, suggesting that all three alleles represent strong if not complete loss of function alleles. The missense alleles dz185 and gk333538 are hypomorphic alleles because their phenotype was (1) less severe when compared to the presumptive gk8 null allele and (2) more severe when placed in trans to a deficiency (hdf17) that spans the genomic region of kpc-1/furin (Figure 1H). The missense mutations affect residues in a perfectly conserved alpha helix that positions the histidine of the catalytic triad in furin [22] (Figure 1G), suggesting that proteolytic activity of KPC-1/furin is necessary for PVD development.
To define when kpc-1/furin function is required during development we first conducted timed RNAi experiments as described [16]. We found that knock down of kpc-1/furin starting at early larval stages when PVD sensory dendrites begin to develop [13] resulted in robust defects in PVD in adult animals (Figure 1I). In contrast, knock down starting at the L3 stage resulted in weaker phenotypes in adults (Figure 1I). RNAi mediated knockdown initiated only 12–24 h later at the L4 larval stage failed to result in any PVD defects. Importantly, RNAi mediated knockdown of GFP in PVD neurons suggested that changes in RNAi efficacy cannot account for lack of defects upon induction of kpc-1 knock down at later larval stages (Figure 1I). Consistent with these observations we find defects in PVD dendrite arborization in kpc-1 mutant animals already during the L3 larval stage rather than later as a result of a maintenance defect (Figure S2). Collectively, these findings suggest that kpc-1/furin functions during the earlier stages of PVD development and may not play a major role in dendrite maintenance.
We next sought to determine whether kpc-1/furin functions in PVD neurons or in surrounding tissues such as the hypodermis (skin) or muscle. To this end we first conducted transgenic rescue experiments. We found rescue of the PVD defects in kpc-1 mutants when a kpc-1 cDNA was expressed under control of a PVD-specific heterologous promoter but not when expressed in muscle or the hypodermis (Figure 2A,B). Expression of this PVD-specific kpc-1 transgene in a wild-type background did not have any detectable effect on PVD architecture (Figure S3). To further investigate where kpc-1 may function we constructed transgenic animals that carried a kpc-1 reporter where a 5.8 Kb fragment of the endogenous upstream regulatory region of kpc-1 drives expression of green fluorescent protein. This transcriptional kpc-1prom5.8::GFP reporter was widely expressed from early developmental stages in the nervous system including in PVD (Figure 2C–E, Figure S4). We conclude that kpc-1/furin acts cell-autonomously to shape the dendritic arbors of PVD neurons.
The observed extensive neuronal expression of the kpc-1 reporter prompted us to examine other neuron types besides PVD for defects in migration, neurite extension and branching in kpc-1 mutant animals. We detected significant defects in multiple neuron types in kpc-1 mutants. For example, the sensory AQR neurons failed to branch appropriately in the nerve ring whereas VC motoneurons showed defects in the extension and formation of characteristic branches near the vulva (Figure 3). Yet, kpc-1 did not appear to be required for all branching processes because branching in AIY interneurons as a result of overexpression of the secreted kal-1 cell adhesion molecule [23] did not require kpc-1 and, AVL neurons showed ectopic branches but no defects in normal branch formation (Figure 3). Whereas DA/DB motoneurons did not display obvious defects, D-type motoneurons showed gaps in the dorsal cord, possibly due to defective neurite extension, and increased numbers of inappropriately positioned commissures on the left side of the animals (Figure 3). Similarly, AIY interneurons displayed a characteristic axonal ‘short stop’ phenotype, likely as a result of defects in extension [23]. In addition, we observed neuronal cell positioning defects of ALM touch receptor neurons and HSN motoneurons (Figure 3). Thus, kpc-1's function in nervous system development extends well beyond its role in dendrite morphogenesis and affects many but not all neurons in the major neuronal classes (sensory, motor and interneurons). Importantly, kpc-1 appears to be involved in both control of neurite branch formation and extension (see also below).
To define the function of kpc-1 in dendrite development in more detail we subjected mutants to a morphometric analysis of PVD dendrites. These studies showed an increase in secondary and ectopic tertiary branches with a concomitant decrease in the number of tertiary and quaternary branches in kpc-1 mutants (Figure 4A–F). The secondary branches in kpc-1 mutants often failed to reach the vicinity of the sublateral nerve cords (Figure 4F) where they normally bifurcate to form tertiary branches [13]. Instead, they frequently sprouted ectopic tertiary branches in places that normally do not support tertiary branches (Figure 4F). Interestingly, the average length of branches was significantly reduced in kpc-1/furin mutants (Figure 5A–E), as was the aggregate length of each branch order (2°, 3° and 4° branches) separately or combined (Figure S5). To distinguish whether this phenotype is a reflection of a decrease or increase in growth or retraction, we conducted time-lapse analyses of kpc-1 mutants and measured growth and retraction of secondary branches. We found significantly slowed growth in kpc-1 mutants compared to wild type animals whereas the speed of retraction remained unchanged (Figure 6, Movies S1 and S2). None of the observed PVD defects were shared in mutants of the three other proprotein convertases encoded in the C. elegans genome, including kpc-2/egl-3, kpc-3/aex-5, and kpc-4/bli-4 [20] (Figure 1J, Figure S6), suggesting that these genes play individually no role during PVD development.
In the hypomorphic alleles kpc-1(gk333538) and kpc-1(dz185) we found that tertiary and possibly other dendritic branches appeared to show substantial overlap and an apparent lack of self-avoidance (Figure 7A,B). These observations could indicate bona fide defects in self-avoidance of dendrites. Alternatively, mutant dendrites could be growing erroneously in three dimensions into hypodermal tissues, rather than growing within a confined two-dimensional plane and directly avoiding each other. Such observations have recently been made for drosophila dendrite arborization neurons whose growth failed to be restricted to a two-dimensional plane in integrin mutants but rather occurred in three dimensions [7], [8]. To distinguish between these two possibilities, we optically sectioned animals bearing a hypomorphic mutation in kpc-1(dz185) and coded each optical section with a different color. Thus, dendrite processes that share a focal plane display the same color whereas processes in different focal planes will display different colors. These analyses suggested that dendrites touch and overlap in the same focal planes in kpc-1 hypomorphic mutant animals (Fig. 7C). Collectively, our data suggest that the normal function of kpc-1 is to promote the formation of tertiary and quaternary branches and to maintain self-avoidance between dendrites. In addition, kpc-1 functions in dendrite extension along the anterior-posterior and dorsal-ventral axes, likely by regulating the speed of growth.
The dendrite phenotypes in kpc-1 mutants were reminiscent of defects in mutants of the ‘menorin’ pathway, namely the cell surface molecule MNR-1/menorin, the neural cell adhesion molecule SAX-7/L1CAM and the leucine rich repeat containing transmembrane receptor DMA-1/LRR [15], [16]. We thus tested the genetic interactions between kpc-1/furin and each of these genes using the same morphometric criteria. Focusing first on the defects in branch formation in kpc-1/furin mutants (i.e. decreased tertiary and quaternary branching and increased secondary and ectopic tertiary branching) we found the kpc-1 mutant phenotype not further enhanced in double mutants with either mnr-1 or sax-7/L1CAM (Figure 4), suggesting that these genes act genetically in the same pathway. In contrast, dma-1 appeared generally epistatic to kpc-1 because the kpc-1; dma-1 double mutant looked more similar to the dma-1 single mutant (Figure 4). A possible exception was the reduced number of secondary branches in the kpc-1; dma-1 double compared to both single mutants and to wild type animals (Figure 4A). This synthetic phenotype reveals possibly a cryptic or parallel function for dma-1 in the formation of secondary branches that is only evident in the absence of kpc-1. To further investigate the genetic relationship between the menorin pathway and kpc-1, we used a previously described misexpression phenotype of mnr-1/menorin. When mnr-1/menorin is ectopically expressed in muscle, the resulting dendritic trees lack the characteristic menorah like morphology and, instead, display muscle-associated dendritic structures we named ‘baobabs’ because of their resemblance to baobab trees (Figure 8) [16]. We found that this mnr-1/menorin gain of function phenotype is fully dependent on kpc-1 (Figure 8) as it is on sax-7/L1CAM and dma-1/LRR [16], demonstrating that kpc-1 is required for the formation of mnr-1-dependent dendritic arbors. Taken together, these results suggest that for branch formation, kpc-1, mnr-1/menorin and sax-7/L1CAM act genetically in the same pathway and that dma-1/LRR is epistatic to kpc-1 in a similar way as it is to mnr-1/menorin and sax-7/L1CAM [16].
With regard to branch length, i.e. branch extension, the situation is slightly different. The kpc-1; dma-1 double mutant was indistinguishable from the kpc-1 and dma-1 single mutants, again demonstrating that both genes act genetically in the same pathway (Figure 5). In contrast, the reduced average length of higher-order branches in kpc-1 mutants was partially suppressed in double mutants with mnr-1/menorin or sax-7/L1CAM but not with dma-1/LRR (Figure 5B–E). Similar observations were made for the primary branches. Primary branch length was measured from cell body to pharynx border and normalized relative to total body length to account for possible variability in body size (Figure 5A, top). The shorter primary branch length in kpc-1 mutants was also partially suppressed by concomitant loss of mnr-1/menorin, sax-7/L1CAM, or dma-1/LRR (Figure 5A). Collectively, these results suggest that kpc-1 acts genetically in the same and, possibly in part, a parallel pathway to mnr-1/menorin and sax-7/L1CAM to control dendrite patterning.
In a screen for genes required for patterning of the PVD somatosensory neuron, we isolated mutant alleles in kpc-1, which encodes the C. elegans furin homolog [20], [21]. The phenotypes in PVD patterning that we observed in kpc-1/furin mutants displayed striking similarities to mutants in the ‘menorin’ pathway, which comprises the three cell adhesion molecules MNR-1/menorin, SAX-7/L1CAM and DMA-1/LRR-TM [15], [16]. Our double mutant analyses revealed genetic interactions that were distinct for branch formation and extension. For branch formation, we suggest that kpc-1/furin acts in a genetic pathway with mnr-1/menorin, sax-7/L1CAM and dma-1/LRR. First, double mutants between kpc-1/furin and either mnr-1/menorin or sax-7/L1CAM were not more severe than the single mutants. Second, dma-1/LRR appeared generally more severe than the kpc-1/mnr-1 or kpc-1/sax-7 double mutants and epistatic to kpc-1/furin. Third, gain of function experiments demonstrated that mnr-1/menorin function requires kpc-1/furin. Overall, these genetic interactions are strikingly similar to those between mnr-1/menorin, sax-7/L1CAM and dma-1/LRR [15], [16] and suggest that kpc-1/furin, mnr-1/menorin and sax-7/L1CAM act in a linear pathway but that a parallel pathway may exist that also acts through dma-1/LRR during branch formation. For branch extension, the scenario is slightly different. First, kpc-1/furin mutant branch extension phenotypes were generally more similar to dma-1/LRR than to mnr-1/menorin or sax-7/L1CAM mutant phenotypes. Second, mutations in mnr-1/menorin or sax-7/L1CAM but not dma-1/LRR partially suppressed kpc-1/furin mutant branch extension phenotypes in higher order branches. This could suggest higher activity of a parallel pathway during extension that is normally inhibited by mnr-1/sax-7-function, but also requires dma-1/LRR. Taken together, the findings presented here suggest that kpc-1 collaborates with the menorin pathway to sculpt the PVD dendritic arbor through distinct genetic mechanisms for branch formation and extension.
Our experiments showed that kpc-1/furin functions cell autonomously to coordinate formation, extension and self-avoidance of PVD somatosensory dendritic branches. However, the functions of kpc-1/furin are not limited to shaping PVD and FLP neurons during development. First, a recent report described a function for KPC-1 in the remodeling of sensory dendrites of IL2s (a set of sensory neurons in the head of C. elegans) as a result of changes in environmental conditions [21]. This report provided also evidence for a role of kpc-1 in the patterning of PVD and FLP neurons. Second, our detailed survey of the neuroanatomy of several classes of neurons in kpc-1/furin mutants revealed that kpc-1/furin plays a more general role in nervous system development than previously acknowledged. For example, AIY interneurons display neurite extension defects and the D-type motoneurons exhibit gaps in the dorsal nerve cord that appear to be the result of defects in neurite extension. On the other hand, kpc-1/furin is required for the formation of certain characteristic neuronal branches like in AQR sensory neurons or VC4/5 motoneurons, while preventing ectopic branching in other neurons such as the AVL neuron. In addition, kpc-1 functions during cell migration of the touch neuron ALM and HSN motoneurons. How could kpc-1/furin regulate such seemingly diverse developmental processes as formation and extension of neurite branches or neuronal positioning? One hint may come from our time-lapse analyses, which established that PVD branches grow slower in kpc-1/furin mutants compared to wild type animals. It remains unclear how KPC-1/Furin controls branch formation, extension or the speed of growth in such diverse cellular contexts on a mechanistic level. One possibility is that KPC-1 regulates extracellular adhesion of the neuron/growth cone to the substrate either directly or indirectly to mediate these functions, much like it has been suggested for metalloproteases (reviewed in [24]).
An important question is hence to determine the target(s) that are proteolytically processed by KPC-1/furin. In vertebrates, furin or furin-like proteases are known to cleave members of the TGFbeta family of morphogens as well as neuropeptides in the secretory pathway [17], [18]. Yet, mutations in genes required for neuropeptide processing and secretion, including the EGL-21/carboxypeptidase E, the PAMN-1/peptidyl-α-hydroxyglycine-α-amidating lyase [25] or unc-31/CAPS which is required for dense core vesicle secretion [26] did not result in comparable defects in PVD (Figure S6). Similarly, neither mutations in genes of the TGFbeta pathway (including the TGFbeta ligands tig-2, dbl-1, unc-129 or the sole type II TGFbeta receptor daf-4) nor in the C. elegans homolog of the repulsive guidance molecule (RGM), known to be cleaved by furin [27] displayed similar defects in PVD (data not shown). Several of the genes that have been shown to be required for PVD development contain predicted cleavage sites for furin-like proprotein convertases (Table S3). Thus, a candidate gene approach testing those genes and alternative genetic or proteomic approaches will be required to identify the in vivo target(s) of KPC-1 that are important for PVD dendritic arborization.
Worms were grown on OP50 Escherichia coli-seeded nematode growth medium plates at 20°C. Strains used in this work include: N2 (wild type reference), kpc-1(dz177), kpc-1(dz182), kpc-1(dz185), kpc-1(gk333538), kpc-1(gk779937), kpc-1(gk8), sax-7(nj48), sax-7(dz156), mnr-1(dz175) and dma-1(tm5159). PVD neurons were visualized by the integrated transgene wdIs52 (Is[F49H12.4::GFP]) and FLP neurons with muIs32 (Is[mec-7prom::GFP]). Transgenic strains for cell-specific rescue were established by injecting the respective plasmids at 5 ng/µl together with rol-6 (su1006) at 50 ng/µl as a dominant injection marker into kpc-1(dz182). The transcriptional reporter kpc-1prom5.8::GFP was injected at 5 ng/µl together with ttx-3prom::mCherry at 5 ng/µl into N2 wild type animals. For details on strains and transgenesis see also Supplementary Text S1.
In a clonal F1 Ethyl methanesulfonate (EMS) screen [16] we identified three alleles of kpc-1. Two alleles with similar phenotypes, dz177 and dz182 were mapped and cloned using a one-step whole genome sequencing approach [28]. Within the mapped region both dz177 and dz182 carried nonsense mutations in kpc-1 on chromosome I at positions 11,676,957 (C to T) and 11,679,245 (G to A) (WS220), respectively. One additional allele, dz185 failed to complement dz182 for the PVD phenotype and contained a missense mutation in kpc-1 at position 11,678,076 (A to T); three additional alleles were obtained from the C. elegans strain collection: the gk8 deletion allele and the missense alleles gk333538 and gk779937, which change 11,678,078 (G to A) and 11,678,071 (G to A), respectively. Transgenic animals carrying a wild type copy of the kpc-1 locus (fosmid WRM0635bG07) fully rescued the PVD defect in dz182 mutants (Figure 2C). For details on mapping and identifying the molecular lesions of different alleles see Supplementary Text S1, and Tables S1 & S2.
The kpc-1 cDNA was amplified with gene specific primers from a N2 mixed stage cDNA sample and cloned KpnI/EcoRI downstream of the ttx-3promB regulatory element [29]. For the cell specific heterologous rescue the kpc-1 cDNA was placed under control of the dpy-7prom, myo-3prom, or ser-2prom3 promoters, respectively. The transcriptional reporter was constructed by cloning 5.8 kb upstream of the predicted kpc-1 translational start site into pPD95.75 (gift of A. Fire).
Synchronized starved L1 larvae were allowed to grow for 30 hrs (corresponding to mid- to late L4) at which time they were mounted and fluorescent images of immobilized animals (1–5 mM levamisol, Sigma) were captured using a Zeiss Axioimager Z1 Apotome. Z stacks were collected and maximum projections were used for tracing of dendrites as described [16]. For time lapse imaging, animals at the L3 stage (by gonadal development) were immobilized as described [30]. PVD neurons were imaged for six to eight hours in 5 min intervals starting at the beginning of secondary branch development. Z-projections (0.5 µm/step) spanning the focal depth of the neuron were collected using a 63× objective. At least four movies per genotype were obtained using an inverted Nikon TE2000-S microscope equipped with a Perkin-Elmer UltraVIEW spinning disk unit. Volocity software (version 6.2.1) was used to collect the raw files. Processing was carried out using the Image-J 1.46r software.
|
10.1371/journal.ppat.1007644 | Genome-wide screen identifies novel genes required for Borrelia burgdorferi survival in its Ixodes tick vector | Borrelia burgdorferi, the causative agent of Lyme disease in humans, is maintained in a complex biphasic life cycle, which alternates between tick and vertebrate hosts. To successfully survive and complete its enzootic cycle, B. burgdorferi adapts to diverse hosts by regulating genes required for survival in specific environments. Here we describe the first ever use of transposon insertion sequencing (Tn-seq) to identify genes required for B. burgdorferi survival in its tick host. We found that insertions into 46 genes resulted in a complete loss of recovery of mutants from larval Ixodes ticks. Insertions in an additional 56 genes resulted in a >90% decrease in fitness. The screen identified both previously known and new genes important for larval tick survival. Almost half of the genes required for survival in the tick encode proteins of unknown function, while a significant portion (over 20%) encode membrane-associated proteins or lipoproteins. We validated the results of the screen for five Tn mutants by performing individual competition assays using mutant and complemented strains. To better understand the role of one of these genes in tick survival, we conducted mechanistic studies of bb0017, a gene previously shown to be required for resistance against oxidative stress. In this study we show that BB0017 affects the regulation of key borrelial virulence determinants. The application of Tn-seq to in vivo screening of B. burgdorferi in its natural vector is a powerful tool that can be used to address many different aspects of the host pathogen interaction.
| Borrelia burgdorferi, the causative agent of Lyme disease, must adjust to environmental changes as it moves between its tick and vertebrate hosts. We performed a screen of a B. burgdorferi transposon library using massively parallel sequencing (Tn-seq) to identify fitness defects involved in survival in its tick host. This screen accurately identified genes known to cause decreased fitness for tick survival and identified new genes involved in B. burgdorferi survival in ticks. All of the genes tested individually confirmed the Tn-seq results. One of the genes identified encodes a protein whose function was previously unknown that appears to be involved in regulating expression of proteins known to be involved in environmental adaptation. Tn-seq is a powerful tool for understanding vector-pathogen interactions and may reveal new opportunities for interrupting the infectious cycle of vector-borne diseases.
| Lyme disease is caused by the spirochete, Borrelia burgdorferi. In nature, B. burgdorferi is maintained in a cycle between mammalian or bird hosts and Ixodes ticks [1] Newly hatched ticks can acquire B. burgdorferi from infected animals during their larval feeding [1]. After molting to the nymphal stage, those infected ticks can transmit the pathogen to a new vertebrate host during their next blood meal [1]. The challenges posed by the vertebrate and tick environments are quite different. B. burgdorferi must adapt to changes in temperature, pH, nutrient availability and immune defense mechanisms between its vertebrate and arthropod hosts [2–6].
Previous studies have shown that B. burgdorferi adapts to its host environments through controlling the expression of proteins that aid in survival at specific points in its life cycle in its different hosts [7–9]. For example, proteins such as outer surface protein C (OspC), variable-major- protein (Vmp)-like sequence E (VlsE) and decorin binding protein A (DbpA) are expressed to differing amounts during particular time points in the mammalian and tick phases of the B. burgdorferi life cycle [10–14]. The regulation of gene expression in B. burgdorferi is complex, often involving multiple layers of control [1,3,6].
Expression of proteins required during the mammalian phase involves two alternative sigma factors, RpoS and RpoN, the enhancer binding protein Rrp2, as well as the transcription factors BosR and BadR [15–26]. In addition to controlling virulence gene expression, BosR also controls expression of genes involved in resistance to reactive oxygen species and affects metal homeostasis, while BadR controls expression of many genes involved in metabolite uptake and utilization [17,27, 28]. Other regulators such as carbon storage regulatory protein A (CsrA) appear to exert their effects outside the RpoS/RpoN axis [29].
Much less is known about gene regulation and proteins critical for B. burgdorferi survival while in its tick host [6]. Histidine kinase 1 (Hk1) and response regulatory protein 1 (Rrp1) are highly expressed during the tick phase and appear to work together to regulate expression of genes involved in tick survival [30–32]. Rrp1 is a diguanylate cyclase required for the synthesis of cyclic diguanylate (c-di-GMP), an important second messenger signaling molecule in B. burgdorferi and other bacteria [32–35] The exact mechanisms by which Hk1 is activated and how Rrp1 is regulated are not known.
Proteins that have been shown to be important in survival in ticks include outer surface protein A (OspA), which binds to the tick mid gut protein TROSPA [6,36]. GuaA and GuaB, two enzymes that contribute to the purine salvage pathway, have also been shown to provide a fitness advantage in the tick host [37]. The glycerol utilization operon (glpF, glpK, glpD) encodes proteins that allow the bacterium to utilize glycerol as the carbohydrate source for glycolysis [33,38]. This operon is upregulated during all tick life cycle stages, and has been shown to be specifically involved with persistence and survival of the molt, but not early colonization [33,35,38]. Another protein shown to be essential for infection of the tick host is the manganese transporter BmtA. This transporter is required for B. burgdorferi to colonize and survive in ticks [39].
In this study, we describe the use of transposon insertion sequencing (Tn-seq) to identify genes that are critical for B. burgdorferi survival during infection of Ixodes scapularis, the tick vector most commonly associated with Lyme disease transmission in North America [1]. Tn-seq is a high- throughput approach that enables the quantification of the frequency of individual transposon (Tn) mutants in a population before and after a selective pressure [40]. Tn-seq has been widely used for in vitro assays of bacterial fitness [41–44]. It has also been used to perform in vivo studies in mice, although in vivo Tn-seq studies are often limited by tight bottlenecks causing stochastic loss of mutants unrelated to the Tn insertion [40–43]. This report represents the first use of a Tn insertional library combined with massively parallel sequencing to identify bacterial genes involved in colonization of an arachnid. Using Tn-seq, we were able to accurately identify a number of B. burgdorferi mutants with impaired fitness for survival in Ixodes ticks. The process is easily scalable though testing additional ticks, which reduces misidentification of mutants that are lost for reasons other than fitness. As opposed to mammalian studies, in which the number of animals is often limiting, we were able to readily screen very large numbers of larval ticks, thereby mitigating bottleneck issues. As part of our studies, we have identified a potential new regulator of B. burgdorferi gene expression, BB0017, which may contribute to expression of genes involved in tick and mammalian survival.
Mice were bred and maintained in the Tufts University Animal Facility. All experiments were performed following the guidelines of the American Veterinary Medical Association (AVMA) as well as the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All procedures were performed with approval of the Tufts University Institutional Animal Care and Use Committee (IACUC, Protocol# B2015-159). Euthanasia was performed in accordance with guidelines provided by the AVMA and was approved by the Tufts University IACUC.
B. burgdorferi strains were grown in Barbour-Stoenner-Kelley II (BSK-II) medium in sealed tubes at 32°C with 1% CO2. Escherichia coli strains (Top10) for plasmid preparation were grown on Lysogeny broth (LB) agar plates or in LB broth at 37oC. E. coli cultures contained either 50 μg/ml spectinomycin or 10 μg/ml gentamicin. The parental strain of the Tn library, the infectious B. burgdorferi strain 5A18NP1, was used as the wild-type strain in all studies and lacks two plasmids (lp56 and lp28-4) [45]. The following antibiotics were used for selection in cultures of B. burgdorferi when appropriate: kanamycin at 200 μg/ml, gentamicin at 40 μg/ml, and streptomycin at 50 μg/ml. Tn mutants were obtained from the arrayed B. burgdorferi library [45]. All individual Tn mutants used in this study were screened by polymerase chain reaction (PCR) at the locus of interest to confirm pure populations as previously described [42, 43]. In cases where mixed populations were identified (i.e. two PCR products indicating the presence of both a wild-type and Tn disrupted locus), the strain was plated for single colonies in semi-solid agarose overlays. Individual colonies were then selected and re-screened to confirm pure populations. All Tn mutants were subsequently plasmid- typed to identify the loss of any plasmids required for murine or tick infection [46, 47]. A description of all individual Tn mutants used in this study is available in Table 1. Other B. burgdorferi strains as well as plasmids used in this study are also described in Table 1.
Infection of ticks with B. burgdorferi from the mutant library was performed using a method previously described [48]. Briefly, before immersion, spirochete cell density was determined by dark field microscopy. Cell suspensions were centrifuged for 10 min at 8,000 x g and were resuspended at the desired cell density and in the desired medium. I. scapularis larvae were obtained from the National Tick Research and Education Resource at Oklahoma State University and were maintained in a humid tick incubating chamber at room temperature. Larvae were used within 4 months of emergence.
Before immersion, I. scapularis larvae were removed from the chamber and allowed to sit at ambient humidity in an air-conditioned room for 2h. The larvae were then transferred with a small brush into 1.5 ml microcentrifuge tubes. B. burgdorferi culture suspension of 108 bacteria in 1 ml was then added to the ticks. For the Tn-seq studies, the suspension consisted of the B. burgdorferi Tn mutant library [42, 43]. For confirmatory experiments, individual mutants, mixtures of mutants and complemented mutants, or mixtures of mutants and controls were used in the suspension. The tubes were gently vortexed to suspend larvae in the culture and incubated at 32°C for 1 h. Tubes were gently vortexed every 15 min to redistribute ticks in the culture. After incubation at 32°C, tubes were centrifuged at 200 x g for 1 min. The supernatant was removed, and ticks were washed once with phosphate-buffered saline (PBS).
Larvae were then transferred from the microcentrifuge tube to a sealed mouse restraining device. The mice were of C57BL/6 background and were all females aged 4-6 weeks old. A mouse was placed into the restrainer containing the larvae, and the larvae were allowed to attach. Mice were removed from the restrainer after 30 min and transferred to cages suspended over water moats. Engorged larvae were collected from the moats 3 to 5 days after placement and transferred into a tick incubation chamber.
Ticks were collected as they completed feeding from the animals. Cages were checked daily to collect ticks. Ticks were batched and processed for Tn-seq up to 48 hrs after collection [48]. Briefly, ticks were washed in 3% hydrogen peroxide, 70% ethanol, and finally in PBS. The ticks were allowed to dry before placement in 500 μl of BSK-II medium with kanamycin and gentamicin. To isolate spirochetes from the ticks, the ticks were crushed with a plastic pestle (Fisherbrand RNase- Free Disposable Pellet Pestles). These tick homogenates were inoculated into 5 ml of BSK-II containing kanamycin and gentamicin. The cultures were incubated for two days to allow outgrowth. Following this, the spirochetes were pelleted by centrifugation for 10 min at 8,000 x g. Bacterial pellets were washed once in PBS, and then the dry pellet was stored at -80°C until further processing for Tn-seq.
Genomic libraries for sequencing were constructed as described previously [42,43]. Chromosomal DNA was sheared using the M220 Focused-ultrasonicator (Covaris) in microTUBEs with a target peak at 350 bp. The first round of PCR amplification was performed using a modified primer with optimized annealing to the Tn (pMargent1A, 5'-ggtaccttaggagaccgggg-3')[43]. Libraries were multiplexed and pooled for analysis. Sequencing was performed on an Illumina HiSeq 2500 at the Tufts University Core Facility as 50-bp single-end reads, as described previously [42,43].
Sequenced reads were clustered by barcode sequence. Data analysis were performed using the Galaxy platform and followed a previously published protocol [43]. We obtained an average of 1.2 x 107 reads per barcode, with 1.7 x 106 reads per condition for analysis after removal of low quality sequences [43]. Reads were mapped to the B. burgdorferi B31 genome using Bowtie, and a custom script was used to count the number of sequence reads corresponding to each insertion site in the genome. Sequence reads were analyzed “by-site” and “by-gene”. Only Tn mutants that were represented by at least ten sequence reads in both input samples were included in the “by site” analysis. In contrast, the “by-gene” analysis included all sequence reads mapping within a particular gene. Genes represented by less than ten sequence reads in both untreated samples were excluded from the “by-gene” analysis. Tn mutants with zero reads in the output samples were assigned a value of one for the purpose of calculation. The frequency of each Tn mutant in a particular condition was determined by dividing the number of sequence reads corresponding to each Tn mutant by the total number of sequences in the barcode. A frequency ratio was then determined by dividing the frequency of a Tn mutant in the output (bacteria recovered from the ticks) sample by its frequency in the input population. For the purposes of prioritizing mutants for follow-up, frequency ratios between 0.5 and 2 were considered neutral.
A plasmid for complementation of the Tn::bb0164 mutant was generated previously by overlap PCR [43]. A plasmid for directing cis complementation of Tn::bb0412 via allelic exchange was generated from three PCR fragments: intact bb0412 with 995 bp upstream sequence (F1), PflgB-aadA for antibiotic selection (F2), and 1022 bp downstream of bb0412 (F3). Primers were designed with approximately 30 bp on the 5' end for overlap, with the numerically assigned PCR products assembled in order into the final construct. See S1 Table for a list of all primer sequences used in this study. Individual PCR fragments were amplified with AccuPrime Pfx (ThermoFisher Scientific, MA) per the manufacturer's instructions. An overlap PCR was performed with equal volumes of the appropriate number of PCR fragments with AccuPrime Pfx reagents per the manufacturer's recommendations. Following PCR, the products were resolved by agarose gel electrophoresis and gel-purified (Zymoclean Gel DNA Recovery Kit, Zymo Research, Irvine, CA) for cloning into pCR-Blunt (ThermoFisher Scientific, Grand Island, NY) following the manufacturer's protocol. The bb0412 complementation vector was designated pAK412 (Table 1).
Plasmid pMR05 was constructed for trans complementation of the Tn::bb0017 mutant was generated by amplifying a DNA sequence containing the bb0017 open reading frame and the 298-bp upstream region from 5A18NP1 genomic DNA via PCR using the primers bb0017FLAG-F-SacI and bb0017FLAG-R-XbaI (S1 Table). The resulting PCR products as well as pKFSS1 were digested with SacI and XbaI, and the resulting PCR product and pKFSS1 fragments were ligated together using T4 DNA Ligase (New England Biolabs), generating pMR05 (Table 1).
Plasmid pBS01 was constructed to direct allelic exchange at the bb0017 locus, resulting in deletion of the bb0017 open reading frame (Table 1). Plasmid pBS01 contains 1835 bp of sequence upstream of bb0017 (amplified from genomic DNA using primers bb0017del1 and bb0017del2), followed by a sequence containing the constitutive PflgB promoter and a streptomycin resistance gene (aadA, amplified from pKFSS1 using primers bb0017del3 and bb0017del4), followed by 1999 bp of sequence downstream of bb0017 (amplified from genomic DNA using primers bb0017del5 and bb0017del6). PCR products were purified using the Qiagen PCR purification kit and were subsequently assembled into the pBlueScript cloning vector using the NEBuilder HiFi DNA Assembly Cloning Kit.
Plasmid pBS02 was constructed to direct expression of bb0017 in the Δbb0017 mutant. Plasmid pBS02 is identical to pMR05, except for replacement of the streptomycin resistance cassette by a gentamicin cassette (Table 1). The streptomycin cassette was excised from pMR05 by restriction digest with AatII and NdeI. The gentamicin cassette had been previously subcloned into the pCR2.1cloning vector. The gentamicin resistance gene was excised from this vector using the same restriction enzymes as above, and the resulting fragment was ligated with the digested pMR05 backbone (Table 1).
Plasmid pMH88R was used to constitutively express the glp operon in the Δbb0017 mutant. Plasmid pMH88R was a kind gift of Dr. Frank Yang [33]. All completed plasmids were verified by restriction digest and dideoxy sequencing.
Plasmids were introduced into B. burgdorferi by transformation as previously described [50,51]. Cis complementation vector pAK412 was transformed into B. burgdorferi Tn::bb0412, and transformants were designated AK102 (Table 1). The complemented strain was screened by PCR for allelic exchange using the forward primer of fragment 1 and the reverse primer for the PflgB- aadA cassette for each construct. The bb0017 deletion construct pBS01 was transformed into B. burgdorferi strain 5A18NP1, generating strain Δbb0017 (Table 1). In order to complement the bb0017 mutation, plasmid pBS02 was introduced into the Δbb0017 background, generating strain Δbb0017 + bb0017 (Table 1). The overexpression of the glp operon construct pMH88R was transformed into the Δbb0017 mutant and was used to generate the Δbb0017 + glpFKD strain (Table 1). Potential transformants were confirmed by PCR with primers designed to detect either a replicating plasmid or a double crossover event, as appropriate, followed by dideoxy sequencing of the PCR product to confirm the expected nucleotide sequence.
To evaluate survival in the tick host, individual Tn mutants of interest were combined with their respective complemented strain or wild type bacteria in a 1:1 mixture. Each strain was grown independently. Cell density was determined by microscopy, and 5 x 107 B. burgdorferi were harvested by centrifugation. The pellets from both cultures were then resuspended in the same 1 ml of BSK-II medium to a final overall density of 1 x 108 cells/ml. Ticks were submerged in the cultures as described, placed on mice for feeding, and collected over three days as described above. Each tick was washed successively with 3% hydrogen peroxide, 70% ethanol and PBS, then crushed into 250 μl of BSK-II containing kanamycin and 5 μg/ml amphotericin B. The cultures were allowed to acclimate in liquid medium for a period of 2 h before plating, to allow the bacteria to escape from the crushed tick into the medium. Plating in semi-solid agarose overlay was performed as previously described [50]. These plates were then sealed in plastic bags and placed at 32°C for 10 days. The plates were removed from the incubator and colonies were enumerated. The ratio of the wild-type or complemented strains to the Tn mutant was determined by counting the colonies on the appropriate antibiotic selective plates. A competitive index was calculated for these experiments by dividing the amount of mutant recovered by the amount of complement or WT that was recovered. In the case where no mutant was recovered, its value was set to one for the purposes of calculation of the competitive index.
Two independent cultures of the 5A18NP1 and Tn::bb0017 strains were grown to mid-logarithmic phase, washed once in PBS, and resuspended in BSK II medium at a concentration of 6 × 107 bacteria/ml. Cultures were incubated at 32°C with 1% CO2 for 1 h. The bacterial pellet was harvested by centrifugation at 9500 × g for 5 min at 4°C, washed once in ice-cold PBS, and frozen on dry ice. RNA was isolated using the miRNeasy kit (Qiagen). The TURBO DNA-free kit (Invitrogen) was used to remove contaminating genomic DNA. Library preparation for RNA-seq analysis was conducted at the Tufts University Genomics Core Facility. Samples were depleted of rRNA using the Gram-Negative Ribo-Zero rRNA Removal Kit (Illumina). Strand-specific libraries were prepared using the TruSeq Stranded mRNA Library Prep Kit (Illumina). Sequencing was performed on an Illumina HiSeq 2500 at the Tufts University Core Facility as 50-bp single-end reads.
Data analysis were performed using the Galaxy platform [43]. Sequence reads were aligned to the B. burgdorferi B31 genome using TopHat for Illumina v1.4.1 with the default settings. Differences in transcript expression were determined using CuffDiff, again using the default settings.
Total RNA was extracted from bacterial cells grown to mid-logarithmic growth phase at 32°C using TRIzol (Invitrogen) following the manufacturer’s instructions. RNA samples were treated with the TURBO DNA-free kit (Invitrogen) to remove contaminating DNA. cDNA was prepared using random hexamers (Promega) and the ImProm-II Reverse Transcription System (Promega). Control reactions were performed in the absence of reverse transcriptase to control for the presence of genomic DNA. Sequences of the primers used to determine the differential expression of target genes are listed in S1 Table, and expression levels were normalized against those of the B. burgdorferi housekeeping gene flaB.
Quantification of target genes from cDNA was performed using the iTaq Universal SYBR Green Supermix (BioRad)[43]. Samples were run in duplicate or triplicate. Analysis of the RT-qPCR data was conducted using the ΔCT method. Data collection was performed using the CFX Connect Real- Time PCR Detection System (BioRad).
Strains were grown in BSK-II at 32°C/1% CO2 until cultures reached early stationary phase. A volume corresponding to 1 × 108 bacteria was harvested by centrifugation at 9500 × g for 5 min at 4°C. The bacterial pellet was frozen at -80°C until processing. To lyse the cells, the bacterial pellet was resuspended in approximately 100 μl of 1X NuPAGE buffer (ThermoFisher) and boiled for 5 min. A 20 μl volume of each lysate was electrophoresed in 4-15% gradient SDS-PAGE gels (BioRad). Proteins were transferred to a polyvinyldifluoride (PVDF) membrane (Trans-Blot Turbo BioRad). Membranes were blocked in 5% milk in Tris-buffered saline containing 0.05% Tween-20 (TTBS). Primary antibodies were diluted as follows in TTBS: anti-RpoS (1:50, courtesy of Dr. Frank Yang), anti-BosR (1:500, courtesy of Dr. Frank Yang), anti-FlaB (1:1000, courtesy of Dr. Xin Li), anti-OspC (1:10,000, courtesy of Dr. Xin Li), anti-OspA (1:1000, Rockland Immunochemicals) and anti- DbpA (1:1000, Rockland Immunochemicals). Appropriate horseradish peroxidase (HRP)- conjugated secondary antibodies were used at a 1:10,000 dilution in TTBS. Detection was performed using the Luminata Forte substrate (Millipore), followed by exposure to film (Denville Scientific) or imaging using a ChemiDoc XRS+ Imager.
In order to use Tn-seq to determine genes involved in B. burgdorferi fitness for survival in ticks, we first needed to decide the number of ticks to include in the experiments. Using immersion feeding, it has previously been shown that approximately 103 B. burgdorferi could be recovered per tick [48,52]. However, these experiments examined a single strain of bacteria, and since bacteria were allowed to replicate within the tick before recovery, the actual number of bacteria entering to establish colonization was likely less than 103. We were also concerned that the bottleneck may be further exacerbated by plasmid loss in the parental strain of the transposon library. It has previously been reported that the mean number of spirochetes per tick with 5A18NP1 lacking both lp28-4 and lp56 plasmids was lower than that in ticks infected with B. burgdorferi harboring all plasmids [53]. The ordered transposon library contains insertions into 45.5% of the predicted protein encoding genes [45]. To ensure sufficient coverage of the pooled library of around 4,000 mutants we chose to target approximately 150 ticks per experiment. If only half the predicted number of bacteria established infection (500 bacteria per tick), this approach should still provide approximately 20- fold coverage of the mutants within our library.
In each of two independent experiments, approximately 300 larval ticks were immersed in a culture solution containing the entire Tn library and then fed on mice. This resulted in the collection of approximately 160 fed ticks per experiment. Fed ticks were processed and cultured in BSK medium for 2 days. The input culture was also cultured for an additional 2 days to match the tick cultures. Bacteria from both cultures were harvested and sequencing libraries prepared.
Reproducibility was high between the two input libraries (Fig 1A, Pearson coefficient R2=0.98). The correlation between the Tn frequencies of the populations recovered from the two groups of tick larvae was also high with a Pearson coefficient R2= 0.85 (Fig 1B). A frequency ratio was calculated for each Tn mutant in the library by comparing its frequency in the output library to its frequency in the input library. For analysis, we included only Tn mutants represented by at least 10 sequence reads (out of a total of approximately 1.7 x 106 sequence reads per experiment) in both replicates of the input libraries. This number was chosen to reduce the risk of stochastic loss after selection in the ticks. Tn mutants that were represented in the input library but had zero reads in the output library were assigned a value of one read in order to be able to calculate a frequency ratio. A frequency ratio less than one indicates that the Tn mutant decreased in frequency after recovery from fed ticks suggesting that the disrupted gene is involved in survival in the larval host. A frequency ratio greater than one indicates that the Tn mutant increased in frequency after larval colonization suggesting that the disruption of the corresponding gene provides a fitness advantage. An overall frequency ratio was also calculated for each gene by aggregating all of the sequence reads mapping to Tn insertions within the same gene (S1 File). A complete list of mutant fitness for larval colonization by gene and by site is provided in S1 File.
In order to validate our screen, we began by analyzing genes that have been previously shown to be essential for tick survival, to ensure that these genes had been identified in the screen. Borrelial genes that have been described in the literature as critical to tick survival for which mutants are present in the library include: bb0419 and bb0420, respectively encoding a response regulator designated Rrp1 and a histidine kinase designated Hk1 [30,35]; guaA and guaB, two genes involved in the purine salvage pathway [37]; glpD, encoding a glycerol 3-phosphate dehydrogenase [33,38]; and bptA, a surface-expressed lipoprotein [8]. In the Tn-seq experiment, consistent with previously published results, insertional mutants in each of these genes except glpD showed attenuated ability to survive in the tick, with median frequency ratios of <0.1 (Fig 2A). GlpD has been shown to be important following the molt from larvae to nymph when carbon sources are less abundant and the organism begins to utilize glycerol, which may explain why Tn::glpD mutants did not exhibit phenotypes in our screen [35].
Further analysis was performed to identify new genes involved in tick survival. A large number of mutants (N=309) had a frequency ratio of less than 0.5 compared with the input library. In order to prioritize mutants with the strongest phenotypes for follow-up analysis, we focused on mutants with a fitness ratio of less than 0.1. Mutants with insertions in 102 genes had an average overall frequency ratio below 0.1 in both experiments (S1 File). However, this group of 102 mutants included some with insertions into genes that have been shown not to be required for tick colonization (e.g. ospC). While many of this group of 102 genes may be involved in tick survival as it includes many of the genes previously identified as involved in tick colonization, to further reduce the chance of false discovery by the screen, we increased the stringency of our criteria and focused on the subset of 46 genes that had >100 reads in the input library but were completely absent in the processed ticks from both experiments (Table 2). These genes would be predicted to have the greatest impact on fitness for survival in larval ticks. Of these 46 genes, many have no predicted function and have not been previously characterized (Fig 3). Approximately 22% of the genes are predicted lipoproteins, while 7% are involved in carbohydrate transport (Fig 3).
To confirm the results of the Tn-seq screen, we chose mutants with insertions in five genes (bb0017, bb0164, bb0412, bb0050 and bb0051) that showed the strongest fitness defects, and that have not previously been reported to be involved in tick survival. Each of these genes was well represented by insertion mutants in the input library and had an overall frequency ratio of less than 0.1 following the ingestion of a blood meal by larval ticks. Competition assays were conducted to assess the capability of each individual mutant to survive the larval blood meal (Fig 2B).
Three of the transposon mutants were competed against a complemented strain, while the remaining two transposon mutants were competed against the parental strain (Fig 2B). The Tn::bb0017, Tn::bb0164, Tn::bb0412, Tn::bb0050 and Tn::bb0051 mutants were all outcompeted by the parental or respective complemented strains, confirming a role for all five genes in blood meal survival (Fig 2B). We were also able to further confirm this phenotype when competing a bb0017 clean deletion strain against its complemented strain. (Fig 2B). The complemented strain greatly outcompeted the deletion strain confirming the role of bb0017 in surviving the blood meal (Fig 2B).
The Tn-seq and competition experiments do not distinguish between the possibilities that 1) the identified genes are required for initial entry into the tick during immersion; or 2) they are required for surviving the blood meal taken by the tick. The competition experiment that was described previously was modified so that the ticks were not allowed to take a blood meal after immersion in the culture containing the two competing strains, allowing us to separate fitness defects due to uptake from fitness defects due to blood meal survival. The ticks were crushed after two hours of immersion feeding or kept overnight and crushed 24 hours post-immersion. The relative frequencies of the mutant and complemented or parental strains were then determined as before. However, in contrast to the competitive defect exhibited by all five Tn mutants after the blood meal, we were able to recover all Tn mutants after immersion feeding in equal or greater numbers compared to the WT or complemented mutants. However, in the absence of a blood meal, as expected, the numbers of bacteria were greatly reduced and B. burgdorferi was not recovered from all individual ticks. The results of these studies are shown in Table 3. These data support a role for these five borrelial genes in surviving changes associated with the blood meal. Also, importantly, because several of these genes have identified roles in ROS resistance and hydrogen peroxide was used in washing the ticks, these experiments confirm that the hydrogen peroxide wash did not affect selection of these mutants.
To begin to better understand the mechanisms by which the genes identified in the Tn-seq screen contribute to tick-phase survival, we performed further investigation into one of the genes identified as critical for survival of the blood meal: bb0017. The gene bb0017 was previously identified in a screen for genes that confer resistance to ROS [43]. BB0017 is highly conserved among both B. burgdorferi sensu stricto and other sensu lato strains (>99% and >94% identity at the amino acid level, respectively). BB0017 homologues are also conserved in the relapsing fever strains (>80% identity) [54]. In the B. burgdorferi strain B31, bb0017 is annotated an integral membrane protein of the YitT family. BB0017 contains four predicted transmembrane domains as well as a C-terminal soluble domain and contains the conserved domain of unknown function DUF2179 (S1A & S1B Fig) [55]. A structure-based similarity search using Phyre2 suggested that the C-terminal domain of BB0017 is structurally similar to PII and PII-like proteins, despite low overall sequence identity (<27% identity) [56]. No high confidence predictions were made for the N-terminal domain of BB0017.
PII proteins are a broadly conserved class of signal transduction proteins found in bacteria, archaea, and plants and are generally small cytoplasmic proteins involved in nitrogen metabolism [57–59]. PII proteins generally function as trimers and control the activity of their regulatory targets through direct protein-protein interactions in response to both post-translational modifications (such as uridylylation) and ligand binding (including ADP, ATP, and 2-oxoglutarate). The long, flexible T- loop mediates interactions with regulatory targets, while a conserved motif in the shorter B-loop is involved in ligand binding (S1C Fig, GlnKEc). More recently, several PII-like families of proteins have been identified in bacteria, including a family of proteins in Gram-positive bacteria that bind cyclic diadenylate monophosphate (c-di-AMP) as well as a broadly conserved family of proteins (CutA) that confer copper tolerance in Escherichia coli and bind acetylcholinesterase in mammals [60–66].
While the PII and PII-like proteins share a common ferredoxin-like fold, the lengths of the T and B loops differ significantly between the different protein families. In the case of the PII-like c-di-AMP binding proteins, the lengths of the B and T loops are reversed relative to the PII proteins and are referred to as the B´ and T´ loops (PstASa, S1C Fig). Structural data suggests that the functions of the B´ and T´ loops are also reversed relative to the PII proteins, with the short T´ loop being involved in ligand binding and the long flexible B´ loop possibly involved in effector binding [61,62,64]. In the case of the copper tolerance protein CutA1, both the B and T loops are truncated, and the same is true for BB0017. Interestingly, BB0017 appears to lack conserved residues involved in ligand binding by both the PII and PII-like protein families. The presence of the N-terminal transmembrane domain also distinguishes BB0017 from the PII and PII-like protein families and suggests that membrane localization may be important for BB0017 function.
Because BB0017 contains a putative signal transduction domain, we hypothesized that the Tn::bb0017 mutant would exhibit global differences in gene expression compared to the parental strain. Total RNA was isolated from the parental and Tn::bb0017 strains, and RNA sequencing (RNA-seq) was used to compare the transcriptomes of both strains. We identified 16 genes that were significantly downregulated more than twofold in the Tn::bb0017 mutant compared to the parental strain and 25 genes that were significantly upregulated more than twofold (S2 Table and Fig 4). It is important to note that bb0017 does not appear in S2 Table. While expression of bb0017 was significantly different between the Tn::bb0017 and parental strains, the difference was less than twofold, and bb0017 expression was actually higher in the Tn::bb0017 mutant compared to the parental strain. Sequence coverage maps confirm that transcription in the Tn::bb0017 mutant is abrogated downstream of the Tn insertion as expected (S2 Fig). However, increased numbers of sequence reads mapped in the 5’ portion of bb0017, likely due to transcription from the strong PflaB promoter contained within the Tn (S2 Fig). It is unclear whether there is translation of the 5’ portion of bb0017 in the Tn::bb0017 mutant, resulting in a truncated protein, but if a truncated protein is produced in the Tn::bb0017 mutant, these results could suggest that the C-terminal portion of BB0017 is the critical portion for survival in the tick.
Strikingly, the putative bb0017 regulon overlaps significantly with that of RpoS, a key regulator of virulence gene expression in B. burgdorferi [15,16,19,20]. RpoS is directly responsible for upregulating a number of genes required for survival in the mammalian host, including dbpA, dbpB, ospC, and bbk32 and repressing expression of genes important for tick survival such as glpD [15,19,20,38,67]. The dbpA, dbpB, ospC, and bbk32 genes are all upregulated in the Tn::bb0017 mutant (S2 Table, Fig 4). Several genes known to be subject to RpoS-mediated repression, including genes located within a glycerol utilization operon important for tick infectivity, are downregulated in the Tn::bb0017 mutant (bb0240-bb0243, S2 Table) [38]; however, other regulators such as c-di-GMP may also affect expression of these genes.
To confirm the results of the RNA-seq screen, we generated a mutant lacking the entire bb0017 open reading frame. A survey of available B. burgdorferi genome sequences revealed two different annotated start sites (S1 Fig). We chose to delete the region encompassing the first start site, which includes a putative 71 bp small RNA (SR0011) in the bb0016-bb0017 intergenic region [68] (S1 Fig). We restored bb0017 expression including the upstream SR11 intergenic region under the control of the native promoter from a replicating plasmid in the Δbb0017 mutant, and confirmed expression by qRT-PCR (S3 Fig).
We performed qRT-PCR on the Δbb0017 mutant as well as the complemented strain to validate the results of the RNA-seq using the transposon insertion strain. We selected a subset of differentially regulated genes from the RNA-seq, as well as some representative regulatory proteins of B. burgdorferi that showed no change in expression. bosR and rpoS levels were not significantly different in the RNA-seq and this phenotype was reproduced in the deletion strain as well as the complement by qRT-PCR (Fig 5). We then confirmed six genes, ospC, dbpA, glpD, bba37, bba25, and bbk32 that were differentially expressed by RNA-seq in the transposon mutant strain. Each showed a similar pattern of expression in the clean deletion strain with recovery in the complemented mutant strain, with the exception of glpD (Fig 5). Transcription of glpD was decreased in the transposon mutant and its complement as well as the deletion strain and its complement in comparison to the wild type making it unlikely that this difference was due to secondary site mutations or polar effects as each of the mutants and complements were created from separate isolations from the parental strain. Of note is that all the changes are small (fourfold) compared to the other genes tested by qRT-PCR. Given the lack of involvement of glpD involvement in tick survival as assayed by the Tn-seq and its established role at a different stage in tick survival, it is likely that the change is not physiologically relevant.
To ensure that we were not missing a role for GlpD in mediating effects of the Δbb0017 mutant, we created a strain that overexpresses the entire glp operon, including glpFKD, in the Δbb017 deletion strain. This construct has previously been used to successfully overexpress GlpD [33]. We confirmed that glpD was successfully transcriptionally over-expressed by qRT-PCR (Fig 5). Using this strain, we then performed a competition experiment between the glp operon overexpressing strain and the Δbb0017 strain. We were not able to recover either strain from this experiment following selective plating from six collected fed larvae. This indicates that increasing the ability for glycerol utilization is not sufficient to rescue the Δbb0017 mutant and the defect in tick colonization is unlikely to result strictly from decreased expression of glpD.
We performed immunoblots for OspC and DbpA in the Tn::bb0017 and Δbb0017 mutants. Levels of both DbpA and OspC were elevated in the Tn::bb0017 and Δbb0017 mutants, confirming the results of the RNA-seq screen (Fig 6A). Restoration of bb0017 expression from a replicating plasmid (which also contains SR0011) in both mutants decreased DbpA and OspC levels to those of the parental strain (Fig 6A). The fact that both the Tn::bb0017 mutant (in which SR0011 remains intact) and the Δbb0017 mutant (in which SR0011 is disrupted) exhibit increased lipoprotein expression, suggests that bb0017 is required for the phenotype, although these results to do not exclude the possibility that SR0011 may also be involved. Expression of OspA, a surface lipoprotein required for infectivity in the tick, was not affected by the absence of bb0017 under the conditions tested.
The expression of dbpA and ospC requires the alternative sigma factor RpoS [15,20]. The regulation of rpoS in turn involves a second alternative sigma factor RpoN (σ54), the enhancer binding protein Rrp2, the Borrelia oxidative stress response regulator BosR, and the small regulatory RNA DsrA [15,16,24,25,69,70]. We hypothesized that BB0017 mediates the repression of ospC and dbpA indirectly by affecting the expression of an upstream regulator. We therefore investigated RpoS levels in the Tn::bb0017 and Δbb0017 mutants. RpoS levels were increased in both mutants, and restoration of bb0017 expression resulted in decreased RpoS levels (Fig 6B). To understand the mechanism by which BB0017 affects RpoS expression, we next investigated production of BosR, a positive regulator of RpoS [24,70]. As was the case for RpoS, BosR levels were increased in the Tn::bb0017 and Δbb0017 mutants, and complementation of bb0017 restored levels to those similar to the parental strain (Fig 6B).
In this paper, we report the use of massively parallel, next generation sequencing technology to identify genes important in survival in the larval tick host. This study represents the most complete survey of B. burgdorferi genes that are required for tick survival performed to date and greatly increases our understanding of this critical phase of the B. burgdorferi life cycle. We have identified many genes that have not previously been associated with tick survival, confirmed the involvement of a subset in tick survival, and began to characterize a mechanism of action for one of the genes, bb0017. Notably, because we were able to quickly and inexpensively screen large numbers of ticks, we were able to minimize bottleneck issues that have arisen in other animal studies, and our results showed a high level of experimental reproducibility. The robustness of the technique is exemplified by our ability to identify genes known to be required for tick-phase survival and our ability to validate the phenotypes of all five mutants we selected for further analysis.
There are several caveats in the interpretation of the Tn-seq screen data. First, the transposon library is not saturated and does not contain insertions into all non-essential genes. Mutants for several genes known to be important in tick colonization are not included in the library. Next, the transposon library we used for this study was generated in the 5A18NP1 background, which is missing two plasmids, lp56 and lp28-4. It is possible that the loss of the genes on these plasmids affects the requirement for certain genes or that the regulatory patterns are altered in their absence. Finally, the mechanism we used to infect the tick larvae, immersion feeding, is artificial and may lead to identification of genes that are not involved in natural transmission or, more likely, miss genes that are involved.
A more general caution about screening techniques such as Tn-seq is that the false discovery rate is dependent upon the stringency of the analysis used. We analyzed the data in two tiers. Using our most stringent criteria of no mutants isolated in either replicate, we did not detect any mis-identification in the subset of genes that were confirmed by additional experimental testing. Using slightly less stringent criteria of a 90% decrease in fitness, we already noted some false identifications of genes that have been previously shown to not be involved in tick survival including ospC. As with any screen, the goal is to enrich the identification of genes that are actually involved in a process while minimizing false identifications, but regardless of the stringency of the criteria, each of the genes will still need to be confirmed by additional testing. There was quite a bit of variability in frequency between genes that were not completely absent. This occurs because of stochastic loss of mutants due to bottleneck issues that can result in differences in the recovered mutants. One way to minimize this variability is to perform more experiments, which in our case, would mean adding more ticks for each replicate. By averaging results over more experiments, stochastic variability will decrease and we would have increased ability to identify genes with partial fitness impacts. At the numbers of ticks we used, the greatest confidence is for an extreme phenotype.
Understanding these caveats, we identified 46 genes whose disruption resulted in complete loss of Tn mutants from the population following colonization of larval ticks. Of these 46 genes, almost all have not previously been reported to be involved in survival in the tick, and thus the current study represents a significant advancement in our understanding of the genetic factors required for B. burgdorferi survival in the tick. Of note, many of the tick-phase genes we identified encode membrane-localized lipoproteins, and a significant portion (14 genes, including the five we selected for follow-up analysis), have been previously identified as important for resistance to reactive oxygen and nitrogen species (Table 2)[43].
We confirmed the phenotype for five of the novel tick-phase genes (bb0017, bb0164, bb0412, bb0050, and bb0051) by individual competition assays for survival in the tick following a blood meal (Fig 2B). Prior to our study, relatively little was known regarding the functions of these gene products, other than their predicted role in ROS resistance. All five gene products are predicted to be membrane-localized, and BB0164 has previously been shown to be involved in controlling intracellular manganese homeostasis [43]. To better understand whether these genes aid 1) the entry of B. burgdorferi into tick larvae during artificial infection, or 2) bacterial adaptation as the tick takes a blood meal, additional competition studies were carried out following the culture immersion step, but before the blood meal. All of the mutants tested survived as well as (or better than) the controls in the competition assays, suggesting that these genes are involved with survival of the blood meal and not with entry into the tick. That survival of the blood meal poses the larger barrier is not surprising. The blood infusion that the B. burgdorferi encounters in the midgut of the tick during feeding creates a rapidly changing environment for the spirochete. During the blood meal, there are changes in pH and temperature and exposure to reactive oxygen species (ROS), natural antibodies, and components of complement that can mediate spirochete killing [71–78]. Our results suggest that the ability to resist oxidative stress is likely critically important for survival in the tick host.
We performed further investigations to better understand the mechanisms by which one of the genes identified in our screen, bb0017, contributes to survival in the tick. Our in silico analysis suggested that BB0017 is part of a larger family of PII and PII-like proteins. However, there are some notable differences that distinguish BB0017 from these protein families, including differences in the lengths of two key loop regions (B/B’ and T/T’) and absence of key conserved residues involved in ligand binding. Thus, if BB0017 does bind ligands such as copper or c-di-AMP as has been previously shown for PII and PII-like proteins, it does so via a unique mechanism. c-di-AMP is produced in B. burgdorferi, although its potential function as a second messenger in this organism remains unclear [79,80]. It is certainly possible that the true ligand for BB0017 is a different molecule as there are significant differences between BB0017 structure and the structure of other PII proteins. Given the downstream effects of BB0017, it is tempting to speculate that it may bind c-di-GMP, which plays a critical role in B. burgdorferi gene regulation, however identification of binding of other molecules by BB0017 requires further experimentation.
RNA-seq analysis revealed that interruption of bb0017 by the Tn insertion results in significantly higher levels of transcription of the genes encoding DbpA and OspC; these results were corroborated by qRT-PCR as well as by immunoblot analyses using strains that had a complete deletion of bb0017 (S2 Table, Fig 5 and Fig 6A). This regulatory effect in the mutant appears to be mediated by increased levels of BosR and RpoS suggesting that BB0017 acts as a potential negative regulator of these important pathways. Repression of genes highly expressed during mammalian infection would be consistent with a role for BB0017 in tick colonization. Previous studies have shown that expression of B. burgdorferi genes that are required for one host may result in a fitness defect in colonization of the other host [69,73,76,77,81–83]. The effects of deletion of bb0017 on RpoS are also likely to affect expression of genes in the glp operon as seen by the RNA-seq studies (S2 Table). However, altered expression of the glp operon does not appear to account for the survival defect of the bb0017 mutant in our Tn-seq experiments as overexpression of the glp operon was not sufficient to restore the ability to survive the blood meal in a Δbb0017 background. This is consistent with the fact that we did not observe a fitness defect for the Tn::glpD mutant and that prior studies have shown that the glp genes are required at a time point later in the tick cycle than was evaluated in our study [33,38].
The elevated lipoprotein expression profile observed in the bb0017 mutant is strikingly similar to the phenotype of a mutant lacking the BmtA manganese (Mn) transporter. The bmtA mutant exhibits decreased intracellular Mn concentrations, which was shown to result in increased levels of ospC expression [84]. In the case of the bmtA mutant, the increased ospC expression is due to an increase in BosR protein levels at the post-transcriptional level, leading to increased transcriptional activation of RpoS [84]. The post-transcriptional regulation of BosR has also been observed in conditions where CO2 is limiting [85]. Our RNA-seq analysis suggests that bosR and rpoS transcript levels are similar in the Tn::bb0017 mutant and parental strains, despite the increase we observe in protein levels (Figs 5 and 6) suggesting that BB0017 affects BosR at the post-transcriptional level. In the bb0017 mutant, RpoS also appears to be upregulated at the post-transcriptional level, suggesting that some mechanism other than direct transcriptional activation by BosR is responsible for increased RpoS levels. There is also precedent for post-transcriptional regulation of RpoS, both in B. burgdorferi and in other bacteria [69].
The reciprocal expression of two sets of genes required for survival in the tick or mammalian hosts in response to a variety of environmental signals is paradigmatic to borrelial pathogenesis. However, the mechanisms by which these external stimuli are sensed remain to be fully characterized [69,73,76,77,81–83]. Given that BB0017 is predicted to be a membrane-localized signal transduction protein, we hypothesize that this protein may sense changes in the environment to regulate downstream effectors accordingly. The nature of the external signal, if any, sensed by BB0017 remains unclear, although it is likely not Mn. We previously showed that Mn levels are similar in the Tn::bb0017 mutant compared to the parental strain [43]. Given that PII and PII-like proteins generally affect downstream targets at the post-transcriptional level via direct protein- protein interactions, we predict that the BB0017 regulon may be larger than the list of genes identified in the RNA-seq analysis.
In conclusion, we have found that Tn-seq is a powerful tool in identifying B. burgdorferi genes important for fitness in surviving the blood meal. We have identified a large number of previously uncharacterized genes involved in the survival of the bacterium in its tick host. These results may provide important new avenues for exploration and understanding how the bacterium adapts to its different hosts. To this end we have further investigated the role of one of the genes identified, bb0017. We propose that BB0017 is a potential global regulator in B. burgdorferi that affects resistance to oxidative stress, survival in the arthropod host, and expression of key virulence determinants. As several of the other genes identified as important for survival of the bacteria during the early stages of tick larval infection also have been identified in prior screens for genes of ROS resistance, our results suggest the importance of ROS resistance in the initial colonization and persistence during the acquisition of the blood meal. Future Tn-seq screens can be tailored to identify genes required for survival during other parts of the bacterial lifecycle within the tick host. This approach will allow investigators to map the network of adaptations used by the bacteria to complete its life cycle.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.